Business & Industry Archives - Newswire — Business, Technology & Startup Press Releases

Inside Sweden’s AI Survey: What 234 Financial Firms Actually Reported

MMDNews — Wed, 06 May 2026 13:38:32 +0000

Editorial · Sector Brief · 16 min read

A structural read of Finansinspektionen’s 234-firm survey on AI in the Swedish financial sector — what the headline numbers say, what the trade press missed, and what the data actually tells us about how a regulated industry is absorbing generative AI in real time.

§ 01 · The Setup

A regulator surveyed 234 banks, insurers, and asset managers about AI. The trade press read the press release.

In December 2024, Finansinspektionen, the Swedish financial supervisory authority, published a working report on the use of artificial intelligence across the firms it supervises. The data underneath the report is unusually rich for the genre. Two hundred and thirty-four firms responded — banks, insurers, payment institutions, asset managers, fund managers, securities companies, electronic money issuers, clearing organisations, and stock exchanges. The response rate was 83 per cent of those surveyed. The report covers employee use of public generative AI tools, in-house AI deployment in production and pilot, sector-by-sector and size-by-size adoption breakdowns, intended investment direction over the coming twenty-four months, and explicit preparation for the EU AI Regulation that becomes applicable on 2 August 2026.

The report was covered briefly by the Nordic financial trade press, mostly through reproduction of the headline statistic that 84 per cent of firms now have employees using generative AI tools. That is, indeed, the most quotable number in the report. It is also, by a substantial margin, not the most interesting one. The interesting numbers sit further inside the document — in the gap between employee adoption and institutional readiness, in the unequal distribution of AI deployment across sub-sectors, in the explicit acknowledgement that most Swedish financial firms believe they are lagging international competitors, and in the regulatory exposure that the survey reveals about specific use cases that will be classified as high-risk under the AI Regulation.

This article is a structural read of the underlying data. We are not reproducing the headline finding the trade press already covered. We are looking at the report as a primary source on how a heavily-regulated industry actually absorbs a fast-moving technology — what gets adopted first, what stays at the pilot stage, where the institutional friction sits, and what the regulator’s framing implies about the supervisory expectations that are about to become operational reality. The Finansinspektionen survey is one of the most useful empirical documents on AI adoption in financial services published by any European regulator to date. It deserves more careful reading than it received.

— The Headline Numbers, Read Together —

84%

Firms with employees using generative AI tools

22%

Firms with AI systems in production or development

32%

Firms with a written GenAI policy for employees

41%

AI-using firms with a formal AI development policy

The four numbers, read together, describe a regulated industry where employee adoption has run far ahead of institutional governance. That gap is the story.

§ 02 · The Governance Gap

Employee adoption ran ahead of institutional readiness. Now the regulator is catching up.

The single most consequential finding in the Finansinspektionen survey is the gap between two adjacent statistics. Eighty-four per cent of firms have employees using generative AI tools as part of their work. Thirty-two per cent of firms have a written policy governing how those tools may be used. The remaining 52 percentage points are firms whose employees are pasting client information, internal documents, draft communications, code, and analytical outputs into commercial AI systems — without explicit institutional rules about what is and is not appropriate. This is not, on its own, evidence of widespread misuse. It is evidence that the institutional governance layer has not caught up to the operational reality, in an industry where the institutional governance layer is normally the differentiating asset.

The gap looks even sharper when you isolate the firms that have moved beyond employee tooling into actual AI deployment. Among the 22 per cent of firms with AI systems in production or development, only 41 per cent report a formally approved policy for AI development and use. The remaining 59 per cent are deploying AI in regulated business processes without a documented policy framework governing how those systems should be built, validated, monitored, or retired. Another 27 per cent of those firms say they are working on a policy but have not yet adopted one, and 30 per cent report no policy and no plan to draft one. From a supervisory perspective, those numbers describe an institutional readiness deficit that is going to become regulatorily expensive when the AI Regulation becomes operationally enforceable.

There is a structural reason this happened. Generative AI tools entered the workplace through a different door than every previous wave of enterprise software. Traditional financial-services technology adoption follows a predictable path: business case, vendor selection, procurement review, IT integration, security review, compliance review, training, rollout. The cycle takes nine to eighteen months and produces an institutional artifact — the policy document, the access control list, the audit log — before a single user touches the system. ChatGPT and its competitors arrived through the consumer browser. The cycle was zero. By the time the policy committee scheduled its first meeting, employees had been using the tools for six months.

That is not a uniquely Swedish phenomenon and it is not a uniquely financial-services phenomenon. What makes the Swedish data interesting is that we now have empirical confirmation of how large the gap is, in an industry that takes governance seriously. The same pattern almost certainly holds at higher magnitude in industries that take governance less seriously. The Finansinspektionen number — 32 per cent with policies in place — is probably the upper bound of governance readiness across the broader European economy, not the lower bound.

Table I — AI Adoption by Sub-Sector of the Swedish Financial Industry
Sub-sector	In Production	Pilot or Experimenting	No Use, No Plan
Banking & Credit Institutions	38%	38%	24%
Insurance	27%	41%	32%
Payments	24%	43%	32%
Securities Markets	16%	51%	33%
All Sectors (Survey Average)	22%	46%	32%

Banking leads on production deployment by a wide margin. Securities markets lead on experimentation but lag on production. The 32 per cent “no plan” share is consistent across sub-sectors, suggesting institutional resistance is structural rather than sector-specific.

§ 03 · Sector Asymmetry

Banks build. Securities firms experiment. The asymmetry is informative.

The sub-sector breakdown in the Finansinspektionen data is the part of the report that received almost no trade-press attention and is, by some distance, the most operationally informative part of the document. Banks and credit institutions report 38 per cent of firms with AI systems in production or development — nearly double the survey average of 22 per cent and more than double the figure for securities markets at 16 per cent. The same banking sub-sector also reports the smallest share of firms with no AI use and no plan, at 24 per cent. By any reading, banking is the part of the Swedish financial sector that has actually moved AI from concept to operational reality at scale.

Securities markets show the opposite pattern. Only 16 per cent of securities firms report AI in production, but 51 per cent are running pilots or experiments — the highest experimentation rate in the survey. The combined “in production plus piloting” figure for securities firms is 67 per cent, the same total proportion as banking, but the distribution is dramatically different. Securities firms are exploring AI; banks are deploying it. The gap is not about awareness or interest. It is about institutional appetite to take the operational, regulatory, and reputational responsibility of putting an AI system into a regulated production process.

There is a straightforward operational explanation for the asymmetry. Banks have the largest existing data infrastructure investment of any sub-sector, the deepest historical experience with traditional machine learning in credit scoring and fraud detection, and the most structured supervisory dialogue with Finansinspektionen on the topic. They were the natural first movers. Securities firms have a higher proportion of relatively recent fintech entrants, smaller average headcount, and trading workflows where the cost of an AI failure is concentrated and immediate — which encourages experimentation in research and back-office workflows but caution in anything that touches order execution. The asymmetry, in other words, reflects rational sub-sector economics, not differential ambition.

The implication for the next twenty-four months is that the production deployment gap between banking and the rest of the sector is likely to widen rather than close. Banks reported the highest planned investment increases across every AI category, and the firms already running AI in production are also the firms most heavily preparing for the AI Regulation. The firms that are still experimenting will face a harder transition: they will need to mature their pilot infrastructure, document their development processes, and build supervisory-grade governance frameworks essentially in parallel. The market will not be patient with that timeline.

The headline is that 84 per cent of firms have employees using generative AI. The story underneath is that 22 per cent have built anything with it, and only 41 per cent of those have a policy governing what they have built.

MMD Newswire · Editorial Read

§ 04 · Where AI Is Actually Being Used

Search, summarise, automate. The unromantic majority of real deployments.

The 44 firms in the survey with concrete AI use cases reported details on 83 specific deployments and aggregate information on a further 184. The most common categories — in declining order — are searching for or summarising information, process automation, customer insights, chatbots and virtual assistants, customer support, marketing and sales, fraud detection, text content generation, anti-money-laundering and counter-terrorism financing, software code generation, credit risk models, and translation. The pattern is unmistakable: the use cases that have actually moved into production are the ones with the lowest regulatory and reputational risk and the most direct productivity benefit. The dramatic, transformational use cases that dominate consultant decks — algorithmic trading, autonomous risk pricing, end-to-end credit decisioning — are conspicuously rare in the actual deployment data.

