A feature on the few who should build
Most should buy. A few should build.
The honest default with AI is to buy, not build. The vast majority of needs are met by configuring mature models and tools that already exist, and writing your own from scratch would be like building your own spreadsheet software — expensive reinvention of a solved problem. But there is a genuine minority where the opposite is true: where the capability is core to your competitive advantage, your own data is the asset, and the tools everyone else uses simply cannot do what you need. This page is about that minority — custom models, built on your data, for your edge, when build honestly beats buy.
An ML engineer training a bespoke model on proprietary data, a performance curve climbing past the off-the-shelf baseline marked on the chart. The moment custom genuinely beats generic. Square aspect ratio, dark studio, focused light, solar-yellow graph glow, a sense of rigour and edge.
Build, not buy — for the few cases where a bespoke model on your data is a genuine, defensible advantage.
It is worth saying plainly, on a page selling custom models, that most businesses should not commission one. The off-the-shelf models and tools available today are extraordinary, cheap relative to building, and improving constantly — and for the overwhelming majority of needs, configuring them well is not a compromise, it is simply the right answer. A firm that builds its own model for a problem the market has already solved has bought itself cost, delay and a maintenance burden in exchange for nothing. The first thing we tell most people who ask for a custom model is that they don't need one.
But the minority for whom the answer flips is real, and for them it matters enormously. There are three signs together. First, the capability is genuinely core to your competitive advantage — not a back-office convenience, but something customers value and rivals would love to copy. Second, you hold data that is yours and yours alone, which is the true raw material of a model others can't replicate. And third, the generic tools demonstrably can't do the job well enough — you've tried, and the off-the-shelf result falls short of what the opportunity demands. When all three hold, a custom model stops being indulgence and becomes a moat.
"Custom" also rarely means "from scratch," and conflating the two is how budgets get burned. Training a large model from nothing is the province of a handful of labs with vast resources, and almost never the right move for a business. The practical path is to take a strong existing foundation and adapt it to your problem and your data — fine-tuning it on your examples, or grounding it in your proprietary knowledge — so you inherit the general capability and add the specific edge. That is where the genuine, defensible advantage lives: not in rebuilding the engine, but in teaching a very good one to do your particular job better than anyone else's.
In this feature
Six things serious custom-model work insists on.
Build only when you should
The honest test comes first: is this core to your edge, is the data yours, and have off-the-shelf tools genuinely fallen short? If not, we say buy. We will talk you out of a custom model more often than into one.
Your data is the moat
A custom model's real advantage isn't the algorithm — it's the proprietary data only you have. We treat that data as the asset, building the model around what makes it uniquely yours and impossible for a competitor to simply copy.
Beat the baseline
A custom model has to clearly outperform the best off-the-shelf option to justify its cost. We measure honestly against that baseline — and if the generic tool gets close enough, we tell you to use it instead of building.
Adapt, don't reinvent
"Custom" almost never means "from scratch." We adapt strong existing foundations to your problem and data — fine-tuning, grounding — so you inherit world-class capability and add your specific edge, without the cost of rebuilding the engine.
Evaluation & honesty
A model you can't measure is a model you can't trust. We define how success is judged before we build, evaluate rigorously, and report what the model can and can't do — including where it fails — because confident, untested AI is a liability.
Owned and maintained
A model is not a one-off delivery; it drifts as the world and your data change. We build it to be yours, documented and maintainable, with a plan for monitoring and retraining — so the edge it gives you holds rather than quietly erodes.
When off-the-shelf
won't fit.
There is a romance to building your own AI that gets a lot of businesses into trouble. It sounds impressive, it feels like ownership, and it flatters the ambition of the people commissioning it. But the market for general AI capability is now extraordinarily good and extraordinarily cheap, and competing with it by hand is a fight almost no one should pick. The discipline of this work begins, perversely, with restraint: a willingness to look a client in the eye and say that the thing they find exciting is the thing they should not do, because a configured off-the-shelf tool will serve them better, sooner and for a fraction of the cost.
What flips the calculation is data. The general models everyone can access were trained on broadly public information; what they have never seen is your proprietary data — your years of transactions, your specialist domain corpus, the labelled examples only your business has accumulated. That data is the one ingredient a competitor cannot simply buy or download, and it is the true raw material of a model that does something the generic tools can't. When a problem turns on knowledge or patterns that live exclusively in your data, a custom model built around it is not reinvention; it is converting an asset you already own into a capability nobody else can match.
The honest discipline that holds the whole thing together is evaluation against a baseline. Before building anything, we establish how well the best available off-the-shelf option already does the job — and a custom model has to clearly, measurably beat that baseline by enough to justify its cost and upkeep. This is the test that protects you from the seductive project that produces a technically-impressive model nobody needed. Sometimes the honest result of the evaluation is that the generic tool, suitably configured, is good enough — and when it is, we say so and stop, because a custom model that wins by a margin too small to matter is a loss dressed as a triumph.
