Niche SaaS Is Bleeding Out as Clients Build It Themselves

Build vs buy internal tools with AI isn't a close call anymore for most back-office software. Thirty-five percent of enterprise teams have already ripped out at least one SaaS subscription and replaced it with something built in-house using AI, and 78% expect to build more custom tools in 2026.¹ The default used to be buy. It just flipped.

Why did the build-vs-buy math flip?

Two years ago, a custom internal tool meant an engineering team, a project plan, and a bill north of six figures before anyone touched production. Newsweek's reporting on the shift is blunt about the before-and-after: "Two years ago, a custom internal tool might take an engineering team weeks and cost six figures. Today, a business operations lead with the right platform can have a working prototype in a day or two."² That's not a marginal efficiency gain. It's a different cost structure entirely, and cost structure is what drives procurement decisions.

Fifty-one percent of builders in the same survey say they've already shipped production software their teams currently rely on.² That's not shadow experimentation. That's live infrastructure, built by people who were never on an engineering headcount.

How many companies are already building instead of buying?

This is measurable, not anecdotal. Alongside the 35% replacement rate and 78% forward intent, 60% of builders say they built software outside official IT oversight in the past year, and a quarter of them do it frequently.¹ That's the base rate for a new category of software creation, happening whether IT signs off or not.

The revenue side backs it up. AlixPartners found more than 100 publicly listed midmarket software companies getting squeezed from both directions: AI-native competitors undercutting them on price, and giants like Microsoft and Salesforce burying them under AI-feature bundling.³ The firm's own words: "We believe many mid-size enterprise software companies will face threats to their survival over the next 24 months."³ Median net dollar retention for enterprise software companies fell from 120% in 2021 to 108% in Q3 2024.³ That number is the quiet leading indicator. Existing customers aren't walking away in a press release. They're just not expanding the seats they already bought, and eventually they stop renewing them too.

Where is this hitting first: internal tools and narrow workflows

The exposure is concentrated in specific categories. Netlify has watched more than 10,000 new sites get built and launched daily through AI coding tools on its own platform, and its own staff have used those same tools internally: an employee survey app, and a revenue operations staffer using Bolt to build a CPQ-style pricing calculator that would otherwise have been bought as SaaS.⁴ Harmonic, a startup discovery platform, had an engineer rebuild a $20,000-a-year third-party tool inside Retool after getting fed up with slow vendor support. The company now runs 33 internal apps that would previously have shown up as separate line items on a SaaS bill.¹ Our own practitioner's guide to replacing $50/mo SaaS tools walks through exactly this pattern at smaller scale.

The pattern across all of these examples is the same: dashboards, CRM and CPQ replacements, admin panels, workflow automation. Software engineer and investor Martin Alderson frames the exposure precisely: "the companies that are at serious risk are back-office tools that are really just CRUD logic - or simple dashboards and analytics on top of their customers' own data... if your product is just a SQL wrapper on a billing system, you now have thousands of competitors: engineers at your customers with a spare Friday afternoon with an agent."⁵ Anything that's fundamentally a form, a table, and some business logic is now a build decision, not a buy decision. We've covered how frontier frontend models are killing niche SaaS dashboards in more depth, and the trend line has only moved further that direction since.

Are agencies and consultancies feeling this too?

This isn't only a SaaS vendor problem. Outside dev shops and consultancies are getting squeezed by the exact same mechanism. The Financial Times reports that banks including UniCredit and Société Générale cut consultant spending by 24% and 9% respectively in 2025-2026, with clients openly saying AI is now doing work they used to pay outside advisers and integrators to do.⁶ One tech executive told the FT the cost of that outside support is simply "collapsing."⁶

The consulting industry's own commentary confirms the mechanism. Buyers can now run a quick AI experiment on a sample of the work and see, in minutes, what used to take a junior consulting team days.⁷ That kills the pricing power that staffing-model consultancies relied on. It's the same dynamic hitting software agencies: clients ask "can we just build this?" before signing a renewal, and increasingly the answer is yes.

Does this mean SaaS is dead?

No. It means undifferentiated SaaS is finished. Products with a real moat are holding up fine. Payment infrastructure, high-availability systems, and regulated systems of record stay comparatively insulated because the risk of getting them wrong is asymmetric and the switching cost is real, not just habitual.⁵ Publicis Sapient is cutting SaaS licenses by roughly half while replacing simpler tools with AI builds, but nobody is proposing to hand-roll their payments stack.⁸ A Salesforce customer walked away from a $350,000 contract to build a custom replacement with Base44, which tells you where the line sits: complex CRM logic on commodity infrastructure is buildable now, but that same customer isn't rebuilding Stripe.⁸ Notably, 76% of AI use cases in 2025 were still purchased rather than built internally, up from 53% the year before, which cuts the other way: AI itself is mostly bought, even as the software wrapped around it gets built.⁸ The moat questions from here are the ones Bain-style frameworks keep coming back to: does the product carry compliance or audit weight, does it need five nines of uptime, does it benefit from network effects, and does it sit on proprietary data the customer can't replicate. If the answer to all four is no, the product is exposed.

What's the hidden cost of building instead of buying?

The part that gets less attention is what happens after the build. Sixty percent of builders created software outside official IT oversight in the past year.¹ Thirty-seven percent of organizations haven't yet established any AI productivity metrics for that work.² That's a governance gap, not a productivity story. A tool an ops lead built in an afternoon still needs an owner, a security review, and a plan for what happens when that person leaves the company. Skipping the build cost doesn't skip the operating cost. It just moves it downstream, usually to whoever discovers the unowned internal app during an audit.

How should you decide whether to build or buy right now?

The decision framework is simpler than it used to be, because the cost side of the equation collapsed. Before signing a renewal or a new SaaS contract, ask:

1. Is this tool mostly CRUD logic on our own data? If it's forms, tables, and business rules with no proprietary dataset behind it, an AI coding agent can likely replace it in days.
2. Does the vendor's moat rest on compliance, uptime, or network effects we can't replicate? If yes, keep buying. If no, price out the build.
3. Who owns this after it ships? A built tool with no named owner and no monitoring is a governance liability, not a savings.
4. What does the renewal actually cost against a day or two of build time? Retool's own reporting frames this trade directly, and for a growing share of mid-priced tools, the build side wins outright.¹²

The vendors and agencies most at risk are the ones that never had a real moat to begin with. They sold convenience, not defensibility, and convenience is exactly what an AI coding agent replicates for free on a Friday afternoon. Companies serious about this shift, including platforms like Remy, are building the tooling to make owning that software the easy default rather than the ambitious exception.