- Build vs buy AI is only half the question. There are four ways to secure an AI capability: adopt, partner, build or buy.
- Place each capability on two questions: how much could it differentiate you, and how available is it in the market? Then check the answer against six signals.
- Adopt the commodities, partner where expertise is scarce, build where your data creates the edge, and buy only when the edge belongs to someone else.
In this article
“Should we build or buy AI?” is the question every leadership team asks about AI. It is also the wrong question, because it offers two answers when there are really four: adopt an off-the-shelf tool, partner with a specialist, build with your own team, or buy a company that already has what you need.
Picking the wrong path is expensive in both directions. Build what you could have rented, and you burn a year re-creating a commodity. Rent what you should have owned, and you hand your competitive edge to a vendor who will sell it to your rivals next quarter.
“
Build what you could have rented, and you burn a year
re-creating a commodity.
We first asked a version of this question for IT in general in Which approach to take: build it, buy it, cloud it? Our answer then still holds: start from strategy, not habit, and buy what is a commodity. AI keeps that logic, but adds two more paths and raises the stakes.
Boston Consulting Group (BCG) tackles exactly this choice in its 2026 M&A Report article “How Non-AI Companies Win at AI Dealmaking” (September 2026). The stakes are rising: BCG counts more than 1,300 new AI alliances in 2025, more than double the 2024 figure, now rivalling M&A as a route to AI. And when nontech companies do acquire AI software firms, investors reward them with announcement returns of about 0.8%, nearly three times the average deal.
BCG’s framework boils the decision down to two questions that place any capability, plus six signals to check the answer. Here’s how it works.
Why the build vs buy AI decision is worth getting right
The sourcing choice matters because the market pays very differently for different kinds of AI. In a June 2026 study of 471 private-equity-backed companies across 31 industries, McKinsey found that companies using AI to run themselves more efficiently traded at roughly the same median revenue multiple as those still piloting. Companies that built AI into what they sell traded at far higher multiples.

Source: McKinsey & Company (June 2026). 471 PE-backed companies, $1M–$250M revenue. Association, not causation.
The split lines up with the four paths. Productivity AI keeps you competitive, and it is mostly bought off the shelf or through partners. The premium sits where AI becomes part of the product or a new business, which is where your own data and IP, and the build and buy decisions, come in. Yet three in four companies in the sample are still at Levels 1 and 2.
Two caveats keep this honest. McKinsey reports an association, not proof that AI causes the higher multiples, since stronger companies may simply be the ones that get there. And only 21 companies reached Level 4, so treat the 31x as directional.
Two questions that place any AI capability
BCG’s framework starts by asking two things about each capability on your roadmap. First, how much could it differentiate you from competitors? Second, how available is it in the market: can anyone license something comparable, or is it locked up with one provider, team or dataset? Plot the answers and each capability falls into one of four boxes.

The build vs buy AI framework: four paths by differentiation and market availability, adapted from BCG (2026), Exhibit 4.
Read it this way: the further right, the more the capability belongs to your strategy; the further down, the harder it is to get anywhere else.
Adopt, partner, build or buy: four different bets
Adopt: use what already exists. When a capability is widely available and gives you no real edge, the smart move is to switch it on and move on. General-purpose employee copilots and standard document or call processing belong here. Your competitors have the same tools, so the win comes from rolling them out quickly and well, not from owning them.
Partner: tailor it with a specialist. Some capabilities aren’t yet a commodity, but they aren’t your secret sauce either. Think of an industry-specific tool that is still maturing, or a vendor product that only works after some co-development. A partner brings the expertise and speed; you bring the context and requirements. Lighter is not easier, though: BCG found that roughly nine in ten partnerships between industrial and technology companies missed their goals, so settle data ownership, IP and exit rights before you sign.
Build: develop it yourself. When the underlying techniques are widely available but your data, workflows or intellectual property make the result unique, building pays off. Forecasting models trained on proprietary data and domain copilots built on company know-how are classic examples. The model is ordinary; what you feed it is not.
Buy: acquire the scarce asset. Occasionally the thing that makes a capability valuable sits with someone else: a dataset only one company holds, or a small team with skills nobody else has. If it matters to your strategy and you can’t build it in time, acquiring it may be the only way to secure it.
