Over the last year, I have spent a huge amount of time speaking to financial analysts, keynoting at industry conferences, and engaging deeply with the venture capital community. Those conversations gave me a clear picture: the market is ripe with opportunity, but only for those who understand the nuances of what is actually happening.

I heard some of the world's leading VCs confidently assume that the 'moat' for legacy ERPs was the inherent stickiness of the system of record. I watched respected financial analysts bundle all 'SaaS' into a single, homogeneous bucket. The instinct is understandable - but it is leading investors and operators to defend the wrong position, and more importantly, to miss the real opportunity.

In early February 2026, roughly $300 billion in SaaS market capitalisation evaporated in a single trading session - triggered, in part, by Anthropic's launch of Claude Cowork, an autonomous AI agent suite capable of managing legal, financial, and HR workflows.² Workflow and orchestration platforms were hit hard. CRM giants saw double-digit declines. The median revenue multiple for public cloud software fell to levels not seen in a decade.

The selloff was indiscriminate - and that is where the opportunity lies.

Morgan Stanley's head of global technology M&A, Wally Cheng, put it well: there has not been a 'thoughtful, detail-oriented approach to sorting through who the winners and losers are.'³ Meanwhile, Scott Galloway called the whole thing 'farcical' and started buying SaaS stocks.⁴ Both are partially right. The truth, as always, is in the detail.

As someone who spends their days building mission-critical AI solutions for some of the world's leading companies, I believe the SaaS-pocalypse is less about destruction and more about differentiation.

The argument I want to make is simple: not all software is equally vulnerable, and the companies that understand which tier they occupy will be the ones that thrive. I see three distinct tiers. The first is software that is primarily an interface layer over a database - vulnerable, and in many cases already being displaced. The second is software that embeds deep, specialised algorithmic intelligence - not only resilient but actively strengthened by the rise of AI. The third is mission-critical infrastructure where engineering rigour is the product itself - untouchable, and increasingly valuable. The rest of this article unpacks each tier, the evidence behind it, and what it means for anyone building, buying, or investing in software.

What Is the SaaS-pocalypse?

The SaaS-pocalypse is a fundamental reshaping of software-as-a-service business models, driven by the rapid commoditisation of code and the rise of autonomous AI agents that sell completed work rather than software seats.

The narrative traces back to December 2024, when Satya Nadella appeared on the BG2 podcast and made an observation that the venture capital community acted on decisively. He described SaaS applications as 'essentially CRUD databases with a bunch of business logic layered on top' and predicted they would 'collapse' in the agent era.¹ It was a bold framing - and one that captured a genuine structural shift, even if the full picture is more nuanced than a single soundbite can convey. AI startups captured $202 billion globally in 2025, nearly 50% of all venture capital, up from 34% the year before.⁵ In the US, AI companies took 64% of VC dollars while comprising only 36% of funded startups.

The strategic reframing is captured in a single statistic from a16z's Alex Rampell: the global software market generates $300 billion in annual revenue, but the US labour market alone represents $13 trillion.⁶ The smartest VCs are not investing in better tools. They are investing in outcomes - replacing the labour that tools were supposed to assist.

Sequoia Capital's Julien Bek crystallised this in one of the era's defining essays: for every dollar spent on software, six are spent on services - and the next legendary company will not sell the tool, it will deliver the outcome the tool was supposed to enable.⁷

This is a massive opportunity for those who can see it clearly. We are leaving an era where software was hard to build and hard to leave, and entering one where value must be earned on merit, not protected by friction. To understand where the opportunity lies, we have to stop treating all software equally and break it down into three distinct tiers.

Tier 1: Interface-Layer Software - Evolve or Be Replaced

I do not wish to trivialise the engineering involved, but a massive portion of today's SaaS - solutions handling HR, Finance, Customer Support, and CRMs - is fundamentally a user interface layered over a database.

Last year, the prevailing bet was that while AI-first startups could move quickly, the sheer friction of migrating a 'system of record' would give slow-moving incumbents the time they needed to catch up. That bet is not paying off.

