In short
Buy the generic layer and own the specific one. Models, infrastructure and the commodity functions every company runs the same way are bought, because a supplier runs them better and cheaper. The workflow and data of a function your business depends on are owned, because that is where your advantage sits and where a bought tool cannot adapt.
The question is new because the cost of building changed. For decades, as our article on vertical integration for mid-sized companies argues, building your own software paid off only at the scale of a Tesla. AI lowered that bar, and the survey numbers below show companies walking through it. The numbers also show most of them tripping. This article gives both sides, and a four-question test for the decision.
xvertical does not work for any of the companies named here. They are public cases a reader can check.
Build or buy: the numbers behind the shift
Just over a third of the builders Retool surveyed had replaced software they used to buy. Retool published its build versus buy report in February 2026, based on “a comprehensive survey conducted in late 2025 across 817 builders, including Retool customers.” Of those builders, “35% of them have already replaced at least one SaaS tool” with something built in-house, and “78% expect to build more of their own tools in 2026.”
The report’s explanation is cost. It is “harder to justify spending thousands on a subscription to a generic tool” when “your team could have a prototype tailor-made to solve your problem in a few days.”
Two caveats. Retool sells a platform for building internal tools, and its respondents are people who already build. The share across all companies is lower. But the direction is the one we described for mid-sized companies: the price of owning a piece of software fell far enough that the decision is now open to companies that never had it before.
Why buying still wins more often
Buying still wins more often. MIT’s NANDA initiative published its report on the state of AI in business in August 2025, summarised by Fortune. Based on 150 interviews, a survey of 350 employees and 300 public AI deployments, it found that “purchasing AI tools from specialized vendors and building partnerships succeed about 67% of the time,” while internal builds succeed at roughly a third of that rate.
Gartner’s April 2026 survey of 782 infrastructure and operations leaders, reported by The Register, points at the causes. Its research director Melanie Freeze said: “The 20 percent failure rate is largely driven by AI initiatives that are either overly ambitious or poorly scoped.” Among the leaders who had seen a failure, 38% named persistent skill gaps and 38% named poor data quality or limited data availability. The full figures are in why most companies stay AI-enabled.
Read together, the two surveys suggest that the companies building more are also the companies whose builds fail more. They are different samples, so this is a reading, not a measurement. The findings are not in conflict either. They describe the same decision made well and made badly. Fortune’s summary of the MIT report calls the cause a “learning gap”: generic tools do not adapt to a company’s workflows, and companies do not learn the tools. A vendor’s tool fails for the first reason. An internal build fails for the second, and for the two Gartner named: the builder did not have the skills, or the data was not there.
Four questions before you own a capability
The decision is not build or buy for the whole company. It is made per function, and four questions settle it. A capability that passes all four is owned. One that fails any of them is bought, or not built at all.
1. Does the business depend on it? A function where a bad week shows up in the numbers. For a restaurant group that is the booking and the reporting behind it. For a publisher it is the IT support line for about 1,600 staff. Payroll and email run the same way in most companies and usually stay bought.
2. Do you own the data and the workflow it runs on? An AI capability is only as specific as the data it sees and the process it sits in. If the data lives in a supplier’s system and the workflow is the supplier’s, you would be building on ground you do not own. Bring the data home first, then decide.
3. Can you run it after it is built? This is the Gartner skill gap. A capability nobody inside the company can change is a bought tool with worse support. The people who run the function have to be able to adjust the workflow, read the data and tell when the AI is wrong. So the build has to include them, or the capability is better bought.
4. Is it specific to how this business works? The removal test from what is an AI-native company applies here. If a generic tool would do the same job for your competitor, buy the generic tool. If the value is in the way your venues, your products or your customers differ, that difference is the thing you own.
What does owning a capability include, and what stays bought?
Owning a capability means owning the workflow, the data and the decisions, not every part. The same was true before AI. Vertical integration means owning a stage of the work, and SpaceX, the clearest public example, still buys many of the parts it assembles.