The technology mix underneath the use cases also flips a common assumption. Generative AI accounts for 45 per cent of the reported deployments and traditional machine learning for 41 per cent, with other AI technologies making up the remaining 14 per cent. Generative AI has, in less than three years, overtaken machine learning that financial firms have been operating in production for over a decade. That is a remarkable rate of adoption for a regulated industry, and it is exactly the kind of compositional shift that creates governance risk — the institutional muscle memory for evaluating, validating, and monitoring traditional ML is mature, and the equivalent muscle memory for generative AI is still being built. The risk is not that generative AI is being deployed in high-stakes use cases; the survey shows it largely is not. The risk is that the supervisory frameworks for evaluating it are still under construction even as deployment volume grows.

Most of the production-grade generative AI deployments in the data are extensions of existing communications and information-processing workflows. Firms are using AI to summarise public press releases, news items, and longer documents into formats suitable for social media and internal distribution. They are using AI to generate code through external services. They are using AI as a wrapper around customer support workflows. The infrastructure underneath these systems is, in every case the survey describes, a third-party generative AI service rather than internally trained models — which means the operational economics are exposed to the underlying provider’s pricing decisions and rate limits in ways that are not always visible at deployment time.

Table II — The 12 Most Common AI Use Cases in Swedish Financial Firms
Use Case	Reported Deployments	Editorial Read
Search & summarise information	28	The clear leader; low risk, high productivity, generative-AI native
Process automation	27	Largely traditional ML; new generative AI layers being added
Customer insights	22	Pattern recognition on structured customer data; established technique
Chatbots & virtual assistants	20	The most visibly customer-facing deployment category
Customer support & help desk	16	Increasingly using generative AI as the response engine
Marketing & sales	13	Content generation and personalisation workflows
Fraud detection	12	Mature traditional ML; generative-AI augmentation emerging
Text content generation	11	Mostly summarisation of public material for redistribution
AML & counter-terrorism financing	10	The regulator’s preferred AI use case; well-established
Software code generation	10	Almost entirely external services rather than in-house models
Credit risk models	9	Watch this one — will be high-risk under the AI Regulation
Translation	9	Largely commodified; minimal regulatory exposure

The use case profile is unromantic and exactly what a regulator hopes for at this stage of adoption: the high-impact, low-risk applications are leading. Source: Finansinspektionen AI Survey 2024, 83 detailed use cases reported.

§ 05 · The High-Risk Exposure

Fourteen firms expect to operate high-risk AI under the new regulation. The number will grow.

The AI Regulation classifies AI systems into risk tiers and imposes substantially heavier compliance obligations on systems classified as high risk. Within financial services, the regulation explicitly identifies two categories as high-risk: AI systems used to evaluate creditworthiness or establish credit scores of natural persons, and AI systems used for risk assessment and pricing in life and health insurance. The Finansinspektionen survey asked firms to self-classify their existing and planned use cases against this framework. Four per cent of respondents reported existing use cases they assess as high-risk, all in creditworthiness assessment. Fourteen firms reported it as probable or highly probable that they will operate a high-risk use case within the next twenty-four months — covering both creditworthiness and life and health insurance pricing.

Fourteen firms is a small absolute number, but the number understates the regulatory weight of the underlying exposure. High-risk AI systems under the regulation must satisfy a comprehensive set of obligations: risk management systems across the lifecycle, data governance covering training data quality, technical documentation, automated logging, transparency to deployers and end users, human oversight, accuracy and cybersecurity standards, conformity assessment before market deployment, and ongoing post-market monitoring. The compliance burden is substantial, and it falls on firms whose existing AI governance — as the survey itself documents — is in many cases not yet mature. The fourteen firms who already see themselves heading into this category are the leading edge. The firms behind them, who will discover their use cases qualify as high-risk closer to the August 2026 application date, will face a tighter compliance timeline.

The survey’s most quietly important data point sits in the responses about preparation. Sixty-eight per cent of firms with AI in production have begun preparing for the AI Regulation, and a further 23 per cent say they will start within three months. That leaves only 9 per cent of production-AI firms who reported neither current preparation nor near-term plans. Among firms only piloting AI, the readiness picture is different and worse: a meaningful share of pilot-stage firms believe their use cases will not be subject to the regulation, an interpretation the survey notes is “in many cases questionable.” Finansinspektionen is signalling, politely, that there will be supervisory follow-up with firms whose self-assessment of regulatory exposure does not survive contact with the actual text of the regulation.

The Nordic compliance picture is further complicated by the operational reality that many Swedish financial firms operate cross-border into the UK, the broader EU, and increasingly into the US. AI governance and disclosure frameworks differ across jurisdictions in ways that are not fully harmonised, and the practical work of mapping a single AI system against multiple regulatory regimes is non-trivial. Nordic accounting and advisory practices — firms like Sveago that handle cross-border invoicing and compliance work for Nordic technology and financial services firms expanding internationally — have started to see meaningful rising demand for AI-related advisory engagements as part of broader regulatory readiness work, particularly from firms that will need to satisfy both the AI Regulation and adjacent regimes simultaneously.

§ 06 · The Infrastructure Layer Underneath

The compute economics that don’t appear in the survey but shape every deployment decision.

The survey does not ask firms about the infrastructure costs underneath their AI deployments, and that absence is itself informative. A regulator focused on AI governance, risk classification, and supervisory readiness has no particular reason to ask about cloud and inference economics. But every operationally meaningful AI deployment in financial services runs on infrastructure provided by one of the three hyperscalers — Microsoft Azure, AWS, or Google Cloud — and increasingly relies on inference services from Anthropic, OpenAI, or a smaller specialised provider. Those infrastructure choices propagate into the unit economics of every deployment, the data residency posture, the latency characteristics, and the regulatory exposure profile.

For firms in the early piloting phase — the 46 per cent of the survey population — AI infrastructure costs are typically not yet a meaningful budget line because pilot volumes are low. For firms with AI in production at scale, the picture changes quickly. A mid-sized Swedish financial firm running embedded generative AI features in customer-facing or back-office processes can plausibly spend several hundred thousand kronor per month on AI inference alone, growing to materially more once usage scales. The market for recovering value from unused AI cloud and credit allocations has emerged in response to this, with brokers like AI Credit Mart matching buyers and sellers of Azure, AWS, GCP, and Anthropic credits across European and global financial services and technology buyers. None of this is captured in the Finansinspektionen survey, but it is the operational reality underneath the survey numbers.

The supervisory implication is real. A financial firm whose AI deployment economics depend on a third-party generative AI provider is exposed to that provider’s pricing decisions, rate limits, model deprecation schedules, and ultimately commercial risk. The AI Regulation places obligations on the “deployer” of an AI system as well as the provider, and a deployer whose infrastructure is concentrated in a small number of third parties is structurally less robust than one with internal capability or with diversified provider arrangements. The survey does not yet probe this dimension; it is, however, the dimension on which a meaningful share of supervisory attention is likely to focus over the next twenty-four months as production AI becomes a more material part of regulated business processes.

Table III — Top Risks and Challenges Identified by AI-Using Firms
Risk Category	Share of AI-Using Firms Citing	Editorial Read
Data quality	63%	The leading concern, well ahead of any other category
Data protection	54%	GDPR overlay; particularly acute for customer-data use cases
Accountability	43%	Reflects the open question of who is responsible for AI decisions
Third-party risks	42%	Vendor concentration is a real concern, especially for generative AI
Regulatory challenges	39%	The AI Regulation is the dominant framework, but not the only one
Lack of AI competence/resources	38%	Persistent talent constraint across the sector
Cybersecurity	33%	Both AI as attack surface and AI in defensive workflows
Compliance	30%	Distinct from regulatory; about the operational compliance workload
Explainability	29%	Higher concern at firms operating generative-AI components
Model deviation & monitoring	33%	The continuous-monitoring obligation that catches firms unprepared
Model validation/approval	28%	Lower than expected; suggests immaturity of validation frameworks
Implementation in production	17%	Quietly significant: nearly one in five cite execution as a top risk

Data quality and data protection lead the risk register. The lower-cited categories — particularly model validation and implementation in production — suggest the institutional muscle memory for AI deployment is still being built. Source: Finansinspektionen AI Survey 2024.

