How to scope an AI agent pilot that your finance team will sign off

By Manoj Gurumurthy · 2026-10-06 · 6 min read

When an AI agent pilot stalls, the technology is rarely the reason. The reason is usually that nobody agreed, before it started, what it was supposed to prove. Then the pilot ends, everyone has an opinion, and the budget conversation goes nowhere.

Here is how we scope pilots so the decision at the end is easy, whichever way it goes.

Pick one conversation, not a department

"Automate customer support" is not a pilot. "Call back every new property inquiry within five minutes and book qualified buyers into a site visit" is. A good pilot covers one type of conversation, with one clear outcome, in one team.

Good first candidates share three traits:

They happen often, at least a few hundred times a month
They follow a predictable script most of the time
Being slow or missing them costs money you can point to

Lead call-backs, appointment reminders and order-status questions usually qualify. Complex complaints and negotiations usually do not.

Measure the baseline before you build anything

You cannot prove improvement against a number you never measured. Spend the first week of the engagement pulling the current figures:

How long does a new lead wait for first contact today?
What share of appointments are no-shows?
How many staff hours a week go into this conversation?

If the data is not in your CRM, sample it by hand for a week. A rough baseline is far better than none.

Agree the success threshold in writing

Before the pilot starts, write down the number that would justify a rollout. For example: "median first response under five minutes, with at least the current share of leads booking a visit." Also write down the number that would mean stopping.

This sounds bureaucratic. In practice it is the most useful paragraph in the whole project, because it turns the end-of-pilot meeting into a check rather than a debate.

Put the guardrails in the scope, not the small print

Decide up front what the agent must never do on its own: quote prices outside a range, promise refunds, discuss medical or legal details. Decide when it hands over to a person and what the person receives. Finance and compliance teams relax considerably once they see these rules written down.

Show the running cost, not only the build cost

A pilot proposal should show what the system costs to run each month at the pilot's volume and at full volume. The usual components are telephony or messaging fees, AI model usage, hosting, and any support plan. Running costs scale with usage, so show the cost per conversation as well as the total. That is the number finance will compare against the cost of the people currently doing the work.

Keep it short and real

Two to four weeks with real customers is usually enough to see a clear result. Shorter, and you are measuring the novelty. Longer, and people stop treating it as a test. Run it in your own accounts, so that if it works, the pilot simply becomes the first version of the production system rather than a demo you have to rebuild.

If you want to talk through a pilot for your team, our engagement model shows how we run them.