The serious version of custom-model work treats the model as a long-term capability, not a deliverable. A model is trained on a snapshot of the world and your data, and both move on — customer behaviour shifts, the domain evolves, new data accumulates — so a model left untouched slowly drifts away from the reality it was built for. We build models to be owned: documented, maintainable, monitored for that drift, and retrained on a sensible cadence. The edge a custom model gives you is real, but it is not permanent by default; keeping it sharp is part of the work, and a model handed over and forgotten is a depreciating asset, not a moat.
A model evaluation chart — a custom model's curve clearly above the off-the-shelf baseline on a domain-specific task, proprietary data feeding it. The proof that build beat buy. Wide cinematic 21:9 crop, dark elegant studio backdrop, solar-yellow accent, a sense of rigour and advantage.
Adapted, not reinvented — a strong foundation taught your data to beat the baseline, and turned into a moat.
Five questions we ask before building a single model.
The first thing we tell most people who ask for a custom model is that they don't need one. The model is never the moat — the data is. The model just turns it into one.
Custom models sit at the deep end of the Artificial Intelligence pillar, reached only when the shallower options have been honestly ruled out. AI strategy is what makes that build-versus-buy call in the first place; a custom model often becomes the engine inside an AI automation or a personalisation system where the generic approach fell short; and the same data discipline underpins them all. It is the most specialised and least frequently-needed discipline in the pillar — which is exactly why the judgement of when to reach for it matters more than the ability to do it.
We approach custom models with a bias against building them, a respect for your data as the real asset, and an honest baseline that any model has to beat before it ships. The brief is a genuine, defensible edge — not an impressive model nobody needed. That combination of restraint, data-first thinking and rigorous evaluation is exactly why custom-model work, done properly, is rare: most should buy, a few should build, and the whole value is in knowing — honestly — which one you are.
A chart showing a custom model's accuracy rising clearly above the off-the-shelf baseline on a niche task, fed by a proprietary dataset. The moment build out-performs buy. Contemporary, shallow depth of field, dark desk, solar-yellow screen glow.
Off-the-shelf couldn't; their data could — a custom model that became a real, defensible edge.
Representative scenario · not a named client engagement
A specialist firm kept hitting the limits of generic AI on its niche problem — until a model built on its own data cleared them.
The firm operated in a specialised domain and had a high-value task — a classification and judgement problem at the heart of its service — that the general AI tools handled disappointingly. They had tried the off-the-shelf options, and each one stumbled on the niche, expert nature of the work it had never been trained for. What the firm did have, accumulated over years, was a large body of its own expertly-labelled examples: precisely the data the generic models lacked, and precisely what a competitor couldn't obtain.
We started, as always, by confirming the case to build was real — it was: core to their edge, data uniquely theirs, generic tools genuinely short. Then we adapted a strong foundation model to their proprietary dataset, fine-tuning it on their expert examples rather than building from scratch. We evaluated it rigorously against the best off-the-shelf baseline throughout, and it cleared that bar by a wide, decisive margin on their specific task — doing what no general tool could, because no general tool had their data.
The custom model comfortably beat every off-the-shelf option on their task — a capability rivals simply couldn't replicate.
The advantage proved durable precisely because it rested on the data, not the algorithm. Anyone could license the same foundation models; nobody else had the firm's years of labelled expertise to adapt them with, so the edge couldn't be copied by simply buying the same tools. We built the model to be owned and maintained, with monitoring and a retraining cadence to keep it sharp as their data grew. For this firm, the custom model was exactly the right call — which is notable only because, for most firms who ask, it isn't. The difference was the three conditions genuinely holding.
The generic AI tools just couldn't do our specialist task, but our own data could teach a model that did. Revolutionize confirmed it was genuinely worth building, then built it — and it beat everything off-the-shelf. They were also clear it's the right call for very few companies. We happened to be one of them.
What serious custom-model work actually involves.
A custom model is an investment in a defensible capability, and it is scoped as a serious build with ongoing ownership. Engagements range from fine-tuning a foundation model on your data for a focused task to developing and maintaining a bespoke capability central to your product, and the scope rises with the complexity of the problem, the state of your data, and the evaluation and maintenance the work demands.
The two largest variables are the state of your data and the maintenance the model needs. A clean, well-labelled dataset and a contained task is one thing; messy data needing preparation, or a model that must be monitored and retrained continuously, is another. We always begin with the honest build-versus-buy evaluation — and a meaningful share of those conversations end with us recommending you don't build at all, which costs you far less than a model you didn't need.
Every engagement includes the full discipline: an honest build-versus-buy evaluation, data assessment and preparation, adaptation of a strong foundation rather than wasteful from-scratch training, rigorous measurement against the off-the-shelf baseline, and a plan for ownership, monitoring and retraining — so the edge is real and it lasts.
We scope every model to the problem rather than to a price list — which is why we don't publish rate cards. Every engagement begins with a free 30-minute scoping conversation, and we will tell you honestly whether you are one of the few who should build, or — far more often — one of the many who should buy and configure instead. We would rather lose the project than sell you a model you don't need.
When you're ready
When buying isn't enough.
Tell us the AI capability you need and the data you hold — and where the off-the-shelf tools have fallen short. We'll respond within 24 hours with an honest read on whether this is genuinely a case to build, or — far more often — one where configuring an existing tool would serve you better.
Begin the conversation →