§ 07 · What The Trade Press Missed

Three findings that didn’t make the headlines.

The first is the international competitiveness signal. Most Swedish financial firms believe they are lagging international competitors on AI adoption — only 17 per cent rate themselves as ahead, and the median self-assessment sits clearly in “behind” territory. The two reasons firms give in the free-text responses are notable. The first is that regulation accompanying new AI implementations is perceived as burdensome. The second is that many firms see themselves as small actors internationally, with limited capacity to build out an extensive AI organisation. Neither reason is fully accurate as a description of the international competitive landscape, but both are real perceptions that drive investment behaviour. The implication is that the Swedish financial sector’s AI investment trajectory is being shaped not just by genuine competitive pressure but by the firms’ mental models of how that pressure works — and those models could be improved by better data on what international competitors are actually doing.

The second is the investment trajectory asymmetry. Firms already running AI in production plan to increase their generative AI investment over the next twenty-four months at materially higher rates than firms still in the pilot stage or with no AI use. Seventy-three per cent of production-AI firms plan to increase generative AI spend; only 26 per cent of firms with no AI use plan to. The same asymmetry holds for machine learning and other AI categories. The structural implication is that the existing AI capability gap between the leading firms and the laggards is going to widen, not narrow, over the next two years. This runs counter to the common assumption that early movers are setting up later movers to leapfrog them with newer technology. The data suggests the opposite: the early movers are pulling further ahead.

The third is the policy lag. Among firms with AI in production, 30 per cent report that they have no formal AI development policy and no plans to draft one. The other 27 per cent are working on it but have not yet adopted one. Only 41 per cent have an approved policy. From a supervisory readiness perspective these numbers are concerning, but the more interesting question is what they imply about the firms themselves. A firm running AI in regulated business processes without a written governance framework is operating on individual judgement rather than institutional process. That works at small scale and breaks at large scale. It is also the kind of operational vulnerability that attracts supervisory attention the moment something goes wrong. The August 2026 AI Regulation deadline is going to surface a lot of these gaps very quickly.

— Reader Questions —

Twenty questions on the survey, answered plainly.

What is the Finansinspektionen AI survey?

A 2024 questionnaire sent by the Swedish financial supervisory authority to 278 firms under its supervision, with 234 firms responding (an 83 per cent response rate). The survey covered employee use of generative AI, in-house AI deployment, sector breakdowns, investment plans, risk identification, and preparation for the EU AI Regulation. It is one of the most empirically rich documents on AI adoption in financial services published by any European regulator.

What is the most important finding in the report?

The gap between employee adoption (84 per cent) and institutional governance (32 per cent of firms with formal generative AI policies for employees, 41 per cent of AI-using firms with formal AI development policies). Employee adoption ran far ahead of policy frameworks, and the supervisory implications of that gap are about to become operationally consequential as the AI Regulation becomes applicable in August 2026.

Are 84 per cent of Swedish banks really using ChatGPT?

More precisely, 84 per cent of Swedish financial firms have employees who use generative AI tools as part of their work. The survey does not break this down by frequency, intensity, or institutional sanction. It is best read as a measure of penetration, not depth. The headline figure has been widely repeated; the more interesting figure underneath is how many firms have governance frameworks for that use.

Why is banking ahead of other sub-sectors on AI deployment?

Banking has the largest existing data infrastructure investment of any sub-sector, the deepest historical experience with traditional machine learning in credit scoring and fraud detection, and the most structured supervisory dialogue with Finansinspektionen on the topic. They were the natural first movers. Securities firms experiment more but deploy less, reflecting the concentrated cost of AI failures in trading workflows.

What does the AI Regulation classify as high-risk in financial services?

Two categories are explicitly identified: AI systems used to evaluate the creditworthiness of natural persons or establish their credit score, and AI systems used for risk assessment and pricing in life and health insurance. Other AI use cases in the financial sector may also fall into the high-risk category depending on context, but those two are the ones the regulation specifically calls out.

When does the AI Regulation actually become operational?

The regulation entered into force on 1 August 2024, but most of its substantive obligations apply from 2 August 2026. Some prohibitions and general-purpose AI rules apply earlier. For firms operating high-risk AI systems in financial services, the August 2026 date is the practical compliance deadline, and meaningful preparation work should already be in progress.

What proportion of Swedish financial firms expect to operate high-risk AI?

Currently 4 per cent of survey respondents report existing use cases they assess as high-risk, all in creditworthiness evaluation. Fourteen firms reported it as probable or highly probable that they will operate a high-risk use case within twenty-four months, covering both creditworthiness and life and health insurance pricing. The number is likely to grow as the regulation is closer to application and firms refine their classifications.

Is generative AI really already more common than traditional machine learning?

In the Swedish financial sector, yes — but with caveats. Generative AI accounts for 45 per cent of reported deployments and traditional machine learning for 41 per cent. That is a remarkable rate of adoption for a regulated industry over less than three years. However, the survey covers in-production deployments only, and traditional ML continues to underpin many of the most operationally critical processes in the sector. Generative AI is more visible; ML remains structurally important.

What are the most common AI use cases in the survey?

Searching for and summarising information leads with 28 reported deployments, followed by process automation (27), customer insights (22), chatbots and virtual assistants (20), customer support (16), marketing and sales (13), fraud detection (12), text content generation (11), AML/CFT (10), software code generation (10), credit risk models (9), and translation (9). The leading categories are exactly the low-risk, high-productivity use cases a regulator hopes to see at this stage of adoption.

What are the biggest risks firms identify in their AI use?

Data quality leads at 63 per cent, followed by data protection (54 per cent), accountability (43 per cent), third-party risks (42 per cent), regulatory challenges (39 per cent), and lack of AI competence or resources (38 per cent). The pattern is interesting: the leading risks are operational and data-related rather than algorithmic. Firms are more worried about whether their data is fit for AI than they are about model behaviour itself.

Are Swedish financial firms prepared for the AI Regulation?

Among firms with AI in production, 68 per cent have begun preparing and 23 per cent plan to start within three months — meaning 91 per cent of production-AI firms are within three months of beginning preparation. That is reasonably high readiness. The picture is less reassuring for pilot-stage firms, where a meaningful share believe their use cases will not be subject to the regulation, an interpretation Finansinspektionen describes diplomatically as “questionable.”

What is “human in the loop” and why does it matter for AI Regulation compliance?

It refers to AI systems where a human reviews, approves, or overrides the AI’s output rather than allowing fully autonomous operation. In the survey, 57 per cent of reported use cases include human in the loop, while 19 per cent are fully autonomous. The AI Regulation places significantly greater oversight obligations on autonomous AI systems, particularly in high-risk categories, which is why most production deployments today retain human review in the decision chain.

How are firms managing the risks they identify?

The most commonly reported measures are evaluating models, services, and suppliers; introducing human oversight; investing in employee training and education; clarifying processes; testing models; monitoring data quality; and conducting other types of data control. Forty-one per cent of AI-using firms report a formally approved AI development policy, with another 27 per cent planning to draft one. The remaining 32 per cent operate without a documented policy framework.

Why do most Swedish financial firms believe they are lagging international competitors?

Two reasons emerge in the free-text responses. First, regulation accompanying new AI implementations is perceived as burdensome and slowing development. Second, many firms see themselves as small actors internationally with limited capacity to build out extensive AI organisations. Both perceptions are real even where they are not fully accurate descriptions of the global competitive landscape, and both shape investment behaviour.

Will the gap between AI leaders and laggards widen?

According to the data, yes. Firms already running AI in production plan to increase generative AI investment at materially higher rates than firms still piloting or without AI use — 73 per cent of production-AI firms plan to increase generative AI spend versus 26 per cent of firms with no AI use. The same asymmetry holds across other AI categories. Early movers are extending their lead rather than being caught by later movers.

What about explainability of AI systems?

Sixty-four per cent of firms consider their AI use cases to have high or very high explainability. Cases rated as having low or very low explainability (13 per cent) primarily use generative AI as a sub-component within larger systems. The pattern is consistent with the broader observation that traditional ML systems are easier to explain than generative AI components, but the firms running them have built up institutional comfort with the trade-offs over time.

Why does the survey not capture infrastructure costs?

Because the regulator’s mandate is supervisory and prudential, not commercial. Cloud and inference economics are not directly within the remit of a financial supervisor, even though they shape every operational AI decision underneath the supervised activity. The absence is itself informative — it points to a dimension of AI deployment that is operationally critical but supervisorially under-examined.

