Skip to content

How to scope an AI agent before anyone writes code

Scope an AI agent by writing down seven things: the job it does, the inputs it reads, the actions it may take, which actions need approval, how it can fail, the real examples it will be tested on and the checks that decide whether it is finished. A one-page scope prevents most expensive surprises.

Why scoping decides whether an agent works

Most agent projects that disappoint were not short of AI. They were short of a clear definition. “An assistant for the office” cannot be tested, priced or finished. “An agent that chases clients for missing year-end documents, logs what arrives and flags anything unusual to the partner” can.

A good scope fits on one or two pages and answers seven questions:

  1. The job: what task, triggered by what, finished when?
  2. Inputs: what does the agent read, and where does it come from?
  3. Actions: what can it do in which systems?
  4. Approvals: which actions need a person first?
  5. Failure modes: what can go wrong, and what should happen then?
  6. Test cases: which real examples will it be tested against?
  7. Acceptance checks: what results mean it is ready?

You can draft most of this yourself before speaking to any developer. It is also the core of the discovery stage in our process.

Step 1: Define one job in one sentence

Write the job as a sentence with a trigger, a task and a finished state: “When a new enquiry arrives on WhatsApp, collect the customer’s details and needs, and book a call with the right adviser.” If you need “and also” more than once, you probably have two jobs. Pick the one with the most volume or the most pain.

Then record how the job is done today: who does it, how often, how long it takes, and the steps they follow, including the unwritten ones. Sit with the person who does it. The exceptions they handle without thinking are exactly what the agent will trip over. This record also becomes your baseline for measuring return later.

Finally, name an owner on your side: the person who answers questions during the build, approves the test results and decides what the agent should do in cases nobody anticipated. Without one, the developer ends up guessing at business rules, and those guesses are the most expensive kind of change to fix after launch.

Draft your brief as you read.

Scope builder

Three questions, one short brief.

Nothing you type leaves this page unless you copy it.

Your brief
Your answers appear here as a brief.

Steps 2 and 3: List the inputs and the actions

Inputs are everything the agent needs to read: the incoming message or email, customer records, a price list, a policy document, a calendar. For each, note where it lives, whether it is accurate and up to date, and who maintains it. An agent answering from an out-of-date price list will be confidently wrong. Note any personal data too, and whether the agent actually needs it; see GDPR and AI agents.

Actions are what the agent can change. Be specific about the system and the level of access, because that drives both the build effort and the risk.

Describing actions at the right level
LevelExample
ReadLook up a customer’s next appointment in the calendar.
DraftPrepare a reply or an invoice reminder for a person to send.
Act within limitsReschedule a booking at least 24 hours ahead with the same practitioner.
Act freelyRarely appropriate in a first release.

For each system, check that it can be reached through an API or integration and what permissions your plan allows. Our integrations guide covers the common questions.

Steps 4 and 5: Decide approvals and plan for failure

For every action, decide whether the agent can do it alone, needs approval, or must hand the whole request to a person. Name the approver, the channel they approve in and what happens if nobody responds. Our guide to human-in-the-loop design goes further.

Then list failure modes. The common ones are:

  • The request is ambiguous or outside the agent’s job.
  • The input is missing, wrong or contradicts the records.
  • A connected system is slow, down or returns an error.
  • The customer is upset, vulnerable or asking for something sensitive.
  • Someone tries to manipulate the agent through the text it reads.
  • The agent completes the task, but the result is wrong.

For each, write the intended behaviour. Usually it is: stop, tell the person what happens next, and pass the request to staff with a summary.

Steps 6 and 7: Collect test cases and agree acceptance checks

Test cases should come from real history: past emails, messages or documents, with personal details removed where possible. Include the routine cases, the awkward ones your team remembers, and a few deliberately hostile ones. For each, write the expected outcome, such as “books Tuesday 10:00 with Dr Byrne” or “hands off: complaint”. This set is used before launch and again after every change.

Quality matters more than quantity. A set that covers every request type, every handoff rule and each failure mode you listed is more useful than a large pile of near-identical routine cases. Add new cases whenever the agent gets something wrong in live use, so the same mistake is caught next time.

Acceptance checks turn “it works” into statements you can verify, for example:

  • Handles every routine test case correctly, and hands off every case marked for a person.
  • Never takes an action outside its permitted list during testing, including the hostile cases.
  • Records every action with enough detail to reconstruct what happened.
  • Replies within an agreed time under normal load.
  • Runs alongside staff for an agreed trial period with corrections reviewed weekly.

Agree these before the build. They tell both sides when the work is done, and they protect you from a demo that only shows the easy cases.

What does a finished scope look like?

That scope is specific enough to estimate, build and test. It also shows what is out: the agent does not prepare accounts or give tax advice. See accountants for similar jobs in that sector.

With a scope like this, a developer can give you a meaningful quote; the AI agent cost guide explains what that quote should contain. If you would like help writing one, book a free review.

Common questions

How long does scoping an AI agent take?

For a single, well-understood job, scoping is usually a short piece of work compared with the build: a few conversations, a look at the systems involved and a written document. It takes longer when the process is undocumented, several teams are involved or the systems’ access is unclear. Time spent here is almost always cheaper than changing direction mid-build.

Do I need technical knowledge to scope an agent?

No. The job, the exceptions, the approvals and the test cases come from the people who do the work today. The technical questions, such as API access and permissions, can be answered by your developer. What helps most is access to real examples and someone who knows how the task is actually done, not how the manual says it is done.

What if the scope changes during the build?

It often does, because building reveals things. Treat changes deliberately: write down the change, its effect on cost and timeline, and agree it before work continues. Small clarifications are normal. A new job, a new system or a new type of action is a new piece of scope and should be priced and tested as one.

Should the first agent do several jobs to justify the cost?

Usually not. One job done reliably is worth more than four done unreliably, and it gives you real data on time saved and error rates. Once the first job runs well, adding a second is cheaper because the integrations, logging and approval patterns already exist.

Bring us the task.

Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.

Start a project.

Two people on a sofa with laptops, planning a project beside a whiteboard of sticky notes

Book a call or send us a message. We read everything.

tash@neu.ie
Working with teams in Ireland and worldwide.

Get in touch

Free 30-minute AI review.