An owned AI capability has the same shape. The model is bought from a lab. The infrastructure is bought from a cloud. Often a proven tool stays where it is, such as the CRM the sales team already knows. What the company owns is the workflow the function runs on, the data that workflow produces, and the decisions about when a person steps in. Those are the parts no supplier can own on the company’s behalf, and the parts that make the capability specific.
That is also the answer to the Retool finding. The teams replacing subscriptions with custom builds are not writing databases. They are owning the layer that used to be a generic tool’s opinion about how their work should run.
What happens when building is cheap and nobody decides?
Retool found that “60% of builders across levels of seniority have built something outside IT oversight in the past year,” and that “25% report doing so frequently.” The reasons given were speed, unmet needs and a slow IT process.
Cheap building means building happens whether or not the company decides it. The report’s warning: “Connectivity, security, and governance will determine which custom builds drive transformational change at the company level.” And later: “Organizations need to prepare with the right systems and guardrails in place.”
For an owner this is the practical consequence of the four questions. A capability the company decides to own gets the data access, the security review and the people to run it. A capability someone built on a Friday afternoon gets none of those, and it still runs part of the business. The decision is being made either way. Making it deliberately, function by function, is the difference between owning a capability and inheriting one.
Key takeaways
- Buy the generic layer, own the specific one. Models, infrastructure and commodity functions are bought. The workflow and data of a function the business depends on are owned.
- Retool’s 2026 survey of 817 builders found 35% had replaced at least one software subscription with a custom build and 78% expected to build more. The cost of building fell.
- MIT found vendor purchases and partnerships succeed about 67% of the time and internal builds about a third as often. Gartner names the causes: scope, skills and data.
- Four questions decide per function: does the business depend on it, do you own the data and workflow, can you run it afterwards, and is it specific to how you work.
- Owning does not mean building everything. Buy the model, own the workflow, the data and the decisions.
- Cheap building happens with or without a decision. Retool found 60% of builders had built outside IT oversight. Decide deliberately, function by function.
Sources
Each quote above is linked where it appears. This is the full list.
- Retool, “The build vs. buy shift: how vibe coding and shadow IT have reshaped enterprise software”, 17 February 2026. The survey of 817 builders, the 35%, 78%, 60% and 25% figures, the cost explanation and the governance warning.
- Sheryl Estrada, Fortune, “MIT report: 95% of generative AI pilots at companies are failing”, 18 August 2025. The MIT NANDA method, the 67% success rate for purchases and partnerships, and the “learning gap”.
- Lindsay Clark, The Register, “Only 28% of AI infrastructure projects fully pay off, survey finds”, 7 April 2026. The Gartner survey of 782 infrastructure and operations leaders, the causes of failure and the quote from Melanie Freeze. The underlying release is Gartner’s press release of 7 April 2026.
Questions this post answers
Should a company build its own AI tools or buy them?
Buy the generic layer and own the specific one. Models, infrastructure and commodity functions such as payroll are bought. The workflow and data of a function the business depends on are owned, because that is where the company's advantage sits and where a bought tool cannot adapt.
How many companies have replaced software they used to buy?
In Retool's 2026 survey of 817 builders, conducted in late 2025, 35% had already replaced at least one SaaS tool with a custom build, and 78% expected to build more of their own tools in 2026.
Why do internal AI builds fail more often than purchases?
MIT's 2025 study found that buying from specialised vendors or partnering succeeds about 67% of the time, while internal builds succeed at roughly a third of that rate. Gartner's survey names the usual causes: a project scoped too wide, people who cannot run what was built, and data that is not there.
Which capabilities should a company own?
The ones that pass four questions. The business depends on the function. The company owns the data and the workflow it runs on. The company can run the result after it is built. And the capability is specific to how this business works rather than generic.
Does owning a capability mean building everything yourself?
No. SpaceX buys many of the parts it assembles. An owned AI capability buys the model and the infrastructure and owns the workflow, the data and the decisions. The ownership is in the process, not in the components.