How concentrated is the third-party AI infrastructure market for Swedish firms?

Highly. Almost all production AI in Swedish financial services runs on top of one of three hyperscalers (Azure, AWS, GCP) and increasingly relies on inference services from a small number of providers including Anthropic and OpenAI. The concentration is not unique to Sweden; it is a global pattern in financial services. The third-party risk concern that 42 per cent of AI-using firms cite reflects this concentration directly.

Are smaller financial firms at a structural disadvantage on AI?

The survey suggests modestly so. Smaller firms are overrepresented in the “no AI use, no plan” category, and the firms that have moved farthest with AI are also planning the largest investment increases. The disadvantage is not absolute — smaller firms can adopt productivity tools and outsource specialised AI capability — but the gap on production deployment in regulated processes is widening.

What should a financial firm do if it has not yet started AI Regulation preparation?

Begin with a complete inventory of AI use across the organisation, including employee use of public generative AI tools. Apply the regulation’s risk-tier framework to each system or use case. Identify any use case that may qualify as high-risk under the regulation. Build a written governance framework covering development, deployment, monitoring, and decommissioning. The August 2026 deadline is closer than it looks once the inventory work is started.

What is the Swedish financial sector’s likely AI trajectory over the next two years?

Continued production deployment in low-risk, high-productivity categories, particularly information processing, automation, and customer support. Growing high-risk exposure in creditworthiness evaluation and life and health insurance pricing as deployments mature. Tighter institutional governance forced by AI Regulation compliance. Concentration of AI capability in firms that are already leaders, with the gap to laggards widening rather than closing. The structural picture is one of accelerating differentiation rather than catch-up.

Source · Primary Document

Finansinspektionen, AI in the Swedish Financial Sector, FI Ref. 24-18158, 6 December 2024.

Published by the Swedish Financial Supervisory Authority. Read the full report: finansinspektionen.se

— Editor’s Note —

On reading regulators slowly.

Reports from financial supervisors are an under-exploited source of business intelligence in the technology economy. They are written carefully, sourced rigorously, and produced specifically to inform institutional decision-making. They are also generally read once, summarised in a paragraph, and forgotten. The Finansinspektionen AI survey deserves longer reading than it received in the Nordic trade press, both because the underlying data is genuinely informative and because the regulator’s own framing of the data tells you something useful about how supervisors are thinking about AI in regulated industries.

MMD Newswire is editorially independent. The interpretations, framings, and structural reads in this article are our own. Readers in regulated industries should treat this as a starting framework for thinking about the survey’s findings, not a substitute for the legal and compliance work that AI Regulation readiness actually requires. The full Finansinspektionen report is publicly available from the regulator and worth reading in primary form, particularly for compliance, risk, and technology functions whose work is about to become noticeably more demanding.

The post Inside Sweden’s AI Survey: What 234 Financial Firms Actually Reported appeared first on Newswire — Business, Technology & Startup Press Releases.

Best Swedish Fintechs in 2026: An Opinionated Watchlist

MMDNews — Thu, 09 Oct 2025 14:32:00 +0000

Editorial · Founder Notebook · 5 min read

Seven Swedish fintechs I keep coming back to — what each of them actually does, what makes me think they have a real shot at category leadership, and where I’m sceptical. Not a ranking. A working list.

Stockholm has become an unusually concentrated fintech ecosystem — over a thousand fintech firms in Sweden according to Dealroom, around 42 per cent of the entire Nordic region’s startup count. Andreessen Horowitz partner Gabriel Vasquez recently mentioned he took nine flights from New York to Stockholm in a single year, and a16z just led a pre-seed into a Swedish AI dental admin startup. There is a reason US capital keeps showing up here. The companies below are seven of the ones I think are worth paying attention to in 2026 — with notes on what each does, why I think they could become the category leader, and what I’m watching for that could derail them.

01 · Stockholm · founded 2020

Quartr

Earnings calls and IR research, packaged as a podcast app and an institutional API.

Quartr is the one I think most people are still underestimating. The mobile app is a podcast-style player for earnings calls, which sounds like a niche tool until you realise that four out of five of the world’s largest hedge funds are using their Pro product and Perplexity, Yahoo Finance, and Fortune are wired into their API. They reported 300 per cent ARR growth year-on-year and raised ten million dollars in 2025 to open offices in New York and Dublin. The thing that makes me bullish is that they’ve quietly become the structured-data layer for qualitative public-company information, and that is exactly the layer AI systems and research platforms need. The risk is competition from Bloomberg or FactSet building a comparable product internally — but Quartr’s data depth is real, and incumbents tend to under-invest in the long-tail content categories that don’t fit their existing taxonomies.

02 · Stockholm · founded 2020

Grasp

Multi-agent AI for investment banks, consultants, and PE firms — valuation, pitch decks, target lists.

Grasp is doing the thing investment banks have wanted forever, which is taking the most painful parts of the analyst workflow — comp tables, target identification, market research, the inevitable PowerPoint — and turning them into AI-driven outputs. They reported a 3.5x ARR jump in 2025, serve nearly two hundred customers including most of the Big Four and major investment banks, and just opened a London office. What makes me think they have category-leader potential is that they’re building inside a workflow that hasn’t materially changed in twenty years and where the buyers (banks, PE) have huge budgets and clear ROI metrics. The risk is the same risk every M&A research vendor faces — specialist competitors emerging in specific verticals (one of which, full disclosure, is Tablestat in Nordic software M&A research), and the incumbents (FactSet, Refinitiv, Capital IQ) bolting AI onto their existing platforms. Grasp’s moat will be how deeply they embed in the analyst workflow, not the AI itself.

03 · Stockholm · founded 2018

Mynt

Corporate cards plus AI-driven expense management for SMEs across the Nordics.

Mynt is the most boring-sounding company on this list and also probably the one with the clearest path to dominating its category. They handle the corporate card and expense workflow for over twenty thousand SMEs across Sweden, Norway, and Finland, and they just announced a partnership with Nordea to launch a joint white-labelled card and spend management product across the entire Nordic region in 2026. Visa took a strategic stake in their Series B. The thing I keep coming back to is that distributing through Nordea gives Mynt access to a meaningful share of every SME in the region without having to pay the customer acquisition cost themselves — and the Visa relationship gives them a path into pan-European banks. The risk is that the SME spend management category is fragmenting fast, with Pleo, Spendesk, Soldo, and Brex all chasing similar deals. But Mynt has the Nordic-bank distribution channel none of those competitors have.

04 · Stockholm · founded 2006

Northmill Bank

A licensed Swedish challenger bank growing the B2B lending book 4.5x in a year.

Northmill is the boring-but-impressive entry. Earnings before tax up 56 per cent year-on-year, the B2B lending portfolio expanding 4.5-fold, 211 thousand card customers (a 3.5x increase), and they just launched a Swedish mortgage product. They’re also a fully licensed bank, which matters more in 2026 than it did three years ago because the regulatory environment around fintech has tightened across Europe and a banking licence is increasingly a competitive moat rather than a compliance burden. The thing that makes me think Northmill is one to watch is that they’ve been quietly building the unsexy parts of a digital bank — lending operations, payment certifications, regulatory infrastructure — while flashier neobanks have been competing on app design. The risk is that Sweden is a small home market and the international expansion is yet to be tested. But the unit economics on the existing business are good enough that the home market alone supports a credible scale story.

05 · Stockholm · founded 2015

Froda

SME lending in minutes through an embedded finance platform — expanding across Europe.

Froda turns the small-business loan from a multi-month ordeal into a few-minute decision, which is one of those problems that sounds incremental until you realise how much SME activity it actually unlocks. They’ve financed over 120 thousand loans, work with fifteen-plus European banks and fintech partners through their embedded finance platform, and raised fifty million euros in Series B last year. The European Investment Fund extended its partnership with them in 2025 to launch a pan-European microfinance guarantee. What makes me think Froda has category-leader potential is that they’re selling into banks rather than competing with them — the embedded finance positioning means every additional bank partnership is a distribution multiplier rather than a customer acquisition cost. The risk is the credit quality of the underlying loan book in a downturn, which the rapid growth makes harder to assess from the outside.

06 · Stockholm · founded 2022

Bits

Compliance automation for KYC, KYB, AML, and fraud across 100+ jurisdictions.

