A 2–4 week AI agent pilot: how to judge payback before a full rollout
A pilot answers one question: does an agent in this process save hours and errors, or not. What to measure, how to turn the effect into money, and when to stop.
Why a pilot, not a programme
A large rollout starts with a hope. A pilot starts with one process, a 2–4 week window and a number you can check.
If that window shows no time saved and no fewer errors, the scope should not grow.
What to count before you start
Hours people spend on the process each week, and the cost of an hour.
Errors, returns and missed deadlines last month, and the cost of one error.
The systems the process cannot live without: 1C, CRM, email, the bank. The agent is judged inside them, not in a separate chat.
What payback looks like
Weekly effect ≈ hours saved × rate + cost of an error × errors removed.
The pilot is worth it when that sum over a quarter exceeds the pilot and the care that follows.
Example: invoice approval. Five people, four hours a week, the agent takes half — the effect shows up without a year-long business case.
What you keep after four weeks
A working agent on one process, or an honest stop.
An action log and an access boundary: what the agent may do alone, and what a person confirms.
A decision: production, a narrower task, or do not build it.