Build vs buy: AI is changing the SaaS question

Since the start of this year, I've been building my own tools.

Some of it is small - scripts that save me 10 minutes here and 20 there, content creation workflows, a bit of data analysis. Some of it is bigger - a mini CRM, a presentation builder, a product management workflow, a legal document review automation.

And now I look back at five months of doing this and realise much of this is replacing products I was already paying for, or would have gone looking for a SaaS product to cover.

Not the big strategic bits of the stack, obviously. Nobody's replacing their accounting software in an afternoon. But a fair number of the SaaS tools that sit in the middle of the stack - the productivity ones, the content ones, the ones half the team uses half of the time - I just didn't need to buy them anymore. I built the 20% I actually used, and it fit the way I work rather than the other way round.

So I keep coming back to the same question. If that's true for me, what does it mean for SaaS?

The SaaS compromise

So, what actually was the SaaS deal?

You identify something your business needs. You do the thinking: what we do, how we do it, what tooling would help. And then you run the numbers on building it yourself. Proper software, built for exactly your workflow. Almost always too expensive. So you do the sensible thing - you go out to the market and look for a product that does most of what you need.

It becomes a game of compromise. You find something that covers your must-haves, does a reasonable job on your nice-to-haves, and comes with a load of features you will never use because they were built to serve a different kind of business than yours. You pay for all of them.

Then the product arrives and you change how you work to fit it. Because it has to be generalised. No SaaS product gets built for one company - it gets built for the median of an addressable market. So people retrain, workflows bend, and the bits that don't quite fit get tolerated or worked around.

For 20 years this has been the best deal on the table. It wasn't great, but it was rational.

Both halves of the deal are breaking

Here's what's changed. Both sides of that deal - the cost side and the adoption side - don't hold the way they used to.

The bespoke build cost is falling as quickly as new AI models are coming out. In our own work at Kyan, what used to cost £200k to build is now costing half as much. It's because the same team, working with modern AI tooling, does in a few weeks what used to take months. And you can get there with fewer people, which in itself gives you a more efficient team because there's less communication overhead.

Meanwhile licence fees haven't been shrinking. The two lines crossed some time in the last year, and for a lot of mid-complexity internal tools, the build-or-buy answer looks different to the answer it looked like 18 months ago. Some things still make more sense to buy - nobody is rebuilding Xero in a weekend. But the set of things where the maths pushes you toward building is growing fast.

The second half of the deal is the bit that doesn't get talked about enough. Even when the cost maths worked for SaaS, the adoption maths often didn't.

The real cost of SaaS has never been the monthly subscription cost. It's been the impact on the business and the people. You ultimately have to make compromises and changes to the way you want to do things. That brings with it the change management, the retraining, the workflows that used to be optimised and are now just a bit more generic.

The thing about software built for how your business actually works is that adoption stops being a fight. You don't have to convince people to bend round the product, because the product already fits. That's a different kind of value, and SaaS has never really been able to offer it.

What breaks, and what doesn't

I don't think this is "all SaaS dies." That's the wrong framing. But the slice that's exposed is a big one.

What's exposed is anything where the moat was essentially "it would be too expensive for you to build this yourself." That moat has gone.

What's protected is different, and it's worth being specific about what actually holds.

There's the data moat - products whose real value isn't the features, it's the aggregated data sitting underneath. LinkedIn, Trustpilot, Rightmove. You can build the UI in a few weeks maybe, but you can't replicate 15 years of accumulated data, and the product would be useless without it.

There's the compliance moat - products living in regulated corners where the risk of doing it yourself is genuinely higher than the risk of continuing to pay.

And there's the deep integration moat, and Xero is a great example. Too much domain knowledge in the code, too much regulation in the detail, over a thousand integrations, bank feeds, tax rules that change every year. Most users actually like it. That combination - deep domain knowledge, complex regulation, sticky integrations, a product people don't hate - is genuinely hard to replicate. It's not a moat that falls because the build cost dropped.

If you run a SaaS company, the useful question isn't "is AI coming for us?" Because we know it's coming for all of us. It's "which of these categories am I actually in?" And if I'm not in any of the three that still have a moat, then how can I get there?

The window is long

One thing to be careful about here. The outlook for the exposed slice is clear, but the timeline isn't short.

Changing software stacks is hard. Gartner has had CRM implementation failure rates at 50-70% for years - and that's before anyone was trying to replace their CRM with their own in-house build. Most companies don't have change management capacity sitting idle, waiting for a good idea. They've got bigger problems to solve first.

The public SaaS numbers bear this out. Bain's latest read has gross retention for public SaaS still sitting at around 90%+. Customers complain about the pricing, they talk publicly about building their own, but they haven't actually walked yet. The narrative is running ahead of the churn data.

So the honest framing is this. The direction of travel is clear. The speed isn't. Enterprise stacks are embedded, switching costs are real, and there's a genuine window here - probably a few years long - for SaaS vendors to evolve and for their customers to think about what they actually want from the next chapter. It's a set of decisions that will play out over the next 5-10 years, realistically.

So a slow churn rate is real, but this probably hurts new business in SaaS companies more immediately.

Evolve, or end up a commodity

Salesforce gave us an interesting version of that answer. At their TDX event they launched Headless 360 - the whole of Salesforce, Agentforce and Slack now exposed as APIs, so that AI agents can drive the platform directly without anyone opening a browser.

"No browser required. Our API is the UI."

The headline is that Salesforce - the company that basically invented modern SaaS, and I'll be honest, a product that I've never enjoyed using (mostly due to the user experience) - is saying their own user interface is no longer the point. The platform becomes infrastructure.

The more interesting bit is what they're putting on top of it. Autonomous agents that go and look for opportunities in your data, action them, and deliver the results. An AI marketing function that spots the thousands of prospects in your pipeline matching a specific signal, writes personalised outreach, sends it, and tracks the response - without anyone asking. Sales agents that do the same for pipeline you haven't had time to chase.

That's a different product. It's still a CRM, but it isn't a CRM that's just a bunch of forms and workflows on top of your sales data. You're paying for the agents to actually do something with the data - turn it into opportunities, and turn those into real sales. That's maybe what a CRM should be in 2026.

The old CRM - the one where you drag opportunities from one column to the next and remember to add your call notes - is something almost anyone could build in an hour. If that's still the product, it's a commodity.

That's the real moat test for SaaS in the exposed tier.

So, would you build one?

Back to the question I started with. Who would build a SaaS company today?

Some, yes - the ones with a real moat. The rest are going to face the same maths their customers are already running.

What I find more interesting is that the default answer for buying enterprise software is being genuinely rethought for the first time since the category was named. For 20 years the answer has usually been "buy SaaS." Now it's a real question again, and the maths you run to work it out has changed.