Bits is the youngest company on the list and the one I think has the highest variance in possible outcomes. They just raised twelve million euros in Series A in February 2026 to fund expansion into the DACH region and the UK. The pitch is that fragmented compliance stacks — one tool for KYC in Sweden, another for sanctions screening in Germany, another for beneficial ownership data in the UK — can be replaced by a single platform that automates the whole workflow. They claim 50–70 per cent reductions in manual case handling and four-to-six-times faster onboarding for clients including Qliro, Alisa Bank, and Walley. The thing that makes me think Bits could be the category leader is that NIS2 and DORA compliance is now a meaningful operational burden across European financial services, and a tool that genuinely reduces it has structural tailwinds for the next few years. The risk is the category is crowded with well-funded competitors (ComplyAdvantage, Fenergo, Persona) and Bits is small enough that a couple of slow quarters could be existential.

07 · Stockholm · founded 2024

Dentio

AI scribe and admin tool for dentists — Andreessen Horowitz pre-seed.

Dentio is on the list because of who backed them rather than what they’ve built so far. A 2.3 million dollar pre-seed from Andreessen Horowitz on a vertical AI tool for dental admin is a small cheque for a16z but a meaningful signal about what Sweden’s pre-seed ecosystem is producing right now. The product itself is an AI scribe for clinical notes, which the founders themselves admit is heading toward commodity status — their bet is that they can extend into the broader dental admin workflow before the scribe layer commoditises. What I’m watching is whether they execute the vertical-deepening fast enough to defend against horizontal AI scribes (and against Tandem Health, the better-funded Swedish competitor that operates across multiple medical specialties). The reason this is on the list anyway is that the a16z signal is real, the SSE Business Lab pedigree is genuine, and the founders cleared the “reached out to zero investors, deal happened through referrals” bar that almost nobody clears. Dentio is the lottery ticket on this list. The other six are the businesses.

— Reader Questions —

Eight questions on the Swedish fintech scene.

Why Sweden specifically?

Sweden has produced an outsized share of European fintech successes — Klarna, Tink, Zettle, iZettle — and the ecosystem keeps generating well-funded follow-on companies. Around 42 per cent of all Nordic fintech startups are Swedish according to Dealroom, and the SSE Business Lab incubator at the Stockholm School of Economics has produced an unusual concentration of breakout names including Klarna, Voi, and Legora. US venture investors have noticed: Andreessen Horowitz alone has been in Stockholm repeatedly over the past year, including a recent pre-seed into Dentio.

Why isn’t Klarna on this list?

Because the question was about companies worth watching in 2026, and Klarna is already the canonical Swedish fintech — everyone is watching it. The interesting names are the ones that haven’t made the global headlines yet, which is why this list focuses on companies one or two stages earlier in their visibility curve. Klarna’s outcome is now mostly about its IPO timing rather than category-defining product moves.

What separates a category leader from a fast-growing fintech?

Distribution, distribution, distribution. Most fintech categories aren’t won on product alone — they’re won by the company that solves the customer acquisition cost problem, which usually means embedding into a bank, a network, or an industry workflow that already has the audience. Mynt’s Nordea partnership is the cleanest example on this list. Froda’s embedded finance positioning is another. Quartr’s API distribution into research platforms and AI systems is the third version of the same idea.

How seriously should I take a16z’s pre-seed cheques in Stockholm?

Seriously, but in the right way. A 2.3 million dollar pre-seed cheque is small relative to a16z’s overall capital base, but the partner travel and the deal sourcing pattern are the real signal. US firms don’t take that many flights to Stockholm if they don’t expect the next wave of breakout European AI companies to come from this region. Treat it as forward-looking flow data rather than a bet on any specific portfolio company.

What’s the regulatory backdrop in Sweden right now?

Tighter than three years ago. The EU’s MiCA crypto framework took effect in late 2024. DORA — the Digital Operational Resilience Act — began applying to financial firms in January 2025. NIS2 was implemented in Sweden through national legislation in January 2026, adding another cybersecurity compliance layer. The combined effect is that fintech operating in Sweden now sits inside three overlapping regulatory regimes, which favours licensed, well-resourced players (Northmill, Mynt) and creates demand for compliance automation tools (Bits).

Are these companies likely IPO candidates?

Most of them are too early for IPO, and the ones that aren’t will probably stay private for a while given current public-market sentiment. Quartr and Northmill are the most plausible eventual candidates on a three-to-five-year horizon. Mynt could go either way depending on how the Nordea partnership performs through 2026 and 2027. Froda has the clearest scale story for a Nordic fintech listing. Bits is too early. Dentio is too early. Grasp is interesting but the public market for AI-services companies isn’t mature enough yet to support a clean listing.

Which of these is the most likely acquisition target?

Quartr, almost certainly — either a major financial data provider (Bloomberg, FactSet, S&P) or one of the AI platforms wiring up financial knowledge graphs. The data layer they’ve built is structurally valuable to anyone trying to do qualitative public-market research at scale. Mynt is a plausible target for a major European bank wanting an SME spend product. Bits could be acquired by a larger compliance vendor if they don’t reach scale on their own. Northmill is harder to acquire because of the banking licence, but a strategic minority stake from a larger Nordic bank is plausible.

What’s the strongest signal you watch for in early-stage Swedish fintech?

SSE Business Lab pedigree — if the founders went through that incubator, the probability of a credible team is materially higher than the average European pre-seed. The second strongest signal is whether the founders raised through inbound rather than outbound, which usually means the product is genuinely good rather than well-marketed. Dentio explicitly described raising through referrals without ever pitching investors directly — that’s rare and almost always a positive signal.

Sources · Further Reading

6 Fintech Startups from Sweden to Follow in 2026, Fintech News Nordics, February 2026: fintechnews.ch

Anna Heim, Have money, will travel: a16z’s hunt for the next European unicorn, TechCrunch, February 2026: techcrunch.com

The post Best Swedish Fintechs in 2026: An Opinionated Watchlist appeared first on Newswire — Business, Technology & Startup Press Releases.

Inside the SaaSpocalypse: A Structural Read for 2026

MMDNews — Tue, 06 May 2025 14:22:00 +0000

Editorial · Market Structure · 17 min read

The SaaSpocalypse is real, partly. The per-seat pricing model is genuinely under pressure from agentic AI, the public market sell-off was overdue, and the IPO window for late-stage SaaS has effectively closed for now. None of that is the same thing as the death of SaaS — and the trade press is conflating four very different stories into one.

§ 01 · The Setup

A trillion dollars in software market cap evaporated in two months. The trade press called it the SaaSpocalypse and moved on.

Between January and early March of 2026, the listed software-and-services sector lost something on the order of one trillion dollars in aggregate market value. Salesforce, Workday, Adobe, Microsoft, Shopify and the rest of the established SaaS bench all took meaningful drawdowns. The selling waves were correlated with specific AI product launches — Anthropic released Claude Code for cybersecurity, and a basket of cybersecurity-adjacent SaaS stocks slid; Anthropic released legal tooling in Claude Cowork, and the iShares Expanded Tech-Software ETF moved with it. The pattern was visible enough that one investor, quoted in TechCrunch, gave it a name: FOBO investing, the fear of becoming obsolete. The public-market trade press settled on a name for the broader phenomenon: the SaaSpocalypse.

The narrative that has crystallised around the term is roughly the following. AI-native tooling has lowered the barrier to building software so dramatically that customers can plausibly choose to build their own internal alternatives rather than license SaaS products. Klarna’s late-2024 announcement that it had replaced Salesforce’s flagship CRM with an internal AI-built system became the canonical reference point. Per-seat pricing — the foundational economic model of the modern SaaS industry — is being undermined as employers replace seats with AI agents. The rapid pace of AI development means new tools can replicate not just the core functionality of a SaaS product but the upsell add-ons that drive net retention. And the IPO window for late-stage SaaS has, in practical terms, closed.

The narrative is broadly correct. It is also, like most narratives that crystallise during a market correction, simultaneously overstating the speed of the structural change and understating the depth of it. The death-of-SaaS reading is overstated. The per-seat-pricing-is-broken reading is understated. The trade press is collapsing four genuinely distinct stories into one undifferentiated headline, and the result is that public market sentiment is moving faster than the underlying economic shift, which is itself moving faster than most SaaS executives have publicly acknowledged. This article is an attempt to separate the four stories.

