AI project workbook · Eric Muriel

AI PROJECT WORKBOOK · 01 / 06

The problem worth solving

Choose one task and record how it works today. Frequency and timing should come from observation or be marked as estimates.

Fictional example: prepare a draft maintenance order from a request. Do not send quotations or assign technicians.

Person doing the task

Input received and result needed

Frequency, complete handling time and measurement source

What is outside the first pilot

ERIC MURIELericmuriel.com

AI PROJECT WORKBOOK · 02 / 06

Input, output and unknown information

Define the record a reviewer receives. Missing data must be allowed to remain missing; specify what happens when sources disagree.

Fictional example: machine_id=M-17, location=unknown, status=needs_review. Do not guess a location.

Required fields and data types

Fields that may remain empty

Example input and expected output

Rule for contradictory information

ERIC MURIELericmuriel.com

AI PROJECT WORKBOOK · 03 / 06

Sources, permissions and review

Write specific operations. Verify boundaries in access controls, not only model instructions. Do not put passwords or keys in this workbook.

Fictional example: read a test catalogue and create drafts in one list. Do not send email, delete documents or expand access.

Source, current version and owner

Account, permitted operation and scope

Action requiring approval and decision owner

How to revoke access or return to manual work

ERIC MURIELericmuriel.com

AI PROJECT WORKBOOK · 04 / 06

Test cases and acceptance criteria

Describe the expected outcome before running a test. Include complete, incomplete, ambiguous and out-of-scope cases. Keep evidence of observations.

Fictional example: receiving S-17 twice should produce one operation. If the first attempt’s result is unknown, review before retrying.

Normal case: input → expected result → observed result

Missing or conflicting data: expected → observed

Unauthorised or repeated request: expected → observed

Acceptance condition and failure requiring a stop

ERIC MURIELericmuriel.com

AI PROJECT WORKBOOK · 05 / 06

Complete cost and sensitivity

Count generation, review, exceptions and maintenance. Released time is capacity; it is not automatically income or cash savings.

Hypothetical example: 600 × 6 ÷ 60 = 60 manual hours. With 20 review hours and 8 maintenance hours, 32 hours are released. These are not market prices.

Monthly volume × manual minutes ÷ 60 = current hours

Review hours + exceptions + maintenance

Tools, usage and implementation cost; source and date

Favourable and unfavourable scenarios; changed assumptions

ERIC MURIELericmuriel.com

AI PROJECT WORKBOOK · 06 / 06

Pilot decision and next review

Collect evidence and choose whether to expand, revise or stop. Identify the process owner and changes that require retesting.

Fictional example: expand to a second location only after repeating permission tests. Keep the manual path available.

Result against the initial objective, with evidence

Known errors, outstanding work and limitations

Decision, owner and reason

Next review and conditions for retesting