What Jim Collins would add
Twenty-five years before the AI boom, Jim Collins reached a conclusion in Good to Great that fits BCG’s framework closely. The great companies he studied never used technology to create momentum. They used it to accelerate momentum they already had, and only when the technology tied directly to their Hedgehog Concept.
The Hedgehog Concept sits where three circles overlap: what you can be best in the world at, what drives your economic engine, and what you are deeply passionate about. Read the 2×2 through that lens:
- Inside your Hedgehog ? build or buy. If an AI capability strengthens the thing only you can do, it deserves ownership and your best people.
- Outside your Hedgehog ? adopt or partner. Everything else should be good enough, fast and cheap. Being merely competent at commodities is a feature, not a failure.
Inside your Hedgehog, build or buy. Outside it, adopt or partner.
Two more Collins ideas sharpen the choice:
- First who, then what. BCG’s “buy” box often holds a specialist capability concentrated in a scarce team. That is a people decision before it is a technology decision. If you can’t attract or keep the right people, no build plan will save you.
- The flywheel, not the doom loop. Companies that chase every new AI tool, switching vendors and strategies each quarter, never build momentum. Pick your few differentiating capabilities, push on them consistently, and let AI make each turn of the flywheel faster.
Six signals to pressure-test your build vs buy AI decision
The 2×2 gives you a first answer. Before you commit budget, BCG recommends running the capability past six practical considerations, summarized below (adapted from BCG, Exhibit 5). Adopt and partner are the lighter-touch options; build and buy give you more control but demand more commitment
| Signal | Adopt | Partner | Build | Buy |
| Differentiation: could a rival copy it easily? | Yes, it’s a utility | It’s useful, but no edge | Your data or workflows make it unique | Someone else holds the source of the edge |
| Urgency: how fast do you need it, and how fast is it commoditizing? | Ready today, standardizing fast | Need speed plus customization | You have time to develop and earn it back | Needed now, and only one target has it |
| Market capability: can you match outside options? | External tools are better | A partner knows more than you | You can match or beat the market | A target performs uniquely well |
| Integration and control: how deep into your stack does it go? | Standard, low-sensitivity use | A vendor can meet your governance needs | Deep integration is where the value is | Control can’t be separated from the data or IP |
| Economics: is there an asset at a fair price? | A cheap service exists | No asset is worth buying | No good target, or building is cheaper | Valuation and synergies justify a deal |
| Lock-in: how costly is changing course? | Keeping options open matters most | You want closeness without permanence | The control is worth the commitment | Owning it removes a critical dependency |
If most of your answers land in one column, you have your path. If they scatter, that’s a sign to split the capability into parts, for example adopting the commodity layer and building only the piece that differentiates.
What we see in the field at Joyful Craftsmen
Frameworks are only useful if they survive contact with real projects. Here is what our data and AI work with clients across Europe adds to the picture.
“Build” stands or falls on your data foundations. An AI agent without business context behaves like a very fast junior analyst who has never seen your company. What makes it reliable is a well-structured semantic layer: clean definitions, modelled relationships and business logic it can draw on. If your models aren’t AI-ready yet, that work comes first. We explain why in Why your data still can’t answer a simple question.
Adopt versus build is often a speed-versus-control trade-off. In the Microsoft stack, out-of-the-box options like Fabric and Copilot tools get you started fastest but limit extensibility and testing. Custom builds on Azure take longer but give you control and validation. Our advice: start with whatever gets you to a testable prototype quickest, then decide after a few weeks of real user feedback.
Commit to a direction, not a destination. When we built a data agent in Microsoft Teams for HOPI’s logistics team, we changed technical approach more than once. HOPI’s CIO, Jan Laš, recommends weighing two or three approaches before committing and re-checking the technology roughly every month. The teams were small: three part-time people on our side, two in IT and one business owner on HOPI’s side. Security took far more effort than planned once access widened beyond a small pilot group. The full story is in Deploying AI in logistics (the unfiltered version).