— The Headline Numbers, Read Together —

~$1T

SaaS market cap erased in roughly eight weeks

70–90%

SaaS gross margins, the asset under threat

$100M

Sierra ARR in under two years on outcome-based pricing

Venture-backed SaaS IPOs filed or expected near-term

Four numbers describing four different stories: a public market correction, a margin profile under threat, a working alternative pricing model, and a frozen IPO window. They are related, but they are not the same.

§ 02 · Story One — The Pricing Model

Per-seat pricing is structurally broken when an AI agent does the work of ten employees.

The per-seat licence is the foundational economic unit of the modern SaaS industry. It produces predictable recurring revenue, scales effortlessly with customer growth, and gives the seller a built-in upsell path through the customer’s headcount expansion. SaaS gross margins of 70 to 90 per cent are produced by exactly this model, and most of the canonical valuation frameworks for the category — revenue multiples, net retention rates, payback periods — were calibrated against it. Per-seat pricing is the reason SaaS, as a category, became the most-favoured business model in technology investing for over a decade. The current correction is, in part, a forced acknowledgement that the assumption underneath this model is now genuinely unstable.

The mechanics are straightforward. A SaaS vendor selling a customer support platform at $150 per agent per month is collecting revenue indexed to the size of the customer’s support team. When the customer replaces eight of those ten agents with an AI system, the vendor’s revenue from that customer drops by 80 per cent unless the contract is restructured. The contract typically cannot be unilaterally restructured, but it can be renegotiated at renewal. And renewals, across the SaaS industry, are now where the structural pressure is concentrating. Customers arrive at the renewal table with both the credible threat of building their own alternative and the credible threat of consolidating the work into fewer seats. Either way, the vendor’s gross dollar retention takes a hit it has not historically had to absorb.

The transition, though, is happening more slowly than the public market reaction implies. Per-seat contracts have multi-year terms, embedded penalty clauses, and procurement-cycle inertia that protect the existing revenue base for the duration of the current contract. Vendors with deep enterprise install bases — Salesforce, Microsoft, ServiceNow, Workday — will absorb the pressure over a renewal cycle of three to five years rather than over a quarter. The public market is pricing the displacement as if it were happening at quarterly speed; the operational reality is that it is happening at contract-renewal speed, which is dramatically slower. That mismatch is part of why the sell-off looks excessive in the medium term even as the underlying structural concern is genuine.

The vendors that will survive the transition with their margins intact are the ones that can credibly migrate from per-seat pricing to consumption-based or outcome-based pricing without losing customer trust during the migration. Sierra, the customer-support agent platform founded by former Salesforce CEO Bret Taylor, has been the cleanest demonstration of an outcome-based model working at scale — reaching $100 million in annual recurring revenue in under two years on a model that charges based on resolved customer interactions rather than per agent seat. The Sierra reference point is the most-cited proof in current SaaS pitch decks that an alternative model is operationally viable. It is also the model that most established SaaS vendors will struggle to migrate to without cannibalising their existing revenue.

Table I — Pricing Models in the Post-SaaSpocalypse Landscape
Model	How it Works	Vendor Exposure
Per-seat Subscription	Customer pays per active user, billed monthly or annually	High — revenue erodes as agents replace seats
Per-tier Subscription	Flat fee for usage tier, with overage charges above thresholds	Medium — insulated from seat compression but vulnerable to consumption shifts
Consumption-based	Customer pays per token, API call, or unit of compute consumed	Medium — tracks usage but exposes vendor to inference cost volatility
Outcome-based	Customer pays per resolved task, ticket, or business outcome	Lower — aligned with value delivered but harder to forecast
Platform Fee + Usage	Base subscription plus consumption overage; the emerging hybrid	Lower — combines predictability of subscription with usage upside
Build-Your-Own	Customer foregoes vendor and builds internal AI-powered alternative	Total — vendor loses the customer entirely

The transition is not from per-seat to a single new model. It is from a near-uniform pricing standard to a fragmented set of approaches whose suitability depends on the specific customer workflow.

§ 03 · Story Two — The Public Market Correction

A correction was overdue. AI was the trigger, not the cause.

The trillion dollars of evaporated market cap is the most quotable number in the SaaSpocalypse narrative. It is also the number least directly attributable to AI. SaaS valuations had been visibly stretched for several years before the AI agent wave gave the market a justification to reprice them. Multiples on forward revenue were running at levels last seen during the zero-interest-rate era of 2020 to 2022, despite interest rates having normalised at substantially higher levels. The cost of capital for SaaS customers has risen, slowing growth budgets across the buyer base. The pace of net new customer acquisition for mature SaaS vendors had been decelerating quietly for several quarters before the AI narrative forced a louder reassessment. The repricing was, in part, a delayed correction that had been postponed by the lack of an obvious catalyst.

When the catalyst arrived, in the form of fast-moving AI agent products from Anthropic, OpenAI, and a long tail of well-funded challengers, the market response had two components mixed together. The first was a justified repricing of terminal value — the recognition that SaaS revenue may not compound at the rates previously modelled because the customer behaviour underneath those models is genuinely changing. The second was the reactive sell-everything pattern that institutional investors apply when a category narrative shifts: sell first, ask questions later, distinguish between stronger and weaker names afterwards. Both happened simultaneously, and the result was a sell-off that overshoots the underlying structural shift in the short term and may undershoot it in the medium term once individual company performance becomes visible.

The Sifted commentary on the SaaSpocalypse made a useful structural distinction that the US trade press largely ignored. The European public-market exposure to this dynamic is materially different from the US exposure. European SaaS companies tend to operate in narrower verticals, with closer customer relationships, tighter capital efficiency, and less reliance on the growth-at-all-costs playbook that inflated US valuations during the zero-interest-rate years. The downdraft has caught European SaaS in its slipstream, but the structural fundamentals are different enough that the European correction is best read as a sympathy move rather than a referendum on European SaaS economics. The distinction matters operationally for any European SaaS company calibrating its response to the US-centred narrative — importing someone else’s existential crisis is a recurring failure mode in cross-Atlantic financial commentary.

A correction that overshoots in the short term and undershoots in the medium term is the most common shape a structural shift takes in public markets. The SaaSpocalypse is not the first such correction. It is the AI-flavoured 2026 instance of a recurring pattern.

MMD Newswire · Editorial Read

§ 04 · Story Three — The Infrastructure Layer

The AI-native challengers have a margin profile that nobody is talking about.

The piece of the SaaSpocalypse narrative that has received almost no serious analysis is the unit economics of the AI-native challengers themselves. Every AI-native company displacing an established SaaS vendor is running on top of one or more third-party large language models — OpenAI, Anthropic, Google, Mistral, or one of a smaller specialised set. Inference costs at production scale for any meaningfully agentic workflow are not trivial. A customer service AI handling a hundred thousand interactions per month is consuming tokens in volumes that translate, at current API pricing, into operating costs that would have horrified a 2018 SaaS investor accustomed to 85 per cent gross margins on infrastructure that cost almost nothing per incremental customer.

The result is that the gross margin profile of the AI-native generation is structurally lower than the gross margin profile of the SaaS generation it is displacing. Pure-play AI-agent products at production scale typically run at gross margins in the 50 to 70 per cent range — meaningfully lower than the 70 to 90 per cent range that defined the SaaS golden era. This is not a temporary state of affairs that will resolve itself when inference costs fall; even as costs decline, the consumption-based pricing model means the customer captures most of the cost reduction rather than the vendor expanding its margin. The AI-native model has structurally different economics from the SaaS model. Public market investors are pricing AI-native companies on growth metrics borrowed from SaaS, and the underlying margin profile may not justify the comparison once growth normalises.

The operational pressure this creates on AI-native startups is real and is already shaping behaviour. Companies running embedded AI features at scale are routinely spending six figures per month on inference alone, with that line growing as user volume grows. The inflexible relationship between usage and infrastructure cost is the single most consequential difference between AI-native unit economics and SaaS unit economics, and it is not yet priced into the public market valuations of the AI-native peers that have started to list. A market for recovering value from unused AI cloud and inference allocations has emerged in response to this pressure, with brokers like AI Credit Mart matching buyers and sellers of Azure, AWS, GCP, and Anthropic credits across the AI-native and SaaS-transitioning customer base. The market for credit recovery exists precisely because the underlying inference economics are tight enough that a few percentage points of effective compute cost matter to the business case.

