Operating Notes: decision sheets for AI work

1. Follow the hour. Where was released time reassigned? Ask the supervisor to show the shift or queue.
2. Find the money event. Did overtime, a vendor bill or fulfilled contribution change? Include review and run costs.
3. Give the board one honest sentence. Name the observed change and the financial decision still outstanding.

Which source would make your controller change the board line?

1. Watch one real order. Log the moment a quote is ready, approved and accepted. Name the longest wait.
2. Try a rule before a model. Route ordinary discounts under a clear rule; send exceptions to a named approver.
3. Price the alternative. Compare time to accepted quote, margin errors and the continuing cost to operate.

What would you change first: the draft, the rule or the handoff?

1. Take the test, leave the records. Share a generic duplicate example and expected decisions. Keep each company’s source records local.
2. Change one detail. Amend the purchase order after approval. Which local role may stop the posting?
3. Buy a replay before a rollout. Ask the second controller to accept a local trial before you fund expansion.

What can the second company reuse without assuming the first company’s rights?

1. Start with a fresh case. Let the backup use the permissions available on shift. The builder observes without coaching.
2. Interrupt the task. Replace an input after a draft is ready. Have the operator stop the stale action and route it.
3. Prove recovery. Restore the known version, run a safe task and keep the receipt. Record any private help needed.

Who can restore the workflow when your usual expert is away?

1. Keep the buyer’s words. Start with a real, de-identified comparison question. Freeze wording, locale, system and date.
2. Open the footnote. Check identity, destination and the paragraph that supports the answer. Log wrong-company links separately.
3. Follow the useful visit. Keep citations, qualified inquiries and held meetings in separate columns. Allow unknown attribution.

Which buyer question would you actually want your company cited for?

1. Separate what transfers. List software, rights, source access and operating responsibility as separate diligence questions.
2. Price the unknown. Record which assumption would change the investment case and what cannot be tested pre-close.
3. Hand off a decision. Give the post-close owner a lawful test, a review date and a reason to stop.

Which assumption in your investment memo needs its own post-close test?

1. Look past the reply. Check which order and payment event each request refers to before recommending a refund.
2. Draw the action boundary. Separate reading an order, recommending a refund and issuing one. Give each action a named role.
3. Test a second attempt. Run a duplicate request in a safe environment. The second attempt should hold or return the original result.

What stops a perfectly plausible second refund?

1. Choose a held-out task. Use a permitted task with a quality gate and a time limit. Keep the task constant.
2. Swap the dependency. Change the adapter in a test environment. Record what besides the model required work.
3. Reprice the exit. Compare review time, cost per accepted task and recovery. Keep unresolved dependencies in the agreement.

What would your next provider inherit besides a repository?

1. Change the example. Test a fresh case, a missing source and a prohibited action. Keep the policy available.
2. Measure correction effort. Track accepted completion, escalation and reviewer time. A fast wrong answer moves work downstream.
3. Give the backup a turn. Repeat independently with the backup. Record which step needs coaching and rerun a fresh case.

Which task can a new operator complete without a trainer?

1. Start with the portable parts. Share questions, acceptance rubrics and generic failure cases before identifying any customer.
2. Try a denied request. Show that a valid local request succeeds and an otherwise identical cross-company request fails.
3. Plan for divestiture. Name who owns export, revocation and support when an entity leaves the portfolio.

Which parts of your playbook still work if the underlying data never leaves the company?

1. Replay one abandoned task. Watch the person who stopped using the tool. Identify the exact point they leave it.
2. Price the checking work. Count time to accepted work with review, rework and escalation included.
3. Make a reversible choice. Repair one dependency, change the process or stop. Name the evidence that would reverse the choice.

When did the last user return to the old workflow, and why?

1. Trace the operating claim. Attach the baseline definition, observed window and changing volumes to the metric.
2. Separate capacity from financial recognition. Mark what the controller accepted, what is modeled and what has not converted.
3. Rehearse buyer verification. Have a permitted independent reviewer reproduce the comparison without the original presenter.

What would you hand to a buyer who wants to test the claim today?