Sales & Automation
AI doesn't just want our jobs. It's coming for our SaaS too
July 16, 2026
6-minute read
The "AI is killing / replacing / about to replace (insert preferred verb) SaaS" narrative sounds, at first read, like pure marketing bullshit (fine, on a second read too). But over the past year or two, and continuing right now, some genuinely exciting things have been happening that fire up the imagination. To borrow a classic line: this bit of technological progress is visible from the Moon.
On the SWE-bench Verified benchmark, which measures how well models solve real-world GitHub issues, the performance of leading models has grown from roughly 4 to 5%in late 2023 to 93.9% in 2026: a more than 20-fold improvement in barely two and a half years. METR's "time horizon" metric (which measures the duration of tasks an AI agent can complete autonomously with a 50% success rate) doubled every seven months between 2019 and 2024. Since 2024 it has been doubling every three to four months, meaning the acceleration of AI's progress is itself accelerating. (METR, March 2026)
That is the pace of technological progress we are dealing with. Extrapolate it a few years forward, multiply it by the computing capacity coming online amid the current AI datacentre building boom, and legacy software vendors and SaaS company owners have every reason to be seriously concerned about the software development capacity about to hit the market.
In early 2026, an Anthropic product update (which really just drew attention to this area of AI progress) wiped roughly $300 billion off the software industry's market value in two days. Salesforce, ServiceNow, Adobe and Workday each fell 7 to 11% in a single day, and the sector's forward P/E dropped from around 39x to around 21x within a few months. The press has since dubbed this story the "SaaSpocalypse ", and it hasn't let up: the IGV software index is 30% below its September peak, and the cumulative market-cap loss is now running at roughly $2 trillion.
What's genuinely interesting is that the market isn't pricing in companies buying less software, and nobody seriously thinks they'll use less of it either. What they can do, though, is pay for the concrete functionality delivered by AI-assisted developers, for greater efficiency and for lower vendor lock-in risk, instead of paying for a rented licence. And that is why the SaaS "recurring revenue" dogma, on which the past decade's high SaaS profits and predictability were built, has suddenly evaporated.
In short, here's the question I ask myself as a business owner: why pay a lot of money to an outside SaaS vendor if, with a one- to two-year ROI (sometimes even within a year), I can already build the technology my company's processes need myself?
The cases uncovered by The Information and summarized by PYMNTS are particularly telling because they include the names of specific companies, specific dollar amounts, and specific dates:
A similar pattern shows up in Retool's 2026 Build vs. Buy survey of 817 enterprise developers and operations professionals: ClickUp's GTM team built six internal AI tools that save $200,000 a year on automation software, while still connecting their Salesforce, Zendesk and Snowflake systems. A company called Harmonic rebuilt a $20,000-a-year tool for itself because it was faster than waiting for a customer-support reply; today it runs 33 internal applications with Salesforce, Gong and Slack integration.
The ERPClaw case points the way: a former Accenture consultant who used to lead SAP implementations for large energy clients (Allegheny Power, E.ON, American Water) single-handedly built a 45-module, open-source ERP system using AI tools. A typical SAP licence costs at least $50,000 a year; ERPClaw runs on a $20-a-month server. The project doesn't promise to replace SAP at enterprise scale (not yet), but it does show that the business model behind software categories that used to carry six- or seven-figure price tags is about to change fast (HackerNoon, 2026)
Aspen Pumps ( (a British air-conditioning parts manufacturer and user of SAP Business ByDesign / SAP Cloud ERP) built 12 automation bots with its partner, including one that extracts data from CAD drawings and automatically produces a bill of materials (BOM). Result: 10,000 hours saved a year in total, with the BOM bot alone saving £25,000 a year by eliminating previous manual-processing errors. (SAP, case study)
Unified Women's Healthcare (a US healthcare network) worked with its partner to redesign and automate its NetSuite environment (script simplification, automation, optimisation), saving more than 1,500 work hours a year. (Rand Group)
Successful switches are documented in the Power BI / Tableau market too: Jean Mandarin, Head of Data & Insights at Matillion, has publicly shown how the team cut reporting-error tickets by 80% by moving away from Tableau to a differently architected, AI-native solution. (Velosio, 2026)
The Swedish fintech's OpenAI-based customer service assistant launched in February 2024 and handled 2.3 million conversations in its first month, equivalent, by the company's own account, to the workload of 700 full-time staff. By the third quarter of 2025, that figure had grown to the "work value" of 853 people and roughly $60 million in annual savings; response time improved by 82%, and its customer satisfaction score (NPS) reached 73 points. (Twig, 2026)
This isn't just an American story: the founder of a Berlin-based cybersecurity startup, DmarcDkim.com, replaced around ten SaaS products with self-hosted, open-source alternatives within a year (Rocket.chat instead of Slack, Twenty instead of HubSpot/Salesforce), partly for cost reasons and partly out of concern for European data sovereignty. The same article profiles the team behind the developer tool Warp, where rebuilding an internal documentation product took only around two days of work and produced a better end product that fit the company's brand more closely. (LeadDev, 2026)
Explosively growing computing capacity, multiplied by exponentially growing AI coding efficiency, raised to the power of the frustration of business owners at the mercy of SaaS and legacy tech companies, equals an AI bubble. Trees don't grow to the sky (only the expectations attached to them do). There are already risks, there will be bottlenecks, human resistance exists and is growing, and state oversight is growing too (quite rightly). We're approaching AI light speed, but we won't reach it in the next two years. In my view.
BUT. And I mean BUTBUTBUTBUT. Just be careful. Just because this is a promising, exciting direction for our business, one with a lot to gain and plenty of problems it can solve, that's no reason to go charging open-mouthed into that particular forest.
There are plenty of risks. In AI usage and in the technology itself. In IT security and data protection. In companies' (AI) readiness actually are, in the quality of their earlier digitalisation, and in how well-developed their processes are. In legal and ethical compliance. In the people involved. And in a thousand other things besides.
In future posts, I'll try to write about these risks too: what they are, how to identify them early, and either prevent them or at least hedge against them.