AI-powered migration tools can now compress months of migration work into days, sometimes for free. As Jason Lemkin of SaaStr put it: switching a marketing automation platform used to be a six-month project; now you can take a prompt from one AI vendor and hand it to another in an afternoon.⁸ Andreessen Horowitz has explicitly identified switching costs as 'the one moat that really is going to change.'⁹ Model Context Protocol is already beginning to weaken data moats by creating standardised interfaces between AI agents and software systems.

But agents will not just make migrating systems of record frictionless. They are beginning to replace the interface altogether. Why log into a clunky dashboard to file an expense, update a client record, or pull a report when a dynamic AI agent can execute the task in the background via a simple conversation?

The companies seizing this opportunity are growing at extraordinary speed. Sierra AI - founded by former Salesforce co-CEO Bret Taylor - reports hitting $100 million ARR in just 21 months, raising $350 million at a $10 billion valuation.¹⁰ Crucially, it charges per completed resolution, not per seat. Companies like Decagon report that clients have reduced support headcount by as much as 80%.¹¹ In sales, 11x.ai's autonomous SDR agents handle outbound prospecting 24/7 - directly challenging traditional seat-based models.

The pricing model shift tells the same story. Seat-based pricing adoption dropped from 21% to 15% in just twelve months. Hybrid usage-based models surged from 27% to 41%.¹² We are seeing this transition first-hand at WPP, where WPP Open Pro operates on a user-plus-usage model - reflecting the broader market recognition that value should scale with outcomes delivered, not headcount served.

It is worth noting a cautionary nuance. Klarna's well-publicised ditching of roughly 1,200 SaaS applications became the poster child for this shift - but CEO Sebastian Siemiatkowski later clarified that Klarna replaced SaaS with a mix of in-house tools and alternative SaaS, not purely with AI.¹³ The lesson is not that replacing SaaS is straightforward. It is that the switching costs that once protected these companies are evaporating - and the incumbents that recognise this fastest will be the ones that reinvent themselves around outcomes rather than access.

Tier 2: Deep Algorithmic Intelligence - The Compounding Advantage

There are SaaS solutions that are infinitely more sophisticated than an interface layer. 'Sticky' SaaS embeds highly technical, specialised intelligence that is incredibly difficult for an LLM - or an LLM-powered agent - to replicate. This is what I think of as the second part of the equation Nadella began to spell out, and it is just as important: not all software is a CRUD database, and understanding that distinction is where the greatest opportunity for compounding advantage exists.

Take last-mile delivery. These solutions utilise bleeding-edge Operations Research to calculate time slots and optimise schedules, solving for highly complex objectives with non-linear constraints. They must calculate incredibly accurate routes in milliseconds, leveraging machine learning to predict 'time-at-door,' analyse driver patterns, find likely parking spaces, and adapt to a constantly changing environment of road closures, adverse weather, and erratic customer behaviour.

This requires a beautiful orchestration of a plethora of sophisticated algorithms, operating in perfect harmony. It is one of the reasons why Satalia - WPP's deep AI research and optimisation lab - arguably has the world's leading last-mile solutions. It is also why supply chain platforms like Kinaxis - whose SaaS revenue grew 17% year-over-year to $362 million while the rest of the sector contracted - are accelerating through the SaaS-pocalypse.¹⁴ Their Maestro Agents delivered a 10x productivity boost for one pharmaceutical customer. Blue Yonder's SaaS revenue grew 37%, with Panasonic planning an IPO valuing it above $7.7 billion.¹⁵ You cannot vibe-code any of this - and that is precisely the point.

Another prime example is WPP Open. I am deeply worried that brands are being told by tech consultancies that they can easily build their own marketing stacks. They can build something, sure - but it will not be sustainable or differentiated. As Chief AI Officer, I am responsible for ensuring the intelligence layer inside WPP Open is highly differentiated. That requires designing, deploying, and innovating across seven distinct parts of our supply chain: dynamic segment analysis, synthetic audience representation, creativity, brand-perfect production-grade content creation, performance prediction, channel optimisation, and macro and micro moments identification.

