Choose Your First AI Workflow

One workflow. One owner. One useful next step.

Bring three recurring tasks your team already does. Rough notes are fine.

The first prompt compares them using what you know and asks no more than three short questions. You’ll leave with one workflow worth testing, or the missing fact to settle first.

Start here: copy Prompt 1 into ChatGPT or Claude. It will ask for your three tasks. Use a tool your workplace allows and leave out sensitive information.

Prompt 1: Choose your workflow

Act as a practical AI adoption advisor. Help me choose one recurring team workflow for a small AI pilot.

Start by asking me to list three recurring team tasks. Tell me plain notes are fine. If I need examples, offer weekly updates, meeting follow-ups, support handoffs, invoice checks, or another task I actually own.

After I answer, ask at most three short follow-up questions, one at a time. Ask only for facts that could change the choice. Compare the candidates using what I provide about:
- how often each task happens
- where people repeat work, wait, lose information, or make avoidable mistakes
- who owns the work
- what inputs it uses
- what data or actions must stay outside the AI tool
- who can verify the output

Do not invent scores, time savings, ROI, error rates, or data quality. Mark missing facts as unknown.

Then recommend one workflow provisionally and explain the choice from my facts. Name the owner, the useful first AI task, the human check, and the important unknowns. Say why the other candidates should wait. If none is ready, do not force a choice. Tell me the single missing fact or boundary to resolve next.

If you tried the weekly-update exercise already, it can be one of your three candidates.

Save all three prompts

The prompts on this page are free to copy. Right Click Prompt is an optional way to keep them together and find them again when the next task comes up.

Prompt 2: Build the pilot plan

Bring your chosen workflow and the notes from Prompt 1. This turns the choice into a small test with an owner, a human check, and a way to tell whether it helped.

Act as a pilot designer. Turn the chosen workflow below into the smallest useful AI test a team can review safely.

Chosen workflow and notes: {{paste the result from Prompt 1}}

Ask at most three essential clarifying questions, one at a time. Prioritize facts needed to run the test safely. Then mark anything still missing as unknown instead of inventing it.

Build a one-page pilot plan with:
- one owner
- the event that starts the workflow
- the input the AI receives
- the single task the AI performs
- the output format
- the person who checks the output and what they check
- data and action boundaries
- the current baseline, or unknown
- one observable success metric
- a small set of real or safely reconstructed test cases
- a stop condition

Do not claim savings or improvement before the test runs. Finish with a short checklist the owner can use for each test case.

Prompt 3: Review the pilot

Use this after a test. If you haven’t run one yet, use the worksheet below to decide what you’ll record.

Act as a skeptical pilot reviewer. Review the plan and test records below without filling any gaps.

Pilot plan: {{paste the plan from Prompt 2}}
Test records: {{paste what happened in each test}}
Baseline: {{paste the comparable before-state, or write unknown}}

Separate the review into:
1. Observed results supported by the records
2. Reasonable interpretations that still need testing
3. Unknowns we cannot answer yet

Compare the pilot with the baseline only when the measures and cases are comparable. Do not invent missing results, averages, ROI, or causes.

End with exactly one decision:
- KEEP
- CHANGE
- STOP
- NEED MORE EVIDENCE

Explain the decision from the records. If the decision is CHANGE, define the next smallest test. If the evidence is thin or not comparable, choose NEED MORE EVIDENCE and name exactly what must be recorded next.

A fictional example

A team leader lists three tasks: the Friday status update, support handoffs, and invoice exceptions. The update happens weekly, has one editor, uses known project notes, and gets reviewed before leadership sees it. The support handoffs contain customer data the team hasn’t cleared for AI use. Nobody can describe the invoice-exception baseline.

A reasonable first choice is the status update. The editor owns it. AI does one thing: flag questions the draft doesn’t answer. The editor checks every question before changing the update. How often blockers are missed today is still unknown. That needs recording before anyone claims an improvement.

Copyable pilot worksheet

WORKFLOW:
OWNER:
STARTING EVENT:

CURRENT PROCESS:
CURRENT BASELINE OR UNKNOWN:

AI RECEIVES:
AI DOES ONE THING:
AI RETURNS:

HUMAN CHECKS:
DATA THAT MUST NOT ENTER THE TOOL:
ACTIONS AI MUST NOT TAKE:

TEST CASES:
SUCCESS METRIC:
STOP CONDITION:

FOR EACH TEST, RECORD:
What happened
What helped
What needed correction
Time or quality measure, if observed

DECISION: KEEP / CHANGE / STOP / NEED MORE EVIDENCE
EVIDENCE FOR THE DECISION:
IMPORTANT UNKNOWNS:

Where to go next

Start with Prompt 1. Once you have a candidate, share the choice and its unknowns with the person who owns the work. That’s the beginning of an AI habit your team can actually use.

Want the full five-prompt process for comparing workflows and building an implementation plan? Paid members can open Stop guessing which process to automate first.

For more ideas, browse the Start Here guide.

Adapt and Create,
Kamil