Whatever the path, gate it on value. We don’t start a use case without an agreed return, a named owner and a plan for getting it into production. Once it’s live, we score it on six branches: Impact, Cost, Quality, Adoption, Data and Control. Then we decide whether to scale it, fix it or switch it off. Of the six, data quietly decides the other five. See How to measure AI value.
Mistakes to avoid
- Building for pride. Engineering teams love to build. But a custom copilot that matches an off-the-shelf subscription is not a strategy; it’s a maintenance bill.
- Adopting your edge away. If the capability runs on your proprietary data, plugging it into a generic tool can quietly turn your advantage into the vendor’s training set or roadmap.
- Treating the decision as permanent. AI capabilities commoditize fast. Today’s “buy” can be next year’s “adopt”. Revisit the map at least once a year.
- Ignoring the exit. Every path has a switching cost. Ask how hard it would be to walk away before you sign, not after.
How to start
- 1
List the AI capabilities on your roadmap, from customer service bots to forecasting.
- 2
Place each one on the 2x2: how much could it differentiate you, and how available is it?
- 3
Run the borderline cases through the six signals.
- 4
Put most of your budget and best people on the “build” and “buy” boxes.
Move fast and cheaply on the rest.
The bottom line
The real question isn’t build or buy. It’s where your advantage actually lives, or in Collins’ terms, what sits inside your Hedgehog. Adopt the commodities, partner where expertise is scarce but strategy isn’t at stake, build where your data creates the edge, and buy only when the edge belongs to someone else. Get that allocation right, and AI spending turns into advantage instead of overhead. BCG’s own rule of thumb is to choose the lightest structure that delivers the control you need, and to keep ownership the exception rather than the default.
Sources: Jens Kengelbach, Daniel Friedman, Lianne Pot, Georg Keienburg and Tobias Söllner, “How Non-AI Companies Win at AI Dealmaking”, The 2026 M&A Report, Boston Consulting Group, September 21, 2026. The 2×2 and six-signals table are adapted from the article’s Exhibits 4 and 5. Hedgehog Concept, “first who, then what” and the flywheel are from Jim Collins, Good to Great (HarperBusiness, 2001). Valuation data from McKinsey & Company, “Beyond productivity: How AI creates value in private equity,” June 23, 2026.
Frequently asked questions
Should you build or buy AI?
Neither by default. Build vs buy AI is only half the question: there are four options. Adopt an off-the-shelf tool, partner with a specialist, build with your own team, or buy a company that already has what you need. Which one fits depends on how much the capability could differentiate you and how available it is in the market.
What is the difference between build, buy and partner for AI?
Building means developing the capability with your own team, which pays off when your data or know-how makes the result unique. Partnering means tailoring it with a specialist, for capabilities that aren’t a commodity but aren’t your edge either. Buying means acquiring a company that holds a scarce dataset or team. Adopting, the fourth option, means switching on a ready-made tool.
When should a company build AI in-house?
When the underlying techniques are widely available but your own data, workflows or intellectual property make the result unique, for example forecasting models trained on proprietary data.
Should you build or buy AI agents?
It depends on what the agent needs to know. Out-of-the-box options such as Copilot tools get you started fastest but limit extensibility and testing; custom builds on Azure take longer but give you control. Either way, an agent is only reliable with business context, so the data foundations come first. Start with whatever gets you to a testable prototype quickest, then decide after a few weeks of real user feedback.
When does buying an AI company make sense?
When the scarce asset, such as a unique dataset or a specialist team, sits with someone else, matters to your strategy, and you cannot build it in time.
How often should you revisit the build vs buy AI decision?
At least once a year. AI capabilities commoditize fast, so today’s buy can be next year’s adopt.
Free Webinar
Time to Make Your AI Pay Off
get ready for the EU AI Act and plan your 2027 AI budget.
Thursday, 26 November, 2026
14:00-14:45 CET • 45 minutes, in English, with Ivan Jelic
Not sure which box your capabilities belong in? At Joyful Craftsmen we help companies map their AI use cases, place them on the adopt-partner-build-buy grid and get the data foundations ready for the ones worth owning. Book a call and we’ll work through your portfolio with you.
Related reading
The post Build vs Buy AI: When to adopt, partner, build or buy appeared first on Joyful Craftsmen.