Here is the hint the consultancies will not give you: you only use LLMs to solve a small fraction of these. And even those that do require deep AI expertise to ensure they are differentiated. True differentiation requires multi-disciplinary algorithmic mastery that an API wrapper simply cannot fake.

The market is already rewarding this kind of depth. While interface-layer SaaS stocks cratered, cybersecurity platforms thrived - CrowdStrike posted $4.81 billion in FY2026 revenue (up 22%), because every new AI agent creates new attack vectors, making security more essential, not less.¹⁶ Veeva Systems maintained 80% market share in life sciences CRM, growing to $3.17 billion in revenue, protected by FDA validation requirements and patient safety regulations that no amount of vibe coding can bypass.¹⁷ Datadog's AI-native revenue grew 253% year-over-year, because more AI means more infrastructure to monitor.¹⁸

Box CEO Aaron Levie articulated the opportunity well: you want some sort of 'church and state' between the deterministic side of your software and the non-deterministic side.¹⁹ The SaaS that thrives will be the deterministic execution layer that AI agents operate on top of. LLMs interpret human intent; deterministic systems execute the actual work. If you are building the execution layer, you are not competing with AI - you are becoming essential to it.

Tier 3: Mission-Critical Infrastructure - Where Rigour Is the Product

I would not recommend vibe-coding software to run a nuclear power station.

It might sound like an extreme example, but the evidence is now abundant. SaaStr founder Jason Lemkin spent nine days building a production app with AI coding tools. The AI agent deleted a live production database during a code freeze, fabricated 4,000 fake users, generated false performance reports, and violated explicit instructions over 11 times - including directives written in all caps.²⁰ Veracode found 45% of AI-generated code contains security vulnerabilities.²¹

A randomised controlled trial by METR studying experienced developers on real-world tasks found that AI actually made them 19% slower - despite the developers believing they were 20% faster.²² It is worth noting that a February 2026 follow-up by the same researchers found early signs of improvement among returning developers, though the authors acknowledged significant selection effects that made the newer data less reliable. The picture is evolving rapidly - but the direction of travel reinforces a crucial point: the gap between what AI coding tools promise and what they deliver in complex, real-world systems remains material.

Google's DORA report found that increased AI adoption correlated with a 1.5% estimated decrease in delivery throughput and a 7.2% estimated reduction in delivery stability.²³ Their 2025 follow-up reversed the throughput finding - AI now shows a positive relationship with speed - but the negative correlation with stability persisted. In other words, AI is getting faster at producing code, but the code is not getting more reliable. For mission-critical systems, that distinction is everything.

Even Andrej Karpathy, who coined 'vibe coding,' has walked the term back, declaring it 'passe' and admitting he hand-coded his own serious project because AI agents were 'net unhelpful.'²⁴

Any system that is truly mission-critical - systems that have a material impact on the physical world and living things - requires engineering rigour that is incomprehensibly beyond the capability of current AI. Even if an AI could perfectly generate the code for these systems tomorrow, it would take years of rigorous testing to legally and functionally trust that they will not fail.

For companies operating in this tier, the SaaS-pocalypse is not a threat - it is a competitive moat that deepens with every headline about AI-generated failures. Cyber-physical systems and high-stakes infrastructure are safe from displacement for the foreseeable future, and the premium the market places on proven reliability will only increase as AI-generated code proliferates elsewhere.

Building the Right Moat

If an interface layer will not save you, what will? The companies that emerge strongest from the SaaS-pocalypse will be those that invest in the moats that compound, not the ones that erode.

Proprietary Data. It is data that makes the AI smart. If you have useful, highly contextual data that your competitors do not have - and that data cannot be easily inferred or scraped by AI - you have a moat. VLex quintupled its revenue by adding AI to 26 years of digitised legal records.²⁵ The data was the asset; the AI was the unlock. The strategic question for every SaaS company is: what data are we uniquely positioned to accumulate, and how do we make AI work harder because of it?

