How to Scope a Custom AI Agent Project

How to Scope a Custom AI Agent Project

A custom AI agent is software that carries out a multi-step task across your systems, such as reading an incoming request, looking up a record, drafting a response and updating your CRM. It can save a lot of time, and it can go badly if the scope is vague. Use these questions before anyone writes code.

Last reviewed: November 2026.

1. What is the one job?

Describe it in a sentence: "When a supplier emails an invoice, extract the details, match it to the purchase order and create a draft bill for approval." If you need several sentences, it may be several projects.

2. What triggers it, and what does it receive?

An email, a form, a file in a folder, a scheduled time. Collect twenty real examples of the inputs, including messy ones.

3. Which systems does it touch?

List every system it reads from or writes to, and check each has a way to connect, such as an API. Systems without one add cost and risk.

4. What may it decide on its own?

Be explicit: which actions it can take automatically, which need approval and which it must never take. See human-in-the-loop AI.

5. What happens when it is unsure?

It should stop and hand the item to a person, with context, rather than guess.

6. How will you know it works?

Pick two or three measures, such as hours saved per week, error rate on a sample, and time to complete the task.

7. What does it cost to run?

Build cost, plus hosting, AI model usage and maintenance. Usage costs rise with volume.

Do you need an agent at all?

If the task follows fixed rules, ordinary workflow automation may be cheaper and more reliable. Agents earn their place when inputs are messy and steps vary.

How we help

We scope agent projects with you, build and test them on your real cases, and set clear limits on what they may do. See custom AI agents.