Choosing your first AI agent: a one page test

Why the first choice matters

A pharmaceutical distributor in Rajkot receives around forty supplier invoices a day, as PDFs on email and as photos on WhatsApp. Two people in accounts open each one, find the matching purchase order in Tally, check quantities and rates, and post the entry. When a rate does not match, they call the supplier. The work is steady, dull and important, because a wrong rate flows straight into GST input credit and the margin on every strip of tablets.

The owner's nephew has read about AI agents and wants to start with something bigger: a sales assistant that predicts which chemists will reorder. That project might work one day. But it depends on clean sales history, it is hard to check, and nobody in the company will know for months whether it is helping. The invoice work, on the other hand, is visible every morning. If an agent matches thirty of the forty invoices correctly and puts the other ten in front of a person with the reason, everyone can see the difference in a week.

Your first agent sets the tone for every project after it. Pick one that people can check, and the second project gets a fair hearing. Pick one that nobody can measure, and the whole idea gets quietly shelved.

The one page test

Take one process you are considering and answer each question with the people who do the work today. A manager can describe the process; the people doing it know the exceptions.

1. How often does it happen?

Daily or weekly work is a good candidate. A task that happens twice a quarter will not give the agent enough practice, or give you enough evidence, to judge it. Forty invoices a day is good. Annual budget preparation is not.

2. Can you write down the rules?

Ask the person doing the work to explain how they decide. If the answer is a list of checks (vendor matches, PO number present, quantity within received quantity, rate within one rupee of the PO rate), an agent can follow it. If the answer is "I just know", spend a week writing the rules down first. That exercise alone often improves the process.

3. Where do the inputs come from?

List every source: email inboxes, WhatsApp, a shared drive, a scanner, a portal. An agent needs a reliable way to receive the inputs. If half the invoices arrive as photos taken at an angle in a godown, note it. The agent can still read many of them, but you will need a plan for the ones it cannot.

4. Where does the output go?

The result has to land somewhere useful: a Tally voucher in draft, a row in a Google Sheet, a ticket in Freshdesk, a reply on WhatsApp. If the output is "a summary someone reads", ask who reads it and what they do next. An output with no next step is a report. An agent leads to an action.

5. How do you know the answer is right?

This is the most important question. For invoice matching, a person can check a match in seconds by looking at the invoice and the PO side by side. For a sales forecast, nobody can tell for months. Choose work where a wrong answer is easy to spot.

6. What does a mistake cost?

A wrongly matched invoice that a person catches before posting costs a minute. A wrong message sent to a borrower by an NBFC collections agent can cost a complaint to the regulator. Both can be automated, but the second needs far more checks. For a first project, pick work where mistakes are cheap and caught early.

7. Who approves the result today?

Every process already has a person who signs off. Keep that person. The agent prepares the work, the same person approves it. This keeps responsibility where it was and makes adoption much easier.

8. Is there personal or sensitive data?

Supplier invoices carry business data. Patient discharge summaries and loan files carry personal data covered by the Digital Personal Data Protection Act, 2023. A process with personal data can still be a good candidate. The data changes where the agent should run and who can see its logs. Note it now.

9. Who will own the agent?

Someone inside the team must own the prompts, the rules and the exceptions queue. If nobody has the time, the agent will drift. A first project needs a named owner who spends a few hours a week on it, at least for the first two months.

10. What will you stop doing?

If the agent works, what changes? Perhaps the two accounts staff spend their mornings on supplier disputes, with the data entry handled by the agent. Write this down. It helps you judge success and it answers the question every team member will quietly ask.

Scoring the answers

You do not need a formula. Look for these patterns.

• Strong first candidates: frequent, written rules, clear inputs, a system to write the output into, easy to check, cheap mistakes, a willing owner. Invoice matching, IT helpdesk intake, leave policy questions, routing incoming support tickets, and pulling fields out of standard documents usually fall here.
• Good second projects: frequent and valuable, but mistakes cost more or inputs are messy. A hospital drafting discharge summaries for a doctor to correct, or a school group's admissions office answering parent queries in Gujarati and English.
• Not yet: rare, judgement heavy, or impossible to check soon. Forecasting, pricing strategy and anything where the rules live in one senior person's head.

Common traps

Starting with a chatbot for everything. A general assistant that answers any question sounds useful, but it is hard to test and hard to improve. A narrow agent that does one job well teaches you far more.

Choosing the process with the loudest complaint. The most painful process is often painful because the rules are unclear or the data is poor. Fix those first, then automate.

Skipping the baseline. Before the agent starts, note how long the work takes and how many errors reach the next stage. Without that, you cannot tell whether the agent helped.

Limits to plan for

A one page test cannot tell you how well a model will read your particular documents. Your supplier invoices may use layouts that confuse it, or your WhatsApp messages may mix three languages in one line. The only way to know is a short trial on real samples, with the people who do the work checking every result. Plan for two to four weeks of that before you decide to keep the agent.

The test also will not make a poor process good. If invoices routinely arrive without PO numbers, the agent will flag most of them, and that is the correct behaviour. Sometimes the first win from an AI project is discovering a fix that needs no AI at all.

Checklist before you start

• One process chosen, with the people who do it in the room
• Rules written down, including the exceptions
• Inputs listed by source and format
• The system where the output lands, and who approves it
• A baseline: time taken and errors today
• A named owner with time set aside
• A note on personal data and where the agent should run

When you have a process that passes, you can sketch the agent in the Agent studio and test it on your own samples before anything goes live.