How to scope an AI engagement from pilot to production

Scope an AI engagement from pilot to production on one page: what live means, the systems involved, acceptance targets, owners and gates. A buyer's guide.

By Published 3 min read

Scope an AI engagement from pilot to production on one page, written before work starts: what “live” means, the systems it touches, the acceptance targets, the named owners, what is out of scope, and the gates where you decide to continue, change course or stop. Scope by outcome and by gate. A list of features alone does not protect you.

This guide is for the person who has a working pilot, a budget conversation ahead, and a vendor, or an internal team, waiting for a brief.

Key takeaways

  • Start with what live means, not with what to build.
  • Acceptance is a small set of targets agreed in advance.
  • Each phase ends with a gate: continue, change course or stop.

1. Define production for this system

Write two or three sentences. Which users, on which live data, doing which task, under which rules? “The claims team uses it on every new case, and a person approves any change to a customer record” is a definition. “The AI works well” is not.

Anything the pilot did on sample data or by hand belongs here as an explicit gap.

2. List the systems it must touch

Name every system the AI reads from or writes to, and who owns each. Integration is where most of the time goes, because every system brings an owner, a data format and a change process. Systems left off this list return later as change requests.

3. Set acceptance targets

Choose a handful of measures that reflect what users need, and agree a target for each before work starts. Google’s SRE book recommends keeping the number of objectives small, starting from what users care about and not from what is easy to measure, and avoiding absolutes such as “always available” (Service Level Objectives, Google SRE book). For an AI system that usually means accuracy on a test set built from real cases, response time, and how often a person must step in.

Agree how the measure is taken and by whom. A target nobody can measure is a disagreement waiting to happen.

4. Map the risks and name the owners

The NIST AI Risk Management Framework organizes the work into four functions: Govern, Map, Measure and Manage (NIST, AI Risk Management Framework). You do not need the full framework to borrow its shape. Write who is accountable (Govern), what the system could get wrong and for whom (Map), how you will tell (Measure), and what happens when it does (Manage).

Name one business owner and one technical owner. They sign the scope and accept the result.

5. Write what is out of scope

New use cases, new user groups, a second language, a model change and a redesign of the process around the AI are the usual items. Writing them down is how you keep the promise clear and still welcome a second phase.

6. Add gates

Split the work into phases that each end in a decision with evidence on the table. A typical shape is: confirm the definition and systems, build and integrate, test against the targets next to today’s process, then release and hand over with a runbook. At each gate someone with authority says continue, change course or stop. A gate that can only say continue is a milestone.

How long will it take?

Count three things: the systems from step 2, the reviews and approvals the system must pass (security, legal, the people who approve its outputs), and the decisions still open. Time follows those counts, not the model. Ask any vendor to show how their estimate follows them.

What should the vendor give you back?

A written response to each of the six sections, with the questions they could not answer yet. A response that skips sections 3 to 6 is a quote for effort, not a plan for production.

Where OrNsoft fits

Bring us the problem, not the specification. Our forward-deployed engineers work directly with your business and technical teams, learn how your operation actually runs, determine what has to be built, and move the system from pilot to production. You explain the objective once.

OrNsoft is an advanced AI engineering firm, founded in Florida in 2009, with applied AI in its software since 2015. If you want to scope your pilot with us, open the contact page and tell us the problem. One call is enough to know whether we can help.

Questions

What is the first thing to write when scoping an AI engagement?
The definition of production for this system: what live means, on which data, for whom and under which rules.
Should scope be a list of features?
A feature list helps, but the scope that protects you is outcomes with acceptance targets and gates, because AI systems change as they meet real data.
How do I estimate how long it will take?
Count the systems the AI must connect to, the reviews and approvals it must pass, and the decisions still open. Time follows those, not the model.
Who should sign the scope?
The accountable business owner and the accountable technical owner, both named.

Bring us the problem.

Book a call