Home / Insights

Licences or your own model: where AI spend pays back

The choice between buying per-seat AI licences and building on your own model is presented as cheap-and-easy against expensive-and-hard. That framing hides the question that actually decides it, which is where your volume, your differentiation and your data sensitivity sit.

The short answer

  • Licences, the per-seat assistants and hosted APIs, are fast, predictable and require little to stand up. Cost scales with seats and usage.
  • A custom build, whether fine-tuning, retrieval over your own data, or self-hosting, costs far more upfront and gives control, fit and data residency.
  • The payback turns on three things: how much volume you run, how much the capability differentiates you, and how sensitive the data is.
  • High volume and a distinctive use case favour building. Broad, general productivity gains favour licensing.
  • The two are not exclusive. Licence for general uplift, build only where it earns its keep.

What each side really offers

Buying licences means paying for access to a capability someone else runs. The per-seat assistants embedded in office software, and the hosted model APIs, get you started in weeks, need little internal expertise, and carry a predictable per-user or per-call price. You give up control of the model, the roadmap and, to a degree, where your data goes, and your cost rises directly with how many people use it and how much.

Building on your own model means investing to make a capability yours: fine-tuning a model on your data, building retrieval over your own knowledge, or self-hosting so nothing leaves your environment. It costs far more before it produces anything, and it demands expertise you may not have. In return you get control of behaviour, fit to your specific task, and the ability to keep sensitive data inside your own boundary. The full weight of that commitment is set out in the line items nobody budgets, because a build is an operating commitment rather than a one-off.

The three questions that decide payback

QuestionPoints to licensingPoints to building
VolumeModest or uncertain usage, spread thinHigh, sustained volume on a narrow task
DifferentiationGeneral productivity everyone can buyA capability that is a genuine advantage
Data sensitivityLow; ordinary business contentHigh; data that must not leave your control

Volume is the clearest lever. Per-seat and per-call pricing is a bargain at low usage and an expanding liability at high usage, because you pay for every unit forever. A build carries a large fixed cost and a low marginal one, so at enough volume it undercuts the licence and keeps undercutting it. Somewhere between low and high usage the two cross, and knowing roughly where your volume sits against that crossing is most of the decision.

Differentiation decides whether building is worth the risk

Volume tells you whether a build is cheaper. Differentiation tells you whether it is worth the effort and the risk. If the capability is general productivity, the kind every competitor can buy off the same shelf, there is little to gain from owning it and a lot of complexity to take on, so licensing is the rational answer. If the capability rests on data or process only you have, and it changes what you can offer, then owning it is the point, and the same logic that governs build, buy or partner applies: build the advantage, licence the commodity.

Data sensitivity can settle it on its own

The third question sometimes overrides the other two. Where data cannot leave your environment for legal, regulatory or security reasons, the option to send it to a hosted service is closed regardless of cost, and self-hosting or a private deployment becomes the requirement rather than the preference. A financial or infrastructure operator with strict data residency may find the build decision is made by the data before volume or differentiation is even considered. Where the data is ordinary business content, this question falls away and the first two decide.

The middle path the framing usually misses

Buying and building are treated as the only two options, which quietly equates building with training or hosting your own model. There is a middle path that gets less attention and fits more organisations: buy the model, build the data layer around it. Retrieval over your own knowledge, where a bought model answers using your documents and records supplied at the point of the query, gives much of the fit and relevance of a custom build without the cost and risk of training a model from your data. You licence the capability that is a commodity, the language model, and build the part that is genuinely yours, the connection to your own information.

This reframes the decision usefully. The more useful question is whether to build the data layer that makes a bought model useful on your specific task, since few organisations should train a model of their own at all. That build is smaller, cheaper and lower-risk than training or self-hosting, and it keeps the advantage, your data and how you use it, on your side of the line while leaving the commodity where it belongs.

The constraint

The binding limit is an honest read of your real volume, differentiation and data sensitivity, rather than the headline price of a licence against a build. Committing to a custom model for a general, low-volume use case spends heavily to own a commodity, and licensing a high-volume, differentiating, sensitive workload leaves advantage and control on the table, because the deciding question is which side of those three tests you are actually on, and only that answer tells you where the spend pays back.

The pattern most enterprises land on

For most organisations the answer is both, targeted. Licence broadly for the general productivity uplift, where speed and low commitment win and the usage is spread across many people doing many different things. Build narrowly, on the one or two workloads that are high-volume, genuinely differentiating, or too sensitive to send out. Spending to build everything wastes money on commodities, and licensing everything caps you at the capability your competitors can also rent. The discipline is to know which of your workloads earns a build, and to resist building the rest.

Sources

  1. RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 2024, on cost against delivered value.
  2. Gartner press releases on generative and agentic AI project outcomes, 2024 and 2025, on cost and value as decision drivers.