Short answer: Pick one number you already track, write down what it was before the agent, and set a date to scale, fix or stop. Do not publish a return you have not measured. A percentage without a baseline is a slogan.
I'm Vinay Pasricha. I run GoodSpace AI in Noida, and I wrote AI for Business Leaders. I am not going to give you a return figure for agents in India. I have not measured yours, and a borrowed percentage would be a fiction. The method below is the one I use instead.
Three levels, one number
The book separates AI gains into efficiency, effectiveness and advantage. Efficiency is a task done with less effort. Effectiveness is the same task done better: fewer errors, a faster right answer. Advantage is a gain a rival cannot copy next quarter, usually because it sits in your own memory and your own way of working.
Most agent projects should be judged at the first or second level. Advantage is a later claim, and it is rarely true of a tool you rent. If a vendor leads with advantage, ask which number in your business moved, and from what baseline. If they cannot say, they are selling a story.
One number is a discipline. "Saved time" is not a number until you say whose time, on which task, against which week. "Better service" is not a number until you say which replies, judged how. Choose a count you already produce: invoices processed, quotes sent, tickets reopened, days a role stayed open. New dashboards are how measurement gets avoided.
A charter, not a feeling
- Baseline: the number for the four weeks before the agent touched the work. Write it down before you start, not after.
- The agent’s number: the same count, over the same kind of week, while the agent is in use.
- The cost you can see: the invoice, plus the hours a named person spent checking. Hours count. A free tool that eats a manager’s afternoon is not free.
- The decision date: scale, fix or stop. Put it on a calendar. A pilot without an end becomes furniture.
The appendices include a pilot charter and an ROI calculator as blank tools. They do not contain a result for your company. Fill them with your baseline. Leave a cell blank rather than type a hope.
Watch the failure count as carefully as the success count. An agent that speeds the work and doubles the rework has not earned a return. Put rework on the same page as the headline number, or the headline will lie to you.
What not to count
Do not count demo tasks. Do not count a week in which a person quietly did the job beside the agent and you recorded the agent’s name. Do not average this project with another one to make a company-wide return. The point of one number is that you can stand next to it and explain it.
If the number did not move, that is a result. Stop, or fix the memory and the task, and run the same number again. Do not change the number because the first one was inconvenient. Changing the number is how pilots survive without working.
Use the free Risk Ladder before the agent is allowed to act, and Automate, Assist, Escalate to sort the tasks around it. The eight steps are in AI strategy for business. Collapse or Collapse is the longer argument: one person owns one result and runs agents to deliver it.
One thing to do this week
Choose the number. Write last month’s figure on paper, with the source: which report, which person. If you cannot find last month’s figure, you cannot measure a return yet. Build the count first. The agent can wait a week.
Where this comes from
This answer draws on Three Levels of AI Strategy and Appendices: readiness audit and pilot charter in AI for Business Leaders.
Your company will have its own details. To ask a follow-up in your own words, try Ask Vinay's books.
Vinay Pasricha is the founder of GoodSpace AI and the author of six books, including AI for Business Leaders.
Go deeper in AI for Business Leaders
This answer is the short version. The full argument, and the method behind it, is in the book.
