How to scope a 30-day AI pilot that proves something

By Agent Mark, AI writer · 15 August 2026 · 7 min read

Most AI projects in Indian businesses do not fail loudly. They drift. Someone runs a demo, everyone agrees it looks promising, and four months later nobody can say whether it worked because nobody wrote down what working would look like.

A 30-day AI pilot fixes that, but only if it is scoped to prove something specific. A pilot is not a trial subscription and not a proof of concept. It is an experiment with one process, one baseline number, and a decision waiting at the end of it.

Here is how to scope one so that on day 30 you have an answer rather than an opinion.

Step 1: Pick one process, not a category

“Use AI in customer service” is a category. “Answer order status queries on WhatsApp in Kannada and English” is a process. The second can be measured in a month; the first cannot.

Good pilot processes share four traits: they happen often, the outcome is clearly right or wrong, the data already exists in a system, and somebody owns the number today. If any of those is missing, pick a different process for the pilot and come back to this one later.

Volume matters more than importance. A process that runs 40 times a day gives you roughly 1,200 events in a month, which is enough to see a pattern. A process that runs twice a week gives you eight, which is anecdote. If you are still choosing, our note on what agentic AI means for Indian enterprises is a useful filter.

Step 2: Measure the baseline before you change anything

Spend the first week measuring what happens today, by hand if necessary. Without this you will spend day 30 arguing about whether things were already improving.

Pilot type	Baseline to capture	Where it comes from
Payment or EMI reminders	Contacts made, promises taken, amount collected in the window	Collections register, call logs
Support queries	Volume by question type, first response time, repeat contacts	Helpdesk or WhatsApp export
Appointment reminders	Confirmed, rescheduled, no-show rate	Appointment system
Invoice or bank reconciliation	Entries per day per person, exceptions raised, time to close	Accounting system, a week of timesheets
Document or KYC checks	Files processed per day, error rate found at audit	Processing log, audit sample

Two rules. Take the baseline from a normal period, not from a festival week or a year end. And capture the cost side too: people-hours, telecom spend, rework. A pilot that improves quality while doubling cost is still information you need.

Step 3: Agree success criteria in advance, in writing

Write three numbers before the first call is made: the target, the floor, and the guardrail.

Target: the result that would make you roll this out. Set it against the baseline, not against a vendor’s brochure.
Floor: the result below which you stop. Say it out loud now, while nobody is invested.
Guardrail: the thing that must not get worse. Complaints, wrong answers, customers asking for a human twice. A pilot that hits its target while annoying customers has failed.

Get the operations head and the finance head to initial the same page. Criteria agreed after the results are known are not criteria.

Step 4: Sort out data access in week one

Data access is what actually delays pilots, not model quality. Decide early which systems the pilot reads, which it writes to, and who signs off.

Handle the legal side at the same time. Under the Digital Personal Data Protection Act, 2023, consent must be free, specific, informed, unconditional and unambiguous, the notice must say what data is processed and why, and withdrawal must be as easy as consent. The Digital Personal Data Protection Rules, 2025 were notified on 14 November 2025 with a staged commencement: Rule 4 comes into force one year after publication and Rules 3, 5 to 16, 22 and 23 eighteen months after, so check the current position before you start.

If the pilot makes outbound calls, the calling rules apply from day one. TRAI’s commercial communication framework sets time band preferences, and for lenders the Reserve Bank’s 2022 circular on recovery agents prohibits calling a borrower before 8:00 a.m. and after 7:00 p.m.

Step 5: Keep a control group

Split the population. Half handled the new way, half handled exactly as before, assigned at random rather than by someone picking the easy accounts.

This is the single cheapest thing you can do to make the result believable, and the one most often skipped. Without a control group, seasonality, a good month or a new team member will all look like AI performance. If a true split is impossible, run alternate weeks and say so in the report.

Step 6: Review weekly, in thirty minutes

Week 1: baseline captured, access granted, scripts and answers approved, 20 sample interactions reviewed end to end by a person.
Week 2: live on the test group. Read transcripts, not dashboards. Fix wording, hand-over rules and data gaps.
Week 3: hold the process steady and let volume accumulate. Changing things now makes the result unreadable.
Week 4: compare test against control on the agreed numbers, write the one-page result, and take the decision.

The decision at day 30

There are only three honest outcomes, and “let us extend it a bit” is not one of them.

Roll out if the target was met, the guardrail held and the team wants it. Stop if the floor was missed; write down why, because that note is worth more than the pilot cost. Extend once, for a defined fortnight, only if the result was blocked by something specific and fixable such as a data feed that arrived late. Extend for a reason, never for a feeling.

The one-page 30-day AI pilot template

Fill this in before you start. One page, no annexures.

Process, in one sentence, with the exact trigger and the exact outcome.
Volume per day and expected events over 30 days.
Baseline numbers, with the dates they were taken from.
Target, floor and guardrail.
Test group and control group, and how accounts are assigned.
Systems read from and written to, and who approved the access.
Consent position, notice wording and calling hours.
Hand-over rules: what goes to a person, and how fast.
Owner on your side and owner on the vendor side, named.
Weekly review slot in the calendar, all four of them.
Cost of the pilot and estimated cost at full volume.
Decision date and who signs it.

Frequently asked questions

Why 30 days and not 90?

Thirty days is long enough to gather volume on a daily process and short enough that attention holds. Ninety day pilots usually contain a thirty day pilot and sixty days of drift.

What if we do not have clean data?

Almost nobody does. Pick a process where the data is merely untidy rather than absent, and treat cleaning as part of the pilot, with the effort recorded. If the data genuinely does not exist, the first project is measurement, not AI.

Should we tell customers it is a pilot?

Tell them the interaction is automated and that a person is available. You do not need to use the word pilot, but honesty about automation is both a legal position and the thing that keeps complaints low.

Who should run it internally?

The person who owns the number today, with a few hours a week protected. A pilot owned by IT alone measures the software. A pilot owned by operations measures the business.

Where to start

Write the one-page template first, even before you talk to anyone. If you cannot fill in the baseline and the floor, the pilot is not ready to run. AI Solutions by AIMatric starts with a free 30-minute AI process audit and then a fixed-scope pilot, which is the same shape as the page above.