How to pick your first AI agent: the four-question test

Most companies that try AI agents fail in the same way: they pick a first project that is too big, too vague or too risky. The agent half-works, nobody trusts it, and the initiative quietly dies.

The first agent matters less for the hours it saves than for what it proves. It should show your team that agents work, that they're safe and that they make people's jobs better. Here is the test we use with every client to find that project.

1. Does it happen at least 50 times a week?

Volume is what turns a small time saving into a meaningful one. A task that takes ten minutes and happens five hundred times a month is worth more than a task that takes two hours and happens twice.

Volume also gives you something else: lots of examples. Past tickets, emails and records are the raw material you'll use to design, test and measure the agent.

2. Could you explain it to a new hire in a page?

If the process lives only in one senior person's head, or changes every week, it's not ready for an agent yet. The best first candidates already have rules, even informal ones: "Refund automatically if it's under 30 days and unused; otherwise ask the customer for photos."

Writing those rules down is often the most valuable part of the project — even before any AI is involved.

3. Is a mistake cheap and fixable?

Your first agent will make mistakes, just as a new employee does. Pick work where an error is an inconvenience, not a disaster. Drafting replies, enriching lead records and categorising expenses are forgiving. Moving money or changing legal documents are not — at least not without approval steps.

This doesn't rule out sensitive work. It means you design a human approval step at the risky moment and let the agent do everything around it.

4. Can you measure "better"?

Before you build, agree on the number that will show the agent is working. Good examples:

• Median time to first response on support tickets
• Hours per week spent on manual data entry
• Percentage of inbound leads researched within an hour
• Error rate on processed invoices

Measure it for two weeks before the agent launches. Without a baseline, every result becomes a matter of opinion.

Putting it together

Score your candidate tasks against the four questions. The winner is usually unglamorous: triaging the support inbox, preparing lead briefs, matching invoices to purchase orders. That's exactly right. Boring, frequent and measurable is where agents earn trust.

Choose the first agent your team will be relieved to have, not the one that makes the best press release.

What happens after the first one

Once the first agent has run for a month or two, something shifts. People start suggesting the next task themselves, because they've seen what it's like to hand off the tedious parts of their day. That's when it makes sense to add a second agent, and eventually to let them hand work to each other.

If you'd like a second opinion on your shortlist, that's exactly what our free discovery call is for.