How to scope an AI agent project in 2027: 9 questions to answer before you build

TL;DR

Most AI agent projects that stall were never scoped. Nobody wrote down the one job the agent does, what it is allowed to touch, or how anyone would know it worked. Answer the nine questions below before you open a builder, and the build gets smaller, cheaper in time, and much easier to trust.

Chatbot, automation or agent?

Teams often ask for an "agent" when a simpler tool would do. Pick the lightest option that solves the job.

If a fixed flow covers 90 percent of cases, build the flow and add AI only to the step that needs judgement.

The 9 scoping questions

1. What is the one job?

Write it as a single sentence with a verb and a finish line. "Read new inbound leads, score them against our ideal client notes, and post the top ones to the sales channel by 9am" is a job. "Help the marketing team" is not.

2. Who uses the output?

Name the person or team who reads what the agent produces. If nobody is waiting for the output, the agent will be ignored within a month.

3. What does it read?

List every source: inbox, CRM, ad accounts, spreadsheets, documents. For each one, note whether the agent needs read access only.

4. What can it change?

This is the most important question. Separate actions into three groups.

Start every agent in draft mode. Promote single actions to "act" only after you have watched it get them right.

5. What does "done" look like?

Define a check the agent or a person can run. A report has every section filled. A lead has a score and a reason. Without a check, you cannot tell a good run from a confident wrong one.

6. What happens when it is unsure?

Decide the fallback before launch: ask a named person, park the item in a review list, or skip it and log why. Silent guessing is the failure you want to rule out.

7. How often does it run?

On a schedule, on an event (a new form fill, a new email), or on request. Event triggers usually beat polling because the agent only wakes when there is work.

8. What will you log?

Keep a short record of each run: inputs, actions taken, anything skipped. Logs are how you debug, and how you build enough trust to give the agent more room.

9. How will you measure it after 30 days?

Pick one or two numbers tied to the job, such as hours saved per week, response time to new leads, or errors caught before a client saw them. Measure them before launch so you have a baseline.

A one page scope template

Fill this in with the people who will live with the agent. If a field stays blank, that is your next meeting, not your next build step.

Common scoping mistakes

1. Too many jobs in one agent. Split them. Two small agents are easier to test than one large one.
2. Write access on day one. Start in draft mode.
3. No owner. Every agent needs a person who reads its output and fixes its instructions.
4. No baseline. You cannot claim time saved if you never measured the time spent.
5. Building before the data is clean. An agent reading a messy CRM produces messy work faster.

Final thought

A good scope makes the agent boring in the best way: it does one clear job, asks when it is unsure, and leaves a record. That is what makes people keep using it.

I build AI agents, chatbots and automation systems for marketing and operations teams, alongside SEO, AEO and paid ads work. More at vickylalwani.net.