Choose your first automation candidate

You will leave with a pilot decision record

This is what you produce. The result at the top of the page is what it is for. These fields carry into the next guides:

• Candidate task candidate
• Reviewer reviewer
• Baseline baseline
• Pass and fail criteria passFail
• Stop and fallback fallback

Before you start

Bring:

• A mapped or clearly described process
• How often each task happens
• Your own observations of time and errors
• How sensitive the data involved is
• Systems you already use

Tools: A chat AI tool that accepts pasted text; Optional later: an automation tool your organisation has approved. No named products are recommended here; tools you already use are often enough. If you need to look further, see Useful external links.

Steps

1. List three repeated tasks

   AI can helpSuggests tasks from your process map that are frequent and well defined.

   You decidePick three real candidates.

   Output: Candidate list

2. Assess volume, stability and exceptions

   AI can helpBuilds a comparison table and writes "unknown" where you have no evidence.

   You decideFill in what you have observed, not what you expect.

   Output: Comparison table

3. Choose a low-consequence candidate

   AI can helpPoints out candidates that touch customers, money or personal data.

   You decideChoose the one where a wrong output is easy to catch and cheap to fix.

   Output: Chosen candidate

4. Estimate the full cost of review and setup

   AI can helpLists cost and effort questions. It does not produce savings figures.

   You decideEstimate setup, review time and subscription costs from your own situation.

   Output: Cost questions answered

5. Define acceptable results

   AI can helpDrafts pass and fail criteria from your description.

   You decideSet criteria the reviewer can actually apply.

   Output: Pass and fail criteria

6. Name the owner and fallback

   No AI neededThis step is for people.

   You decideName who owns the pilot, who reviews outputs and how you go back to manual.

   Output: Owner and fallback

7. Run a manual or sandboxed trial first

   AI can helpProduces drafts from copied, approved sample data.

   You decideCompare outputs with your baseline before connecting any system.

   Output: Trial results

Prompt to copy

Paste it into the AI tool you use, replace the parts in square brackets, and read the result before you use any of it.

Compare [three tasks] as first automation candidates using frequency, business value, process stability, data sensitivity, exception rate, review effort and integration complexity. Use "unknown" when evidence is absent. Recommend a bounded pilot or recommend postponing automation. Include a non-AI option, representative test cases, human approval points, failure conditions, fallback and a baseline measurement plan. Do not claim savings without measurements.

Template: Pilot decision record

Check before you use it

If any of these fail, the output is still a draft.

• The task has a stable, clear finish condition.
• Representative real examples exist for testing.
• Reviewing outputs is affordable in time.
• Wrong outputs can be caught and corrected before they reach anyone outside the team.

• The task has a stable, clear finish condition.
• Representative real examples exist for testing.
• Reviewing outputs is affordable in time.
• Wrong outputs can be caught and corrected before they reach anyone outside the team.

Common mistakes

• Starting with the most visible task instead of the safest one.
• Counting time saved without counting review time.
• Connecting live systems before a manual trial.

Measure it: comparable-case results

For comparable cases: total time including review, share of outputs accepted, and errors found. Time recovered is not automatically money saved.

Baseline first. Measure the task manually for a few weeks before the trial, including any rework.

Use this next

• Plan a small-team AI pilotNot in the library yet. Until then, keep your checked output from this guide ready for it.Carry forward: candidate reviewer passFail. Only carry forward a candidate the owner has approved.
• Test and evaluate a pilotNot in the library yet. Until then, keep your checked output from this guide ready for it.Carry forward: candidate baseline passFail fallback. Test cases must come from real, representative examples.
• Monitor and improve an automationNot in the library yet. Until then, keep your checked output from this guide ready for it.

Plan your next step

A short worksheet: result, why it matters to you, baseline, finish line, first action, owner, review date, AI role and human check. Suggestions from this guide are included to get you started.