Algorithmic Talent. If your software requires complex algorithmic transformation or decision-making, you have an advantage - but only if you can attract and retain the elite mathematical talent required to build and continuously improve those algorithms. A PE firm recently prototyped a competing vertical SaaS product in two weeks that was reportedly 'cleaner and faster' than the original. Feature-based differentiation is collapsing; only deep algorithmic complexity endures. The companies investing in this talent now will own the next decade.

Business Model Velocity. In a world where code is increasingly commoditised, the ability to capture markets at speed becomes decisive. Y Combinator's latest batch includes companies reaching $10 million in revenue with teams of fewer than ten people - and roughly 25% of their codebases are 95% AI-generated.²⁶ Cursor crossed $2 billion ARR, doubling in 90 days, making it arguably the fastest-growing software company in history.²⁷ Whether through open-source adoption, aggressive pricing, or relentless execution, speed is becoming its own moat.

But there is a final moat, and it is arguably the most enduring.

In a world where software code can be created instantly, where talent is highly mobile, and where executing business models is increasingly formulaic, how do you stand out?

Creative connection. This is the moat I find most interesting - and not simply because I work in the marketing industry. It is because every other moat I have described has a half-life. Data can be replicated over time. Talent can be poached. Speed advantages are temporary by definition. But the depth of creative connection between a brand and its audience - the trust, the emotional resonance, the sense that this company understands me - compounds in a way that no algorithm can shortcut. As the barriers to building software drop to zero, the premium on human connection, brand trust, and creative communication goes through the roof. When everything else is commoditised, creativity is the only moat left.

The Real Opportunity

The SaaS-pocalypse is not the end of software - Forrester still projects global SaaS spending rising from $318 billion in 2025 to $576 billion by 2029.²⁸ What is ending is a specific archetype: narrow, UI-dependent workflow tools sold on per-seat pricing with switching costs as the primary moat. That is not a tragedy. It is a correction - and a long overdue one.

Here is the insight that even the disruptors should heed: the very dynamics reshaping interface-layer SaaS apply to AI agent companies too. At $100 million ARR, AI agent startups reportedly face 82% gross retention versus 92%+ for traditional SaaS.²⁹ Prompts are portable. Configurations are transferable. The disruption is recursive - which means the opportunity to build durable, defensible platforms has never been greater for those who understand what actually holds value.

The companies best positioned are those converting from tool vendors into outcome platforms - using proprietary data, regulatory moats, and deep algorithmic intelligence to become the essential execution layer beneath an agentic future. The winners will not sell you software. They will deliver the result the software was supposed to help you achieve.

The SaaS-pocalypse is not the end of software. It is the beginning of what software was always supposed to be: not a tool you log into, but an outcome you can rely on.

Notes and References

¹  Nadella, S. Interview on BG2 Podcast with Bill Gurley and Brad Gerstner, December 2024. Nadella described SaaS applications as 'essentially CRUD databases with a bunch of business logic' and predicted they would 'collapse' in the agent era.

²  UBS strategist estimates and multiple financial reporting outlets confirmed approximately $285-300 billion in software market value lost in a single trading session, early February 2026. See also Inc.com reporting, February 2026.

³  Cheng, W. Quoted in Reuters, reported by Inc.com, 11 February 2026. Cheng is Morgan Stanley's head of global technology M&A.

⁴  Galloway, S. Interview with Inc.com, 26 February 2026. Galloway called the idea that companies would cancel enterprise subscriptions in favour of prompts 'farcical' and disclosed he was buying SaaS stocks.

⁵  PitchBook and CB Insights reporting, 2025. AI startups captured approximately $202 billion globally, representing nearly 50% of all venture capital. In the US, AI companies took 64% of VC dollars while comprising 36% of funded startups.

⁶  Rampell, A. (a16z). Widely cited framing: global software market ~$300 billion annual revenue versus US labour market ~$13 trillion.

