Two calls, and you will know where your AI actually stands.

You describe the problem and we ask about it for an hour. Not a qualification script — we are trying to establish what you are trying to move, what you have already tried, where you sit on your own roadmap, and where the process around the work is leaking. It helps enormously if someone who does the work day to day is on the call alongside whoever owns the budget. You do not need to prepare anything, and there is no deck from our side.

No, and this is the most common shape of a first call. A large part of what the report says is usually which candidates are worth doing at all and in what order — and for perhaps a third of the organisations we talk to, the honest first answer is that one or two of their problems are not AI problems and would be better solved another way. Hearing that in week one is worth more than a roadmap that has you starting in the wrong place.

Both, and deliberately so — separating them is how a lot of AI programmes stall, with a strategy document on one side and nobody accountable for making it real on the other. The advisory work produces the sequence, the cost model and the benchmarks; the delivery work is our engineers building it alongside your team. You can buy the first without the second, and it is priced separately, so a recommendation is never a route into a build contract.

From your side of the business, in the second call, before anything is built. A useful benchmark is almost always a number you already report and already argue about — how long a case takes to close, how often something has to be reworked, what proportion of a queue clears without a person touching it. Model-level scores are a diagnostic we use internally, but they are not what we ask you to judge the engagement on, because they can move without anything in your business moving at all.

Usually, and often substantially — but we would want to check it is the right thing to fix first. The first production version of an AI workflow is nearly always more expensive than it needs to be: oversized context, the expensive model used everywhere, no caching, and no attribution of spend to the workflow that caused it. That is straightforward work. What we look at alongside it is quality, because a system that returns the wrong answer often enough costs far more in checking and rework than it ever costs in tokens, and cutting the model bill while that is true makes the real problem harder to see.

Yes. The same workflow deploys three ways — inside your own cloud tenant, on-premise on hardware you own, or fully isolated with no external network at all. We currently run seven fully isolated deployments. What changes between them is where the model runs and what, if anything, crosses your boundary; the access controls and the audit record are identical in all three. If your procurement or security team wants the detail behind that, we will walk them through a real one rather than describe it.

Practitioners on the first call, not account managers. The engineers who would do the work are the ones who scope it and the ones who write the report, and they stay on the engagement afterwards. Nobody is handed over to a delivery team they have never met after the contract is signed.