For most of my career, “should we build this or buy it” was a question with a fairly stable answer. Building was slow and expensive, buying was fast and predictable, and unless the capability was strategic you bought. I have a financial background, and I have built a lot of these models. The spreadsheet almost always agreed.
This year the inputs changed, and I think a lot of operators are now re-opening every capability decision they have with assumptions that were true in 2024 and are not true anymore. Some in one direction, some in the other. Let me go through four numbers from McKinsey’s 2026 State of AI survey, because they describe the situation better than I could.
32% of organisations have now decided not to buy at least one software product because they could build it in-house with AI coding tools.
This is the big one. A third of companies looked at something they would previously have licensed and said, we can do this ourselves now. And quite often they are right. Agentic coding has genuinely changed how much engineering effort a build needs. Something that used to take a team of five for six months might take two people for two.
What I would at least say is that the part that got cheaper is the part that was never the expensive part. In my own experience building machine learning products, writing the code was a small fraction of the work. The expertise, the research, the data, the iteration after the first version disappoints you, the maintenance when the world underneath the model changes. None of that got two and a half times faster. The line on the spreadsheet that shrank is the wrong line to get excited about.
20% say AI operating costs, including tokens, are already constraining how much they use it.
One in five. This is a new cost category that did not exist in anyone’s model two years ago, and it behaves differently from a licence fee. A vendor subscription is a fixed number you can budget. Token consumption scales with usage, with complexity, with how ambitious you get, and McKinsey notes that even as per-token prices fell, consumption grew faster. So the “build” column now has a variable operating cost that many finance teams are still treating as a rounding error. I would be cautious about that. It is the kind of line that is small in month three and surprising in month fourteen.
37% report any enterprise-level financial impact from AI. Unchanged from last year.
So the tools are everywhere, individual productivity is up (80% say so), and the company-level number has not moved. This is the finding I find most relevant to the build-or-buy question, because it tells you where value actually comes from. It is not from having the capability. Almost everybody has the capability now. It is from what the organisation does around it: who owns it, how the workflow changed, whether anyone measures it.
That is the same for a built capability and a bought one. Buying does not save you from this problem. It only saves you the twelve months of building before you discover you have it.
High performers are 6% of the sample, and nearly three quarters of them fundamentally redesigned their workflows around AI. Among everyone else, one in four.
The high performers are also more likely to build, more likely to scale agents, and, interestingly, three times more likely to report being cost-constrained on coding agents. They are spending more, building more, and hitting the ceiling more. That is not a contradiction. It is what it looks like when a company has decided something is strategic and is prepared to pay for the control. The other 94% are mostly doing the opposite: light adoption, no redesign, and a surprised look when the impact does not arrive.
So what does all this add up to?
I think the honest reading is that the build-or-buy decision has become harder, not easier. Building is more accessible than it was, which makes it more tempting for things that were never strategic. Operating costs have a new variable component that most models ignore. And the evidence says the capability itself is rarely the differentiator; what you do with it is. Which means the question is no longer “can we build this,” because increasingly you can. The question is whether owning it will still be worth the cost when the novelty wears off and the maintenance begins.
None of that is comfortable when the capability in question is a single product. It gets much less comfortable when it is a function that runs around the clock, in every market you are licensed in, with real money at risk while it does.
That is the decision I will be modelling, with the hidden costs included, at the Acceler8 Lab at SBC Summit Lisbon. Join me for the session “The Trading Desk Question: Build In-House, Buy Managed, or Run Hybrid” on October 1, 11:00: https://gr8.tech/acceler8-lab/
Viktoriia Grygorenko, CEO at The Playa