⁷  Bek, J. 'In the Service of Software.' Sequoia Capital, 2025. sequoiacap.com/article/in-the-service-of-software/

⁸  Lemkin, J. Commentary on SaaStr, 2025-2026. Multiple posts and conference remarks on AI migration tools, switching costs, and seat-based pricing disruption.

⁹  Andreessen Horowitz. Multiple partners have identified switching costs as 'the one moat that really is going to change' in the agent era. 2025-2026.

¹⁰  Sierra AI. '$100M ARR in 7 Quarters.' Sierra blog, 21 November 2025. sierra.ai/blog/100m-arr. Raised $350 million at $10 billion valuation, September 2025, led by Greenoaks Capital. Revenue figure as reported by the company.

¹¹  Decagon and 11x.ai metrics sourced from company disclosures and industry reporting, 2025-2026. Performance figures as reported by the companies.

¹²  OpenView Partners. 'SaaS Pricing and Packaging Benchmarks.' 2025-2026. Seat-based adoption declined from 21% to 15%; hybrid usage-based surged from 27% to 41%.

¹³  Klarna. CEO Sebastian Siemiatkowski clarified in subsequent interviews that Klarna replaced SaaS with a mix of in-house tools and alternative SaaS, not purely with AI.

¹⁴  Kinaxis FY2025 earnings. SaaS revenue grew 17% year-over-year to $362 million.

¹⁵  Blue Yonder SaaS revenue growth of 37% reported in FY2025 financials. Panasonic IPO valuation estimates above $7.7 billion.

¹⁶  CrowdStrike FY2026 earnings. Revenue of $4.81 billion, up 22% year-over-year.

¹⁷  Veeva Systems FY2026 earnings. Revenue grew to $3.17 billion with approximately 80% life sciences CRM market share.

¹⁸  Datadog AI-native revenue growth of 253% year-over-year reported in quarterly earnings, 2025.

¹⁹  Levie, A. (Box CEO). Remarks on 'church and state' between deterministic and non-deterministic software. CNBC interview, early 2026.

²⁰  Lemkin, J. 'I Spent 9 Days Vibe-Coding a Production App. Here is What Happened.' SaaStr, July 2025.

²¹  Veracode. '2025 State of Software Security.' Found 45% of AI-generated code contains security vulnerabilities.

²²  Becker, J., Rush, N., Barnes, E., Rein, D. 'Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.' METR, July 2025. arXiv:2507.09089. Randomised controlled trial with 16 experienced developers across 246 tasks. A February 2026 METR follow-up found early signs of improvement among returning developers, though the authors acknowledged significant selection effects making the newer data less reliable.

²³  Google. 'Accelerate State of DevOps Report.' DORA, 2024. Found that every 25% increase in AI adoption correlated with a 1.5% estimated decrease in delivery throughput and a 7.2% estimated reduction in delivery stability. The 2025 DORA report found AI's relationship with throughput had reversed to positive, though its negative correlation with stability persisted.

²⁴  Karpathy, A. Coined 'vibe coding' on X, 2 February 2025. Declared it 'passe' in a February 2026 post. Separately, in October 2025, Karpathy disclosed that his Nanochat project was 'basically entirely hand-written' because AI agents were 'net unhelpful.'

²⁵  VLex. Revenue growth reported in company disclosures, attributed to adding AI to 26 years of digitised legal records.

²⁶  Y Combinator Winter 2025 batch. Approximately 25% of companies had codebases 95% AI-generated. Reported March 2025.

²⁷  Cursor. Reported crossing $2 billion ARR, with revenue doubling in approximately 90 days, early 2026. Multiple financial outlets.

²⁸  Forrester. Global SaaS spending projected to rise from $318 billion in 2025 to $576 billion by 2029.

²⁹  AI agent gross retention rates versus traditional SaaS sourced from industry benchmarking analyses, 2025-2026. The 82% gross retention figure for AI agent startups at $100 million ARR has circulated in VC community analyses.