Your first AI agent: a 90-day plan

Days 1–15: pick the right task

The best first project is not the most exciting one. It is the one that is easy to measure and safe to get slightly wrong. Score your candidate tasks against four questions:

• Does it happen at least dozens of times a week?
• Does it follow rules you could write down for a new employee?
• Is the data it needs already in digital systems?
• If the agent makes a mistake, will a person catch it before it matters?

Lead replies, document checks, invoice matching and weekly reporting usually score well. Pricing decisions and complaints from your largest customer usually do not.

Days 15–30: measure today’s baseline

Before any build, record how the task performs now: hours spent each week, response time, error rate and backlog. Without a baseline you cannot prove the agent helped, and you will struggle to justify the next one.

Collect 50 to 100 real past examples too. They become the test set the agent must pass before it touches live work.

Days 30–60: build with guardrails

Write the agent’s job description as you would for a new hire: what it does, which systems it may use, what it must never do and when it must ask. Then build in that order:

• Connect read-only access first, then the minimum write access the job needs.
• Run the agent on your test examples and review every failure.
• Add approval steps for anything risky or irreversible.
• Log every action so any decision can be traced and explained.

Days 60–90: pilot, then widen

Run the agent alongside your team on live work. In the first weeks a person approves every action. As accuracy holds, approvals move to exceptions only. Compare the results with your baseline every week.

At day 90 you should have a clear answer: keep, adjust or stop. A good pilot also leaves you with a reusable foundation, including integrations, logging and evaluation, which makes the second agent much faster to build.

Common mistakes to avoid

• Starting with the hardest process because it promises the biggest win.
• Skipping the baseline, then arguing about whether the agent helped.
• Giving the agent broad system access “to save time later”.
• Treating the pilot as finished at launch, instead of measuring it weekly.