Build, buy or partner for AI capabilities, scored on four tests
Public guidance says what to check but not how to choose a route. Score a proposed AI capability on data advantage, speed, cost of ownership and reversibility before anyone signs.
In our experience, most AI capability decisions are made by the team with the most urgent demo, and the sourcing question gets answered afterwards by whoever signs the contract. There are three routes. A company can build the capability, buy a finished product, or partner with a supplier that shares the work and some of the risk. This framework gives a proposed capability a score on four tests, so the route is chosen on paper before the vendor deck arrives.
We looked for published studies of how enterprises decide between building, buying and partnering for technology, and found none in the sources available to us. The four we reviewed were NIST, the UK government, the National Bureau of Economic Research and Brookings. None of them measures build, buy or partner costs, timelines or reversibility, and none offers a data-advantage measure. What follows is therefore editorial judgement, offered as a working tool rather than a tested method.
What the public guidance does cover
The most directly relevant government text is the AI Playbook for the UK Government, published by GOV.UK. It explains what AI is, including its capabilities, limitations and risks, and how to select, buy and deploy AI in government. It sets out 10 principles that civil servants should uphold when using AI. Its stated scope stops at select, buy and deploy, and it gives no scoring method for choosing between building, buying and partnering.
NIST, the US standards body, published version 1.0 of its AI Risk Management Framework on 26 January 2023. According to NIST, it is a voluntary framework for managing AI risks to individuals, organisations and society. It was developed through a request for information, draft versions, public comments and workshops. It helps leaders think about risk. It does not tell them whether to build a system or buy one.
The research sources sit further from the decision. A National Bureau of Economic Research working paper, Chinese Sputnik Moments?, discusses a low-cost, high-performing AI model released in 2025, which some observers cite as evidence of China's rapid technology advance. Brookings's AI pages carry headlines on how the tax code treats AI spending and on fiscal policy, but we had no body text for them. Neither source offers a company a sourcing test.
Before you start
Three things must be true before scoring begins. First, the capability must be named as a business outcome, such as a claims triage step or a supplier-risk alert, rather than as a technology. Second, one executive must own that outcome and its budget. Third, someone must list the data the capability would use, where it sits and who may use it. Without these, every score is a guess dressed up as analysis. Two or three candidates are usually enough for a first round. More than that and the session turns into a tour of vendor demos.
Chapter one: 1. Data advantage
This test asks whether the capability depends on data that only your company holds, and whether that data would produce a result a competitor could not buy. A vendor's general model is the same for every customer who licenses it. Your claims history, maintenance logs or customer conversations are not. Write down the three data sources the capability needs, then ask whether a rival using the same vendor product would get the same result. If so, your data advantage is low, and buying is probably the sensible route.
Where the data is yours and valuable, the choice moves toward building or partnering on terms that keep it yours. Score 1 for a capability any rival could buy, and 5 for one that runs on data nobody else has. Read the contract before the score is final. Who owns the outputs, whether the supplier may train on your data, and what happens to that data at exit can move a high score down. Those are questions for your legal counsel, not for a scoring session.
Chapter two: 2. Speed
Speed asks how soon the business needs the capability, set against how long each route takes to deliver a version people will actually use. The test is a dated window. Write the date by which the capability must work in the operation, then ask whether your team can build, integrate, secure and test it before then. If the answer is no, buying or partnering is the only route that meets the window, whatever the other scores say.
The usual error is to estimate build time from the demo. A working prototype can take days. The work that follows, cleaning data, connecting systems, passing a security review and training staff, takes far longer. The public guidance we reviewed gives no timelines for any route, so every duration in your plan is an estimate. Ask the team that will do the work to write its own estimate, then ask someone outside the project to challenge it. Score 1 for a window measured in years and 5 for one measured in months.
Chapter three: 3. Cost of ownership
Cost of ownership is the full cost of running the capability over its expected life, not its purchase price. For a bought product, that includes licence fees, usage charges, integration, monitoring and the staff who manage the supplier. For a built one, it includes the engineers who build it and, more often forgotten, the engineers who maintain it after launch. Price all three routes over the same period, usually five years, and include the cost of stopping. We found no published figures for build, buy or partner costs, so this model has to be built from your own numbers.
Compare like with like. The most common mistake is to set a vendor's annual fee against the salary of one internal engineer, which makes building look cheap. A fair comparison includes the people needed to keep the system running, the hosting, and the time spent by security and legal teams. Ask your finance lead to review the model, and ask your accountant how the costs should be treated for accounts and tax. Those treatments vary by jurisdiction and should be confirmed by a qualified adviser, not by this article.
Chapter four: 4. Reversibility
Reversibility asks how hard it would be to stop, change or leave the capability within a year. A capability you can switch off within a quarter, with your data exported in a usable format, scores high. One that depends on a supplier's proprietary format, a model tuned on a single platform, or workflows nobody else in the company understands scores low. Test it by writing the exit plan before signature: which data comes out, in what format, what switching off would cost, and who would carry it out.
Low reversibility does not rule a route out. It means the route needs a written exit plan and a named owner before the contract is signed. Once all four tests are scored, read them together. High data advantage with low reversibility points to building or partnering, with strict exit terms. High urgency with low data advantage points to buying, with the exit plan written first. Score each test in a room with the sponsor, finance, IT and at least one person who did not write the proposal, because the sponsor's scores are the least reliable.
Before the route is chosen
Editorial checklist from the four tests. Each yes or no answer should be written down with the name of the person who gave it.
- Can you name the data this capability needs, and who owns it?
- If a rival bought the same vendor product, would it get the same result?
- Is there a dated window for the capability that is shorter than a build your team could st
- Have you priced all three routes over five years, including the people who run the system
- Can you export your data and switch the capability off within one quarter?
- Does the contract state who owns the outputs, whether training on your data is allowed, an
- Have your finance lead and your accountant reviewed how the costs should be treated?
- Has a named owner signed off the scores, with at least one scorer who did not sponsor the
Outgrace checklist
Where it breaks
The framework fails in three predictable places. The first is scoring by the sponsor alone. A project's champion tends to score speed high and cost low. Use at least two scorers, and have them disagree in writing before they agree. The second is treating the scores as a verdict. They are a structure for argument. A close result should send the team back to the data and the contract, not to a vote. The third is assuming the scale is calibrated. The 1-to-5 scale is our judgement. No study we reviewed tests it against outcomes, so the first two rounds in your company should be used to calibrate it.
intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems.
That description is about building trustworthiness into AI systems. It is not a method for choosing a route to them. Keep the two questions apart in your papers, so that a risk review does not stand in for a sourcing decision, and a sourcing score does not pass for a risk assessment. Each needs its own owner and its own record.
On Monday, pick the one AI capability already in front of your board, and book 90 minutes with a sponsor, a finance lead, someone from IT and one person who did not write the proposal. Before anyone opens a vendor deck, write the data list, the dated window and the five-year cost for each route. Score the four tests, then ask the supplier for its data, ownership and exit terms in writing. The meeting's output should be a score sheet and a list of open questions, not a decision.