Is AI Killing SaaS? What I Tell Clients Who Ask
No. What is dying is the seat, not the software. And those are very different problems.
I have had this argument twice in the last month with two prospects who looked at the same evidence and reached opposite conclusions. One was about to buy a point solution because it had AI in the name. The other will not buy any software at all, because he is convinced he can build it faster himself. Both of them were answering a question that stopped mattering.
I have seen this movie before
I spent about twenty years selling and delivering sales methodology. Master certified in Strategic Selling, Large Account Management, SPIN, Conceptual Selling and Professional Selling Skills. Which means I was in the room during the last time a technology shift got confused for a strategy.
Companies bought Salesforce. They bought HubSpot. And a lot of them genuinely believed that buying the CRM had given them a sales process. It had not. Those are two different things, and the confusion cost a generation of sales organizations real money. The tool captured what happened. It never told anybody what to do next.
We are doing it again, faster.
What actually happened in February
Roughly a trillion dollars of enterprise software value disappeared in about a week. The Doomsters jumped on it with all the zeal of a vulture finding fresh roadkill. AGENTS KILL SaaS. SEATS ARE DEAD. Everyone builds their own.
Then came the correction. NVIDIA's Jensen Huang called the software-is-dead narrative "the most illogical thing in the world," and he was right. Systems of record did not go anywhere. Avenir's January 2026 research found 63% of enterprise buyers expect their existing software vendors to benefit from generative AI, against only 8% who expect them to lose.
Good correction, well supported by data, and about as newsworthy as rain falling in the middle of the ocean. Nobody covers the correction. The panic gets the clicks.
So here is what I think actually happened, having watched a few of these cycles now. The selloff was not a verdict on software. It was a verdict on how software gets billed.
For twenty years, one seat roughly equaled one person's output. Charging per person tracked the value delivered, and everybody understood the deal. AI broke that link. When one person with AI does the work of five, the customer needs one license instead of five. Same output, same value to the business, eighty percent less revenue to the vendor. The meter stopped measuring anything real.
The people selling seats already know it. A Cruxy survey of 300 SaaS CEOs in April 2026 found 97% plan to retire seat-based pricing within two years, while 94% said seat pricing still matches their product's value today. Read that twice. Both cannot be true.
It reminds me of the old Looney Tunes bit where Bugs talks Elmer into such a knot that Elmer ends up politely negotiating the terms of his own shooting. That is a room full of CEOs saying the model is finished and the model is fine, in the same breath, because they know the meter is broken and they have not built its replacement.
Meanwhile the other meter broke too. Model inference prices fell more than 280-fold between November 2022 and October 2024, from about twenty dollars per million tokens to roughly seven cents. In plain terms, having a model read and reason across eight or nine novels' worth of text went from twenty dollars to pocket change. The model is a purchased input now. It is priced like bandwidth, not like an advantage.
So two reliable places to make money disappeared at the same time. Charging per person stopped working, because headcount no longer predicts output. Charging for access to AI stopped working, because AI got about as cheap as bandwidth, and nobody has built a business selling long distance in thirty years.
Which leaves exactly three things that did not get cheap. Your data. Your workflow. And whether your people can actually operate.
Four questions I ask about any AI product, including my own
- Could a competent person get most of this by pasting a prompt into a model they already pay for? If yes, it is not a product. It is a feature that has not shipped.
- Does it hold data that compounds? Defensibility lives in the data a product generates through its own use and the workflow it becomes the official record for. A tool that touches your data without becoming the record for anything has neither.
- What breaks if you rip it out tomorrow? If nothing breaks, you did not buy a system. You bought a subscription to a convenience.
- Can one manager cancel it alone? Tools bought on discretionary budget die at the first budget review. Workflows the whole team runs on do not.
Buy, build, or standardize
| Buy the point solution | Build it yourself | Standardize the record, build the fluency | |
|---|---|---|---|
| What it costs | Per seat, monthly | Engineering time, forever | Licence plus a capability project |
| What you own at the end | Access until renewal | Code and a maintenance obligation | Your data, your workflow, your people's skill |
| Where it breaks | The model provider ships it natively | The person who built it leaves | Nothing breaks, but people have to change how they work |
| Best fit | A narrow, genuinely bounded task | A real proprietary process nobody sells | An organization sitting on systems nobody uses well |
I run the third one, so weigh the table accordingly. The other two are right in the right situation.
How to choose
- Narrow and bounded, buy it. Transcription, scheduling, document conversion. Commodity job, commodity tool. Building your own is vanity.
- Genuinely proprietary and nobody sells it, build it. But be honest about what building means now. CNBC's Deirdre Bosa recreated a working version of Monday.com, Gmail integrations included, in about an hour. If a competitor can rebuild your tool in a weekend, it is not an asset. It is a maintenance schedule with a single point of failure.
- Systems already in place and nobody uses them well, standardize and build the fluency. This is the most common situation I walk into and the one people diagnose wrong most often, because it does not look like a software problem. The CRM is there. The methodology was bought and trained. The tooling is fine. Nobody can operate any of it at the level the tools assume.
That is the work we do at SalesXecution. Not another platform. We standardize the team on one model, instrument the sales motion they already own, and build enough AI fluency that people can actually run it.
Where I would tell you not to bother
- If your system of record is genuinely broken, fix that first. Reasoning over bad data produces worse decisions faster. (And of course we can help with that)
- If you have no process at all, do not start with AI. You will automate improvisation. Get it on paper first, then instrument it. (Again, we have a proven approach to help you address this situation)
- If you are under fifteen people and everybody already knows every deal, the coordination problem does not exist for you yet. Go sell. (And, of course again, we can help you leverage AI and your existing tools to enhance and accelerate deals even if everyone is aware of every deal.)
Frequently asked questions
Is per-seat pricing actually going away?
Not entirely and not fast. Kyle Poyar's 2026 State of B2B Monetization survey of more than 230 software companies found 37% now run hybrid as their primary structure, and HubSpot, Figma, Adobe, Cursor and Lovable have all added credit models without dropping seats. Hybrid is winning, not pure outcome pricing.
If models are cheap, why not build everything?
Because cheap building is cheap for your competitors too. Whatever you build is replicable in days. The durable assets are the data your operation generates, the workflow that becomes official, and the judgment of the people running it. None of those come from writing code.
Should I cancel my CRM?
No, and this is where the narrative does real damage. AI makes your system of record more valuable, because the model has to get facts from somewhere. What should change is that you stop paying for seats nobody uses and start caring whether the record is complete enough to reason over.
How do I tell an AI product from an AI feature?
Ask the vendor what their tool does that you could not get by pasting a well-written prompt into a model you already pay for. A good one answers directly. A weak one starts talking about the roadmap.
What should I actually do this quarter?
Put the whole commercial team on one model, paid tier, shared workspace. Audit which seats are genuinely used. Then find the process everybody knows but nobody runs, and instrument that one. Cheaper than most point solutions, and it compounds.
Steve Cadley is Managing Partner of SalesXecution in Alpharetta, Georgia. Master certified in Miller Heiman Strategic Selling, Large Account Management, SPIN, Conceptual Selling and Professional Selling Skills, he now works with mid-market and private-equity-backed organizations on the AI fluency that makes their existing methodology, process and technology investments operational. Anthropic partner and HubSpot Solutions Partner.