Sample course outline — adapted after a planning conversation.

AI as an Employee

Design and test a supervised, repeatable AI responsibility using fictional records. Define when it can draft, when it must stop and who reviews its work. “Employee” describes delegation; a person remains responsible.

SAMPLE DURATION240 minutes

two 120-minute working labs, each including a 10-minute break. Live deployment and production integration are outside this sample.

TOOLS FOR THIS EXAMPLE

a laptop and a team-approved text AI tool. Tool and access requirements agreed during planning; practice uses pasted fictional text. A shared text document or worksheet holds the manual run log; no scheduler, connector or autonomous execution is assumed.

We agree on timing, tools, and deliverables with your team before booking.

Plan this workshop →

Your starting point.

Teams that already delegate and verify bounded AI tasks and want to explore an ongoing workflow.

Experience writing task instructions and checking AI outputs.
Understanding of the sample workflow’s business rules and its human approval process.
No coding required for the manual prototype shown here.

What you can leave with.

A role brief and manual supervised prototype.
An evaluation case table and run log.
A draft review, recovery and production-readiness question list.

Skills you will practice

Define a narrow AI role with permitted actions, stop conditions and a human owner.
Run a manual prototype against normal, duplicate and exception cases.
Record evidence for a go, revise or stop decision without treating a demo as deployment approval.

ILLUSTRATIVE EXERCISE · NOT A CUSTOMER RESULT

Route a disputed overdue invoice

Example output

Record: INV-201.
Action: Escalate; no reminder drafted.
Evidence: The supplied record is marked disputed=true.
Human next step: Accounts owner reviews the dispute and confirms whether follow-up is appropriate.
Nothing sent.

Review: The dispute rule overrides overdue status. The output escalates without writing a reminder, inventing a resolution or contacting anyone.

See the input and try the exercise

Fictional practice material

Fictional case, as of May 15, 2026: INV-201, Elm Cafe, $420, due May 1, unpaid, disputed=true. Run log: no prior draft. Rule: disputed invoices must be escalated to the accounts owner; do not draft or send a reminder.

Example prompt

Apply this draft-only follow-up rule: disputed invoices must be escalated to the accounts owner, with no reminder drafted or sent. Case as of May 15, 2026: INV-201 Elm Cafe $420 due May 1 unpaid disputed=true; no prior draft. Return record ID, action, evidence and human next step. Do not infer a resolution or contact the customer.

SAMPLE AGENDA / 240 MINUTES

What you will work on.

Lab 1: Define the responsibility

40 min

Write rules for drafting overdue-invoice reminders and escalating exceptions.

Take home: A role brief and acceptance rules.

Practice and review details

Practice: Use fictional invoice-follow-up cases. Define approved input fields, human owner, drafting rules and stop conditions for missing payment data, disputes or duplicate requests.

Example: Contrast a draft-only role with authority to contact a customer.

Check: The role cannot send messages; missing data, disputes and duplicate drafts have explicit handling.

Before this exercise: Experience with checked AI task workflows.

Lab 1: Build a manual prototype

70 min

Simulate one repeatable run with a log and bounded review.

Take home: Two logged sample runs and a draft-only handoff.

Practice and review details

Practice: Manually run one eligible and one disputed invoice through the instructions. Record invoice ID, input version, result and review state. Limit AI drafting to one attempt and one revision per case.

Example: Walk through input → eligibility check → draft or escalation → human review → log.

Check: Disputed case produces an escalation; eligible case produces an unsent draft linked to its source.

Before this exercise: Role brief from the first section.

Break10 min

Lab 2: Test failures and repeats

60 min

Detect missing information and avoid preparing a duplicate draft.

Take home: A four-case evaluation table with observed results.

Practice and review details

Practice: Run four cases: eligible overdue invoice, disputed invoice, missing payment status and invoice already logged with a pending draft. Compare each result with written expectations.

Example: Show a repeated record with an existing draft log and a record missing payment status.

Check: Eligible case drafts; disputed and incomplete cases escalate; pending duplicate is held. Any mismatch is recorded.

Before this exercise: Saved role brief, prototype instructions and run log from Lab 1.

Break10 min

Lab 2: Review, recover and decide

50 min

Revise a failed rule and document what must be checked before any live use.

Take home: Revised prototype, rerun evidence and a go, revise or stop recommendation for further supervised testing.

Practice and review details

Practice: Review the four results. Tighten one ambiguous rule and rerun affected cases. If all rules pass, test and correct this seeded flawed rule: “Draft a reminder for every overdue invoice, even if disputed.” Rerun the disputed case against the corrected escalation rule. Write stop, recovery and human approval steps, plus a proposed cadence for later discussion.

Example: Have a separate reviewer inspect source evidence and logs; explain that reviewer agreement does not prove correctness.

Check: Recommendation cites test results; failure routes to a person, run limits are explicit and no live schedule or deployment is enabled.

Before this exercise: Four completed evaluation cases.

What to bring and how we can adapt the session

What your team provides.

Choose one recurring responsibility and name its human owner.
Describe stop conditions, approval rules and the cost of a wrong action.
Confirm allowed tools and data; bring anonymized exceptions or use supplied fictional cases.
Identify who could evaluate permissions, logging and recovery before any later production work.

Use anonymized examples or our fictional material. Keep confidential records and credentials out of your inquiry.

Choose a useful focus.

Use a status-change monitor with a fixed supplied source instead of invoice follow-up.
Add a separate draft reviewer while retaining one human decision owner and a fixed revision limit.
Use an approved automation tool for a sandbox prototype only after features, permissions and time are confirmed.