Short answer: Choose one repetitive, reversible job with a named owner and a number you already watch. Leave judgement, money movement and customer promises for a later project. The first project is a test of your company, not of the model.
I'm Vinay Pasricha. I run GoodSpace AI in Noida, and I wrote AI for Business Leaders. The wrong first project is usually the most impressive one in the demo. The right one is slightly dull, happens every day, and can be undone.
A filter, in order
Start with the 90-day shape: diagnose, design, build, then decide. The diagnose step is the filter. You are looking for a job, not a vision.
- Repetitive: it happened often enough last month that you can remember ten cases without a workshop.
- Reversible: a wrong output can be caught before it leaves, or recalled without harming a customer.
- Owned: one person already does it, and will check the agent. A job nobody owns is a bad first project.
- Counted: there is a number. If there is no number, you will argue about feelings at day 90.
Anything that fails one of the four waits. That includes the project your board asked for if it fails them. You can explain a dull first win. You cannot explain a bold first failure that touched a customer.
Write the rejected ideas down. They are not discarded. They are queued behind the first project, and the queue stops the company from starting three at once. Three first projects is how none of them get an owner.
Sort the tasks, then pick one
List the tasks inside the job. Mark each automate, assist or escalate. The first agent project should be an automate task or a tightly checked assist task. An escalate task is a person, by definition. If your favourite idea sorts entirely to escalate, it is not an agent project. It is a management project, and that is a fine thing to do instead.
Prefer a task that reads from memory you can tidy in a week. A price list. A set of approved answers. A folder of past cases. If the memory would take a quarter to assemble, the project is really a memory project. Do that first, and do not call it an agent.
Prefer a task whose user is inside the company. An internal user will tell you the output is wrong. A customer will simply leave. Internal first, customer-facing later, after the Risk Ladder says the worst case is one you can carry.
How the first project ends
It ends on a date, with a decision. Scale means the same owner runs it on the next batch, with the check still in place. Fix means the memory or the boundary was wrong and you will change that before you continue. Stop means you learned the job was a bad fit and you will not quietly keep paying.
Do not add a second agent until the first decision is written down. A second project is how the first one escapes judgement.
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
Write three candidate jobs on a page. Strike any that a customer would see before a person checks it. Strike any with no owner. Strike any with no number. Start with whatever remains. If nothing remains, your first project is to create the count, not to buy the agent.
Where this comes from
This answer draws on The 90-Day AI Leadership Playbook and Automate, Assist, Escalate 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.
