Buy, build or prompt: why companies are replacing SaaS tools with in-house AI-built apps

For twenty years the build-or-buy question had a default answer for anything that wasn't your core product: buy. Building meant a developer for weeks, and a developer was the scarcest thing in the company. AI coding tools have moved that line. A form, an approval flow or an admin screen that used to be a month's work is now a few days, sometimes an afternoon. So teams are doing the sum again, and some of them are cancelling subscriptions.

What the numbers say

Retool surveyed 817 builders in late 2025, from startups to Fortune 500 companies, for its 2026 Build vs. Buy report. 35% had already replaced at least one SaaS tool with something custom, and 78% expected to build more in 2026. The tools most often replaced were workflow automations (35%), internal admin tools (33%), BI dashboards (29%), CRMs and form builders (25%), project management (23%) and customer support tools (21%) (Retool).

The same report has the less flattering half. 60% of builders had made something outside IT's oversight in the past year. Only 44% said they test thoroughly before deploying. Maintenance was named as a blocker by 26%, and security and compliance concerns by 41%. Retool sells a platform for building internal tools, so read the headline numbers with that in mind; the warnings are more convincing for coming from a company that would rather you built.

The Klarna headline, and what was underneath it

The story everyone remembers is Klarna "replacing Salesforce and Workday with AI" in 2024. Its CEO later corrected it: "No, we did not replace SaaS with an LLM, and storing CRM data in an LLM would have its limitations." What Klarna actually did was consolidate scattered data into an internal stack it built itself, move HR to another vendor, and switch off tools it no longer needed. He also said they learned that auditing, versioning and access control were features they couldn't do without (CX Today).

That's the realistic version of "replacing SaaS": a mix of building the parts that are specific to you, buying the parts that aren't, and dropping the ones nobody used. (Klarna has its own story about how far AI could replace people, which we told in Klarna, Duolingo and the AI-first memo.)

Building got cheap. Owning didn't.

Software engineering has known this for decades: most of what a system costs comes after it's built. Estimates put maintenance at up to 70 to 80% of the total over a system's life, and a common rule of thumb budgets 15 to 25% of the original build cost every year (Edana).

AI makes the first number smaller. It barely touches the second, because most maintenance isn't typing code. It's the payment provider changing its API, a library with a security fix, a browser update that breaks a date picker, a new colleague who needs access, and the question nobody can answer when the person who built it has left. A tool that took an afternoon to prompt still needs an owner for years.

A rough sum

Take a team paying $40 a seat a month for 25 people on a tool they use one feature of: $12,000 a year. Suppose an AI-assisted build of that one feature takes a developer 40 hours including testing and deployment, at $100 an hour: $4,000. Then budget four hours a month to keep it running, plus $300 a year of hosting: about $5,100 a year.

Year one: $9,100 against $12,000. Over three years: about $19,300 against $36,000. The build wins, comfortably. Now change one thing: the tool grows to eight features and four integrations, and upkeep becomes two days a month. The three-year saving disappears. These figures are illustrative; plug in your own. The shape of the answer is what matters: building pays when the thing stays small.

When building pays

You use a sliver of an expensive tool. One workflow out of a platform priced for fifty. This is the case behind most of Retool's replacements.

Per-seat pricing punishes occasional users. Forty people who open a dashboard once a week, each costing a full seat. We wrote about why vendors are moving away from seats; until yours does, a small internal page can be cheaper.

The process really is yours. The way you quote, schedule or approve is part of how you compete, and every off-the-shelf tool makes you bend it.

Someone will own it. A named person who knows the code, has the passwords and gets the alert when it breaks.

When to keep buying

The vendor carries a liability for you. Payroll, tax, e-signatures, payments, anything with a compliance certificate. The subscription is partly insurance.

It's security plumbing. Sign-in, password resets, permissions, backups. Building these is easy; building them correctly is the whole job.

The value is the network or the data. A marketplace, a shared calendar with clients, a CRM whose enrichment data you can't recreate.

Nobody would own it. If the builder is a contractor who leaves next month, or an enthusiastic analyst with a full-time job, the tool will be orphaned within a year.

The third option: prompt

Some of the tools teams are building shouldn't be apps at all. A weekly report, a one-off clean-up of a customer list, a summary of last month's support tickets: these are jobs, not products. An AI assistant pointed at an export or a spreadsheet, with a saved prompt, does them without anything to host, secure or maintain. If a task runs once a week and one person looks at the result, try prompting it for a month before you build anything around it.

A simple way to decide: prompt what one person does occasionally, build what a team does every day the way only you do it, and buy what has to be right for legal, security or money reasons.

If you build, finish the job

An internal tool holds the same customer data and the same keys as a product, often with less care, because "it's only for us". The same checks apply: access rules on every table, no keys in the browser, the server re-checking what the screen checks, limits on anything that sends email, a backup you've actually restored and an owner with the repository in an account the company controls. We walked through why in your AI-built prototype works, check these six things, and turned it into a free checklist you can run in five minutes.

Sources

Retool, The Build vs. Buy Shift: AI, Shadow IT, and the SaaS Replacement Era (2026 report, survey of 817 builders, late 2025); CX Today, Klarna CEO explains what really happened with Salesforce; Edana, Estimating software maintenance costs (January 2026). The cost example is our own illustration, not survey data. Hero photo from Pexels (Pexels License).