Treat build vs. buy as a strategic decision, not an IT purchase.
The choice locks in cost, talent, data and vendor dependencies for years, and much of it is hard to undo once workflows and data pipelines are built around it. The gap between demo and daily use is where most of the risk sits: a Gartner survey of 699 respondents in the US, Germany and the UK, fielded in late 2021 and published in 2022, found that on average 54% of AI projects made it from pilot to production. The data predates generative AI, so treat it as a rough benchmark. Before choosing, separate the parts of the decision you can reverse cheaply from those you can't; our guide to reversible vs. irreversible decisions shows how.
Buying is not automatically the safe option.
In June 2025, Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs and unclear business value, and estimated that only about 130 of the thousands of vendors marketing "agentic AI" are real, as reported by Reuters. These are an analyst forecast and an estimate, not measured outcomes, but they are a good reason to test vendor claims hard. Our checklist on how to evaluate an AI vendor before you buy covers what to ask.
Use five questions to decide.
Answer each one as "build" or "buy/partner":
- Differentiation: Would a competitor using the same off-the-shelf tool erase your advantage? If the use case is generic (meeting notes, support triage, document search), buy.
- Data: Does the value depend on proprietary data that cannot leave your environment, or on deep integration with internal systems? That points toward building, or toward a partner who works inside your stack.
- Ownership: Do you have, or can you hire and keep, people to run, monitor and update the system for three years, not just ship version one? Name the owner before you decide.
- Total cost and time to value: Compare three-year cost, including integration, security review, maintenance and model changes, with the cost of waiting six to twelve months for an internal build.
- Exit and lock-in: Can you export your data, prompts and configurations if the vendor fails, raises prices or is acquired? If not, price that risk in. If fewer than three answers say "build", your default should be buy or partner.
More on AI Strategy — every week.
Join 2,400+ leaders getting the decision intelligence briefing.
Consider the third option: buy the platform, build the thin layer.
Most mid-sized companies don't face a pure either-or. A common pattern is to buy the commodity layer (models, platform, hosting), build the thin layer that encodes your own process and data, and bring in a partner for integration. You keep the differentiating logic in-house without maintaining infrastructure you don't need to own. Whatever you choose, set a budget and kill criteria up front; our guide to setting a budget for an AI pilot walks through it.
Hypothetical example.
A 180-person B2B software company wants AI to draft answers to customer tender documents. Differentiation is medium: good answers depend on its own product data and past winning bids. That data must stay in its EU environment. It has two engineers with machine-learning skills, but both are committed to the core product roadmap. A vendor can go live in eight weeks; an internal build is estimated at six to nine months. The vendor allows full data export. The leadership team decides to buy the platform, build only the retrieval layer over its own bid library, and review after 90 days against hours saved per tender and win rate. This is an illustration, not a customer case.
Write the decision down before you sign or staff it.
Record the option you chose, the assumptions behind it (adoption, cost, accuracy), the owner, and the triggers that would make you revisit it, such as a vendor price increase above an agreed threshold or accuracy below target after 90 days. That turns a one-off bet into a decision you can review. If you want the trade-offs pressure-tested first, a NeuroAgent council has up to 16 AI advisors debate build, buy and partner against your constraints, with an AI-assisted boundary check testing the result against your boundaries; you can try it free without an account.
This article is general guidance, not legal, tax or financial advice. Contract terms, data protection and EU AI Act obligations depend on your specific situation.
Research context
Stanford's 2025 AI Index reports that 78% of organizations used AI in 2024, up from 55% in 2023.
Stanford Institute for Human-Centered AI: The 2025 AI Index Report (2025)This independently published statistic provides context for the topic; it does not substantiate every claim in this essay.
One decision framework, every week.
The week's best articles on making and executing high-stakes decisions, backed by sourced research.