The implication for the SaaS-versus-AI-native framing is that the comparison is more nuanced than the trade press allows. SaaS has a margin profile under threat from AI-native challengers; AI-native challengers have a margin profile that is structurally lower than the SaaS profile they are displacing. Both can be true, and both are. The eventual market structure will probably be a hybrid in which mature SaaS vendors migrate to lower-margin consumption-based pricing and AI-native challengers migrate to higher-margin proprietary infrastructure or vertically-integrated model training. The end-state economics may sit somewhere between the two starting points, which is a less dramatic story than the SaaSpocalypse framing implies but a more accurate one.

Table II — SaaS vs AI-Native Unit Economics: A Structural Comparison
Dimension	Classic SaaS	AI-Native
Typical gross margin	70–90%	50–70%
Marginal cost per user	Near-zero	Material; scales with usage
Pricing logic	Per seat, predictable	Per outcome or per token, variable
Infrastructure dependency	Cloud hosting; commoditised	Inference providers; concentrated
Net retention dynamics	Seat expansion + add-on upsell	Usage growth + workflow expansion
Switching cost	High — embedded data, integrations	Lower — less mature integration footprint
Compliance posture	Mature; audit-friendly	Still emerging; regulators catching up
Time-to-revenue maturity	2–4 quarters typical	Faster initial growth, slower steady state
Defensibility	Data ecosystem + workflow lock-in	Proprietary models + vertical specialisation

The two models do not converge on identical economics. They converge toward each other from opposite starting points, with hybrid pricing and infrastructure strategies emerging in the middle.

§ 05 · Story Four — The Frozen IPO Window

There are no SaaS IPOs in the pipeline. There are also no SaaS IPOs in the pipeline a year from now.

The most concrete operational consequence of the SaaSpocalypse is the closing of the IPO window for late-stage SaaS companies. Recent industry analysis indicates there are no venture-backed SaaS filings on the immediate horizon, and none expected to file in the near term. The IPO market has thawed for some other sectors — AI infrastructure, certain biotech segments, and a handful of consumer plays — but SaaS specifically is locked. Late-stage private SaaS companies that would have been the natural 2025 or 2026 IPO class — Canva, Rippling, and a handful of others — are visibly hesitating. The IPO window is not closed because there is no demand to list; it is closed because the receiving market is too volatile, the comparable public valuations are too unstable, and the incremental risk of going public into the current sentiment is greater than the marginal benefit of liquidity.

The downstream effect on the private market is real. Late-stage SaaS companies are finding it difficult to raise extension rounds at terms that approximate their previous private valuations. Some are taking flat rounds; some are taking modest down rounds; some are doing structured deals with investor-friendly liquidation preferences and ratchet protection that protect the headline valuation at the cost of future founder and employee dilution. None of these structures shows up cleanly in the funding announcements that get picked up by the trade press, which means the public picture of late-stage private SaaS pricing looks more stable than the underlying reality. M&A research platforms like Tablestat that track deal terms structurally rather than just headline valuations have been registering the divergence; the press releases have not.

The asymmetric outcome is that late-stage SaaS companies will stay private for longer than they planned, with implications for employee liquidity, secondary sale activity, and the eventual public-market reception when they do list. The first SaaS IPO that prices well after the current correction will reset the comparable set for the rest of the cohort, in either direction. That first IPO will probably not happen in 2026, will probably happen in 2027, and will probably involve a SaaS company that has visibly successfully migrated from per-seat pricing to a hybrid model and demonstrated sustained net retention through the migration. The companies positioning for that scenario right now are visible in the trade press; the companies positioning for it but not telegraphing that fact are the more interesting ones to watch.

Table III — SaaS Categories Ranked by Structural Exposure to the AI Transition
SaaS Category	Exposure Level	Why
Customer support & help desk	Very High	The canonical agentic-AI displacement target; high seat count, repetitive workflow
Sales engagement & outbound	Very High	AI agents already automate prospecting, drafting, follow-up at scale
Marketing automation	High	Content generation and personalisation increasingly handled by general-purpose AI
Generic CRM	High	Klarna case proved internal alternatives are viable; pressure on per-seat model
Project management & collaboration	Medium	Workflow lock-in is real; but task automation is encroaching
HR & payroll	Medium	Compliance complexity is the moat; AI augments rather than replaces
Accounting & finance ops	Medium	Regulatory audit trail requirements limit pure AI replacement
Vertical SaaS (industry-specific)	Lower	Domain expertise embedded in workflows that general AI cannot replicate
Cybersecurity & identity	Lower	Compliance, audit, and adversarial complexity protect incumbents
Developer tools & observability	Lower	Deep integration with code and runtime; AI-augmented but not displaced
Database & infrastructure SaaS	Lowest	The plumbing under the AI-native generation itself; benefits from the wave

Exposure scales inversely with workflow complexity, regulatory friction, and infrastructure depth. The most exposed categories are exactly the ones AI agents are best at automating; the least exposed are the ones AI itself depends on.

§ 06 · What The Trade Press Missed

Three observations that did not make the headlines.

The first is the difference between category death and category transition. The trade press has framed the SaaSpocalypse as the death of SaaS as a business model. That framing is wrong in a specific way: SaaS as a category is not dying, but the per-seat-priced, recurring-subscription, growth-at-all-costs version of SaaS that defined the 2015 to 2022 era is genuinely transitioning into something else. The vendors that successfully navigate the transition will continue to be SaaS companies in the broad sense — subscription-based software businesses with recurring revenue. They will not look like the SaaS companies that listed in 2018, and the public market valuations they will eventually command will reflect a different margin profile, a different growth profile, and a different defensibility story. Calling that the death of SaaS is convenient for headlines and inaccurate for procurement decisions.

The second is the asymmetry between visible disruption and invisible defensibility. The most visible AI disruption is happening at the surface layer of SaaS — customer support agents, marketing copy generation, sales outreach automation — precisely because that is where the user experience is most easily replicated by an AI agent with a chat interface. The least visible parts of the SaaS stack — data architecture, integration depth, workflow lock-in, compliance posture, audit trails, identity management — are where the actual defensibility lives, and they are remarkably hard to replicate from scratch. Companies focusing only on the visible layer of vendor selection are positioning themselves to commit to AI-native alternatives whose hidden infrastructure is materially less mature than the SaaS incumbents they are replacing. The mismatch will produce a wave of disillusionment in 2027 as the early replacements run into compliance, audit, and integration problems that the vendor pitches did not flag.

The third is the nature of the competitive question. The SaaSpocalypse trade press has framed this as SaaS-versus-AI-native, with the implicit assumption that the two camps are fixed and the contest will produce a winner. The actual competitive picture is more complicated. The most interesting vendors in 2026 are SaaS incumbents that have credibly migrated key product surfaces to AI-native architecture without abandoning their SaaS economics, and AI-native challengers that have credibly built the unglamorous infrastructure layer underneath their headline products. The winners of the next phase will be hybrid by construction, and the SaaS-versus-AI-native dichotomy that defines current commentary will look as quaint in five years as the cloud-versus-on-premises dichotomy looks today. The wave is not erasing the previous category; it is reshaping it into the next one.

— Reader Questions —

Twenty questions on the SaaSpocalypse, answered plainly.

What is the SaaSpocalypse?

An informal label for the public market correction in software-and-services stocks that began in early 2026, in which roughly a trillion dollars of aggregate market value was erased from listed SaaS companies over a period of weeks. The correction was triggered by a series of AI agent product launches that called into question the durability of per-seat SaaS pricing, but it reflects deeper structural concerns about valuation multiples, growth deceleration, and changing customer behaviour.

Is SaaS actually dying?

No. SaaS as a category is transitioning, not dying. The per-seat, recurring-subscription, growth-at-all-costs version of SaaS that defined 2015 to 2022 is genuinely under pressure, but subscription-based software businesses with recurring revenue will continue to exist. They will look different — different margin profile, different pricing model, different defensibility story — than the SaaS companies that listed at peak valuations.

Why is per-seat pricing under threat?

Because AI agents can now do work that previously required multiple human users. A vendor selling at $150 per agent per month sees its revenue from a customer drop sharply when that customer replaces eight of ten support agents with an AI system. The contract is typically protected for the duration of its term, but the renewal will be renegotiated at materially lower seat counts or restructured into consumption- or outcome-based pricing.

