Generative AI consulting, as we practice it, is an engineering engagement, not an inspiration session. The questions that decide whether your system works are unglamorous: which model, chosen by measured performance on your data rather than by brand; what the model is allowed to read and to touch; what a correct answer is and how you would know; and what each call costs once the volume is real. Get those four wrong and no amount of prompt polish recovers it.
The failure we are built to prevent has a familiar shape. A pilot demos beautifully — clean inputs, a friendly audience, ten hand-picked examples — and then meets production: the scanned PDF at an angle, the customer who writes in two languages, the record that contradicts the CRM, the question the documents do not answer. The pilot was never tested against any of that, so nobody finds out until it is live and wrong in front of a customer.
So we invert the order. Before rollout, we define what “it works” means — pass and fail, written down, measured on a sample of your real data including the ugly cases. A system that has not been evaluated on your data has not been evaluated; it has been rehearsed.
If a generative model is the wrong tool for your problem, we will say so on the call. A decent share of what gets pitched as a GenAI project is a lookup, a rules engine, or a form redesign — cheaper, faster, and it never hallucinates.
Scope
Which model, which size, hosted or API, is decided by running candidates against your task and your data and comparing the results — accuracy, latency, cost per call. Not by which vendor's keynote was most recent. That comparison has a habit of landing on a smaller, cheaper model than the one you walked in asking for — and sometimes on a different model per step.
A model answering from its training data will answer confidently about your business and be wrong. Retrieval pins its answers to your documents — contracts, policies, product data — with citations back to the source, so a wrong answer is traceable to a wrong or missing document instead of being a mystery. The unfashionable work is here: parsing, chunking, and deciding what happens when the documents genuinely do not contain the answer.
An agent that can act is an agent that can act wrongly, at machine speed. Anything irreversible — money out, records changed, a customer contacted — passes a human gate. Every action is logged with the inputs that produced it, and the allowed-action list is written down before the agent exists, not discovered after an incident.
We build the test set from your real traffic — including the malformed, the ambiguous, and the adversarial — and agree the pass bar in writing before anything ships. The same set runs again after every prompt or model change, so “did the update make it worse” is a measurement, not a feeling.
Per-token pricing means the architecture is the bill. Caching, routing simple cases to cheap models, keeping deterministic steps out of the model entirely — these are day-one design decisions. A system designed without a cost budget demos fine and then produces an invoice that ends the project.
Ours run from the decision to the handover: whether a generative model is the right tool at all, model selection measured on your data, retrieval over your documents, guardrails on anything the system can do to the world, an evaluation set with a written pass bar, and a runbook your team can operate without us. Generative AI consulting services that end at a recommendation leave you holding the hard half — this engagement ends when the system is in production and you can run it.
That page covers the wider question — where AI pays in your business at all, which is sometimes answered by a report, a rules engine, or a tool you already own. This page is the generative-specific build: models, retrieval, agents, evaluation, cost. If you already know the problem is generative — documents in, drafts or decisions out — start here. If you are not sure, start there; the call is the same either way.
By duration of need. A generative AI consulting company makes sense for getting the first system scoped, built, and evaluated — a bounded engineering problem with an end date. A hire makes sense when the work is permanent. The honest failure mode in both directions: a consultancy that stays forever, or a hire made before anyone knows what the job is. Our engagements end with a handover, and the runbook exists so you can leave.
Usually no, and the sequence matters because fine-tuning is by far the most expensive option to build and to maintain. Prompting and retrieval cover most problems that get described as fine-tuning problems — a model that lacks your knowledge needs retrieval, not training. Fine-tuning earns its cost for a narrow, high-volume task where the evaluation set proves prompting has hit a ceiling. That proof comes first: fine-tuning without a measured baseline is spending money to make an unmeasured number move.
The one that measures best on your task, your data, and your constraints — cost, latency, data residency, hosting. We hold no reseller position and no allegiance, so the answer is allowed to be boring: often a mid-tier model with good retrieval beats a frontier model without it. We also design so the model is swappable, because whichever is best this quarter may not be next quarter.
No one can, and a vendor claiming otherwise is telling you something about themselves. What engineering can do: constrain answers to retrieved documents with citations, return “not found” instead of a guess when retrieval comes back empty, keep the model out of arithmetic and record-keeping, and gate anything irreversible behind a human. The design goal is not a model that is never wrong — it is a system where a wrong model output is visible, traceable, and cheap.
We will not quote a timeline before scope exists, because a date given before scope is a date that moves. What we can say about sequence: evaluation criteria and the test set come first, the build second, rollout last — and the first two are where thin projects cut corners. The scope document you sign carries the delivery date; nothing starts before that document and the deposit.
Founder · he answers these himself
Fixed fee, scoped in writing before anything starts. No hourly billing, no verbal scope.
Related: AI consulting · LLM & RAG integration · AI agent development · AI workflow automation