AI consulting that starts with your data, not a model
Most AI consulting starts with a technology recommendation. Ours starts with a less enjoyable question: can your data actually support the thing you want to build. Answering that in week one is a great deal cheaper than discovering it in month four.
Noise resolving into a bearing
The problem
Where AI projects actually stall
It is rarely the model. The usual cause is a roadmap agreed before anyone checked whether the data underneath could carry it. A support assistant needs support history that somebody wrote down. A document assistant needs documents that are current, rather than five conflicting versions in a shared drive.
The second cause is cost arriving late. A per-request price that looks trivial in testing turns into a real line item at production volume, and by then the architecture responsible for it is expensive to unpick.
You might recognise
A pilot that impressed everyone
and then stopped moving for three months
Nobody can say what it costs
per request, at the volume you actually expect
Quality is argued, not measured
there is no test set anyone trusts
A quote you cannot evaluate
because the technical assumptions are unstated
Method
How we run a consulting engagement
We work from a representative sample of your real data, not a curated one. We build a small evaluation set early, because without it there is no honest way to compare two approaches. Then we cost the system at the volume you expect, including the retries and long context windows that nobody budgets for.
If the conclusion is that you should not build it, we say so and explain why. That is a legitimate result and a much cheaper one than finding out later.
What you leave with
- A written technical assessment with its assumptions stated plainly
- An evaluation set built from your data, which is yours to keep
- Cost and latency modelling at your expected volume
- A costed delivery roadmap, or a clear recommendation not to proceed
Questions
Frequently asked
- How long does an AI consulting engagement take?
- A discovery sprint runs two to four weeks. That is normally enough to establish viability, build a first evaluation set and produce a costed roadmap. Anything longer is build work rather than consulting.
- Do you recommend building or buying?
- It depends whether the capability is core to your product. If an off-the-shelf tool covers most of it and the remainder is not a differentiator, buy. We have told clients to buy.
- What do you need from us to start?
- A representative sample of real data, one person who knows that data well, and a statement of what success looks like in numbers. The third is usually the hardest to produce and the most useful.
Also from us
- AI integration servicesConnecting a model to your existing stack without breaking your permission model.
- AI agent developmentTool-using agents with real boundaries, human approval where it matters and a full trace.
- Hire AI developersSenior engineers inside your team, committing to your repo and teaching as they go.
Not sure whether it is worth building?
That is the question a discovery sprint exists to answer. Tell us what you are considering and what is currently in your way.