Propose your first GenAI pilot at work on one page
In this article (8 sections)
Related programme: the Generative & Agentic AI course (3 months).
A first GenAI pilot proposal should fit on one page and answer six things: the task, who does it today, what the tool would do, how you will measure the result, what data it touches, and when you will stop. Approvers rarely reject a small, bounded pilot for being unambitious; they reject proposals that are vague about data, success and exit.
Why one page
A one-page proposal forces decisions you would otherwise hide in appendices. It also matches how most managers approve small experiments: they read it in the few minutes before a meeting and either sign or ask two questions. If your pilot needs ten pages to explain, it is not a first pilot.
The template
Copy this into a document and fill every line. If a line is hard to fill, that is the part of the idea that needs more work.
PILOT: <name, in plain words>
OWNER: <you> SPONSOR: <manager who approves> DATES: <start> to <end, max 6 weeks>
1. THE TASK TODAY
What: <one sentence>
Who does it, how often: <role, frequency>
Current effort: <how you measured it, e.g. timed 10 instances>
Current pain: <errors, delays, backlog — observed, not assumed>
2. WHAT THE TOOL WILL DO
Tool: <name, and whether it is on the approved list>
Step it performs: <draft / classify / summarise / extract>
Step a human still performs: <review, approve, send>
3. DATA
Inputs: <documents / fields>
Contains personal or client-confidential data? <yes/no; if yes, how handled>
Leaves our environment? <yes/no; which approved service>
4. MEASUREMENT
Primary measure: <e.g. median minutes per case, including review>
Quality check: <e.g. reviewer rejects or edits; errors found in a sample>
Comparison: <same task type, same people, before vs during>
5. STOP CONDITIONS
Stop immediately if: <data incident, a harmful output reaching a customer>
Stop at end if: <quality below current process, or no time saved after review>
6. WHAT HAPPENS NEXT
If it works: <who decides on wider use>
If it doesn't: <what we will have learned; how we switch off>A filled example
This example is invented for illustration. Replace everything with your own facts.
PILOT: Draft first replies to vendor invoice queries
OWNER: Accounts payable analyst SPONSOR: AP team lead DATES: 3 weeks
1. THE TASK TODAY
What: Vendors email asking about payment status; an analyst checks the
ERP and replies.
Who, how often: 2 analysts, roughly 40 emails a week (counted from the
shared inbox over the last 4 weeks).
Current effort: Timed 10 replies; median 9 minutes, mostly finding the
right invoice and writing a polite reply.
Current pain: Replies go out 1-2 days late during month-end.
2. WHAT THE TOOL WILL DO
Tool: Company-approved assistant inside our email suite.
Step it performs: Drafts a reply from the email plus the invoice status
the analyst pastes in.
Step a human performs: Analyst looks up the status in the ERP, reviews
the draft, edits and sends. Nothing is sent automatically.
3. DATA
Inputs: Vendor email text; invoice number, amount, status.
Personal data: Vendor contact names only; no bank details pasted.
Leaves environment: Only to the approved assistant.
4. MEASUREMENT
Primary: Median minutes per reply, including review, over 3 weeks.
Quality: Team lead reviews 10 sent replies per week for errors or tone.
Comparison: The 10 timed replies above.
5. STOP CONDITIONS
Immediately: Any draft containing an amount or status not supplied by
the analyst that reaches a vendor.
At end: If median time is not lower, or the lead finds more errors than
in the baseline sample.
6. NEXT
Works: AP lead decides whether to extend to the second inbox.
Doesn't: Switch off; share the log of where drafts failed.What approvers read first
Most sponsors look at section 3 and section 5 before anything else. Data tells them the risk; stop conditions tell them you have thought about failure. The measurement section is what they will ask about at the end, so make it something you can actually collect without a new system.
Notice what the example does not include: a projected annual saving, a headcount number, or a claim that quality will improve. With three weeks of data from two people, any such figure would be a guess. State what you will know by the end, not what you hope.
Getting the baseline right
The weakest line in most first proposals is "current effort". A guess such as "this takes ages" gives the sponsor nothing to compare against later, and a flattering pilot result will be doubted. Measure before you ask:
- Time ten real instances of the task with a stopwatch or calendar notes, including the lookup and checking steps, not just the typing.
- Note how many of those ten had an error, a delay or a rework.
- Record who did them. If the pilot is run by a faster colleague, the comparison is unfair in the tool's favour.
Ten timed instances is not a statistical study, and you should not present it as one. It is enough to show that you measured rather than assumed, and to spot whether the tool changed anything large.
Reasons first pilots get refused
| Refusal | What it usually means | Fix before resubmitting |
|---|---|---|
| "Not now" | Timing clashes with a busy period | Propose a start date after it |
| "Check with IT first" | Tool or data question unresolved | Get the approved-tool answer in writing |
| "Too big" | More than one task, or longer than six weeks | Cut to one task and three weeks |
| "What's the benefit?" | Baseline or measurement is vague | Add the timed baseline |
Sections to leave out of a first pilot
- Vendor comparison matrices. Use the tool you already have approval for.
- Architecture diagrams. If the pilot needs one, it is too big.
- Return-on-investment projections. Report measured results after the pilot instead.
- Future phases. One line in section 6 is enough.
After you get a yes
Send a short update halfway through with the numbers so far, even if they are unflattering. Mid-pilot honesty is what earns you the second pilot. At the end, report against the stop conditions you wrote on day one, not against new criteria you chose afterwards.
Before you submit, expect governance questions; the questions your company will ask before approving an AI project lists them with model answers. The early groundwork of building credibility is covered in the 90-day plan for becoming the AI person on your team. Scoping and evaluating assistants like this one is practised with synthetic data in the Generative & Agentic AI add-on.
Pankit Kumar has 10 years in Data Science & AI, building and shipping production systems in regulated pharma and clinical environments. He is a freelance trainer at Boston Institute of Analytics, AnalytixLabs and Scaler, and has taught this material to thousands of working professionals.
This article is part of our Generative & Agentic AI programme — 3 months. Add practical GenAI, retrieval and agent-building skills to your existing toolkit.
Explore Generative & Agentic AI