What is outcome-based pricing?

A pricing model where the vendor charges based on a measurable business outcome — resolved support tickets, completed transactions, generated leads — rather than per user or per unit of consumption. Sierra, the AI customer-support startup founded by former Salesforce CEO Bret Taylor, has demonstrated that outcome-based pricing can scale, reaching $100 million in annual recurring revenue in under two years on this model.

How big was the actual market cap loss?

Reporting in early 2026 cited drawdowns approaching one trillion dollars in aggregate across listed software-and-services stocks over a period of weeks, with multiple individual sell-off waves of several hundred billion. The exact total varies by which stocks are included in the basket and which window is measured, but the order of magnitude is consistent across observers.

Was the sell-off justified?

Partly. SaaS valuations had been visibly stretched relative to the higher cost of capital that emerged after the zero-interest-rate era ended. The AI agent narrative provided a catalyst for a repricing that was overdue on its own terms. However, the speed and uniformity of the sell-off suggest the market overshot the underlying structural shift — a common pattern when sentiment moves faster than operating reality.

Is the European SaaS picture different from the US picture?

Materially yes. European SaaS companies tend to operate in narrower verticals, with closer customer relationships, tighter capital efficiency, and less exposure to the growth-at-all-costs valuation inflation that characterised the US zero-interest-rate era. European SaaS has been caught in the downdraft as a sympathy move, but the underlying fundamentals are different enough that the correction reads more as imported sentiment than as a referendum on European SaaS economics.

What was the Klarna-Salesforce moment?

In late 2024, Klarna announced it had replaced Salesforce’s flagship CRM product with an internally built AI-powered alternative. The announcement became the canonical reference point for the build-versus-buy shift, demonstrating that a sophisticated enterprise customer could plausibly substitute internal AI tooling for established SaaS products. The case is more nuanced than the headline suggests, but the symbolic weight has been substantial.

Are AI-native companies actually more profitable than SaaS?

No. The AI-native generation typically operates at gross margins of 50 to 70 per cent, materially lower than the 70 to 90 per cent margins that defined the SaaS era. The lower margin profile is a structural consequence of inference costs scaling with usage rather than a temporary state. Public market valuations of AI-native peers that price them on SaaS-comparable metrics may not be sustainable as growth normalises.

Why are SaaS IPOs frozen?

Because the receiving public market is too volatile and the comparable valuations are too unstable. Late-stage private SaaS companies that would have been the natural 2025 to 2026 IPO class are visibly hesitating. The IPO window is not closed because there is no demand to list; it is closed because the marginal benefit of going public into the current sentiment is outweighed by the marginal risk. Most observers expect the freeze to persist into 2027.

Which SaaS categories are most exposed to AI displacement?

Customer support, sales engagement, and marketing automation are the most exposed because their workflows are repetitive, the user interface is replicable by AI agents, and the per-seat counts are high. Generic CRM is exposed by the build-your-own threat. Vertical SaaS, cybersecurity, developer tools, and infrastructure SaaS are materially less exposed because the workflows are deeper, the compliance moats are real, or the categories actually benefit from the AI wave.

Will SaaS companies survive by adding AI features to their existing products?

Mostly not. The piece that the trade press has flagged correctly is that bolting AI onto an existing SaaS product does not address the fundamental structural pressure on per-seat pricing or the threat of AI-native challengers building from scratch. SaaS companies will need to do harder work — migrating pricing models, rebuilding internal architecture, sometimes cannibalising their existing revenue — to compete in the next phase.

What is FOBO investing?

A label coined by an analyst in early 2026: the fear of becoming obsolete. It describes the institutional investor reaction to the visible AI disruption of established SaaS, where stocks sell off on every related AI product launch on the assumption that the underlying SaaS business is structurally threatened. FOBO is best understood as a category-level correction in sentiment rather than an evidence-based reassessment of individual companies.

How are AI infrastructure costs affecting the picture?

Significantly, but invisibly to public investors. AI-native companies running embedded inference at scale are spending six figures per month or more on third-party model APIs. The cost is structurally part of the unit economics in a way that hosting costs were not for classic SaaS. A market for recovering value from unused cloud and inference allocations has emerged in response, helping companies stabilise effective compute costs — but the underlying margin pressure is real and persistent.

What does the build-versus-buy shift actually mean?

It means that customers can now plausibly choose to build internal AI-powered alternatives to SaaS products rather than license them, where ten years ago the build option was prohibitively expensive. The shift is real but bounded — many customers will still prefer to buy because of compliance, support, integration, accountability, and outsourced liability. The shift increases buyer leverage at renewal more than it produces actual build outcomes.

Are M&A activity and consolidation likely?

Yes, and accelerating. The combination of frozen IPO markets, structural revenue pressure on incumbents, and well-funded AI-native challengers creates near-perfect conditions for strategic and private equity acquisition activity in late-stage SaaS. Expect continued roll-up activity in vertical SaaS, defensive acquisitions of AI-native challengers by SaaS incumbents, and a gradual privatisation of weaker public SaaS names through PE buyouts.

What should a SaaS founder do right now?

Audit the pricing model honestly. If the business is per-seat priced, model what happens at renewal when customers reduce seats by 30 to 60 per cent and stress-test the assumptions in the financial plan. Identify which parts of the product create defensibility through workflow, data, or compliance and double down there. Resist the impulse to chase AI features that do not address the underlying structural exposure. Capital efficiency matters more in 2026 than it did in 2022.

What should a SaaS investor do right now?

Distinguish carefully between SaaS categories with genuine AI exposure and SaaS categories where the AI threat is mostly narrative. Underwrite gross retention and unit economics rather than ARR growth. Take the SaaSpocalypse narrative seriously enough to reprice at-risk holdings; take it sceptically enough to recognise that the bulk of well-built SaaS businesses will adapt rather than die. Watch for the first successful migration to a hybrid pricing model from a major incumbent — that will reset the comparable set.

Could the public market reverse on SaaS?

Plausibly, on a delay. The same dynamic that produced the overshooting sell-off in early 2026 typically produces an overshooting recovery once a few high-profile names demonstrate that the structural fears were exaggerated. The catalyst will probably be a major SaaS incumbent reporting strong renewal performance under the new pricing structure, or a successful IPO from an AI-native company at multiples that look reasonable on its own economics. Neither is imminent; both are plausible within twelve to eighteen months.

What is the most likely end state for the SaaS industry?

A hybrid landscape in which mature SaaS vendors migrate to consumption- and outcome-based pricing while retaining their compliance, integration, and workflow moats; AI-native challengers mature into more defensible products with proprietary infrastructure or vertical specialisation; and the SaaS-versus-AI-native distinction that defines current commentary loses its operational meaning. The category itself will look more like the financial services industry — a mix of incumbents and challengers operating with similar economics — than like the cloud-versus-on-premises transition of the previous era.

Source · Primary Coverage

Dominic-Madori Davis, SaaS in, SaaS out: Here’s what’s driving the SaaSpocalypse, TechCrunch, March 2026.

Original reporting on the SaaSpocalypse with venture investor commentary on the build-versus-buy shift, per-seat pricing pressure, and IPO window closure: techcrunch.com

— Editor’s Note —

On the durability of the headline.

A reader who arrived at this article looking for a single tidy take on whether SaaS is dead or alive will leave disappointed. The SaaSpocalypse is, in the structural read above, four genuinely distinct stories that the trade press has compressed into one. The pricing-model story is real and accelerating. The public market correction is partly justified and partly overcorrection. The infrastructure cost story is the underreported piece that will shape AI-native economics for years. The IPO freeze is a near-term constraint that will eventually thaw on a timeline most observers are not pricing in. Each of the four stories is interesting on its own; none of them is the SaaS apocalypse the headlines describe.

MMD Newswire is editorially independent. The interpretations, framings, and structural reads in this article are our own, and we have no commercial relationship with any of the public or private companies mentioned. Readers making procurement, investment, or operating decisions on the basis of the SaaSpocalypse narrative should treat this article as a starting framework rather than a substitute for direct due diligence on the specific companies and contracts involved. The category-level read is informative; the company-level read still has to be done one company at a time.

The post Inside the SaaSpocalypse: A Structural Read for 2026 appeared first on Newswire — Business, Technology & Startup Press Releases.