Scope, acceptance criteria and the fee, fixed in writing before anyone starts: which workflow, which systems, what counts as done, and what it costs, quoted on application before you commit. A fixed arc helps. A short sprint that baselines one workflow, then a build with a hard finish, makes the price of delivery the provider’s problem, not yours.
Approved partner · Anthropic & OpenAI
How to Choose aProduction-Agent Consultancy.
Five criteria that predict production
A first-call checklist inside
S.01 · The Short Answer
Judge the consultancy, not the demo.
What actually predicts production?
Five criteria: fixed scope and a fixed fee; a definition of production written down: real traffic, changes gated by evaluation results, monitoring, governance; IP and ownership from the first commit; an evaluation harness treated as a first-class deliverable; and an embedded delivery pod, not a slide deck.
Most selection goes wrong the same way: the shortlist is built on brand, and the contract is signed on scope nobody can enforce. The five criteria above are checkable on a first call, before you have spent anything. Nugawi is an applied-AI consultancy in London working across the UK, Europe and the US, so weigh the source; the criteria hold whoever you hire.
- 01
Fixed scope, fixed fee.
- 02
Production, defined in writing.
- 03
IP and ownership from the first commit.
- 04
Evaluation as a first-class deliverable.
- 05
Embedded pods, not slide decks.
If you are still deciding between firm types, our comparison of Claude consultancies and generic AI consultancies covers who does the work; this page covers what predicts production. Run both.
S.02 · Criterion 01
Fixed scope, fixed fee.
Day-rate metering dressed up as discovery, phases without end, a proposal that prices the assessment but goes quiet on the build, and scope that moves after signature with change orders as the business model. Those engagements are built to grow into the budget.
S.03 · Criterion 02
Production, defined in writing.
A definition of production in the statement of work: the agent running on real traffic, changes gated by evaluation results, monitoring and tracing, a named owner, and governance shipped with the system rather than promised after it. If it is not written down, it is not a definition; it is a mood.
A provider whose production means the demo worked, a pilot with no acceptance criteria, or a roadmap where production is always the next phase. Ask what, specifically, will be running on live traffic by the end. Vague is an answer.
S.04 · Criterion 03
IP and ownership from the first commit.
Code, prompts, evaluation suites and infrastructure as code committed to your repositories from day one, deployed into your accounts, with a written exit path. A consultancy that keeps winning workflows on merit has no reason to hold your system hostage.
Vendor-held repositories, proprietary platforms you license back, frameworks your own engineers cannot run, ownership that starts at final invoice. Where the first commit lands decides who owns the negotiating position.
S.05 · Criterion 04
Evaluation as a first-class deliverable.
An evaluation harness built from your real cases: task-level suites run on every change, quality gates wired into the pipeline, and a day-zero baseline so improvement is measured, not asserted. Ask to see a suite the consultancy built and what it gates today.
“The demo looked good.” No regression suite, no baseline, no monitoring in the plan. A consultancy that cannot show evaluation evidence is selling you its own confidence, and confidence is not an artefact you can audit.
S.06 · Criterion 05
Embedded pods, not slide decks.
Senior engineers who embed with your operators and ship into your real systems: a named pod, a strategist who owns the workflow’s economics, and capability transfer so your team ends up owning what lands. The engagement should leave your people more capable, not more dependent.
Rotating teams that hand over documents and bill for workshops, sellers who vanish after signature, a model where the people who scoped the work are never the people who do the work. If nobody embeds, nothing transfers.
S.07 · The First-Call Checklist
Questions to ask on the first call.
Ten minutes of these questions separates production partners from presentation partners. Ask for the answers in writing.
- 01
What is fixed before we start: scope, fee, acceptance criteria? Show me where in the contract.
- 02
What does production mean here, in writing: real traffic, changes gated by evaluation results, monitoring, governance?
- 03
Where does the first commit land, and whose account is it in?
- 04
Show me an evaluation suite you built. What does it gate today?
- 05
Who is on the pod, how senior, and how much time do they spend with our operators?
- 06
What did your last production deployment measure against its day-zero baseline?
- 07
What does handover look like: repositories, operational runbooks, evaluation suites, and who can run it without you?
S.08 · FAQ
Choosing a production-agent consultancy: FAQs
The questions buyers ask us most when they are shortlisting.
A named workflow, written acceptance criteria, an evaluation harness, a deployment plan with monitoring and governance, and a fee fixed and quoted before the work starts. Fixed scope makes delivery the provider’s problem: the deliverable is a production system, not a document.
Agents doing real work on live traffic, with changes gated by evaluation results, monitoring and tracing, a named owner, and governance shipped with the system. A pilot with none of those is a demo with a schedule.
You do, from the first commit: code, prompts, evaluation suites and infrastructure as code in your repositories, deployed into your accounts, with a written exit path. Ownership agreed in the architecture is a design constraint; ownership negotiated at the end is a renegotiation.
Test the five criteria: fixed scope and fee, a written definition of production, IP from the first commit, evaluation as a first-class deliverable, and embedded delivery pods. For the firm-type angle, our comparison of Claude consultancies and generic AI consultancies covers who does the work; this page covers what predicts production.
S.09 · Test Us
Run the checklist on us.
Nugawi is an applied-AI consultancy in London, serving the UK, Europe and the US. Our engagements run fixed scope and fixed fee, quoted on application, and our builds put the first consequential workflow into production in 90 days. We would rather you ran the checklist on us than took our word for it. Start with a Discovery Sprint, or point an Agent Audit at what you already run.