Module 09 / 10 · 40 planned minutes
Build a small reliable workflow
Design and test a bounded workflow with evidence checks, stop conditions and a named approval owner.
Combine AI07's reusable method and AI08's access discipline into a controlled sequence. Evaluate the completed, checked result rather than the speed of a first draft.
First pass: draw Cedar's path from input to draft to review. Move on when your workflow stops on a missing or conflicting source, resists C06's attempted action and measures the whole effort honestly.
Learn 14 planned minutes
A workflow is more than a long prompt
A workflow connects steps: receive instructions, gather permitted inputs, extract, draft, check, review and hand over. An ordinary automation follows a configured sequence. An agent may choose intermediate steps or tools. The more freedom and access it has, the more important it is to state boundaries and enforce them outside the prompt where possible. ANTHROPIC
Begin with a narrow result and a clear input set. “Manage this matter” is not a beginner workflow. “Prepare an internal status draft from these three identified documents and stop for review” is inspectable. Some steps are better done with a calculator, a template or a person; using AI everywhere is not the objective.
Use clear states. A draft can be prepared, source-checked, awaiting review or approved for a specified next step. A label produced by an AI tool does not move a document into an authorised state. The human reviewer must see the correct version and relevant evidence.
Design the unhappy path
What happens if the signed file is missing? If two versions disagree? If a connector fails? If a source asks for an unauthorised action? A useful process stops, records the limitation and identifies the next human decision. It should not silently switch to web guesses, widen access or send an incomplete answer to meet a deadline.
For live operation, configuration should limit accessible sources and actions, record meaningful events and preserve a manual fallback. Review gates should bind the exact output and action being approved. A broad “go ahead” at the beginning is not evidence that the reviewer inspected a later changed draft. No external action is permitted in this course. NIST
Test what matters
Define the expected behaviour before running a test. Include an ordinary example, a missing/conflicting-input example and an attempted unauthorised action. Check material facts, omissions, source traceability and stopping behaviour. A system that catches every bad statement but omits a critical obligation still needs correction.
Use a versioned record of the prompt or skill, inputs, result and changes. Repeat after a material change. A few familiar fictional tests support learning, not a claim of reliability across real matters. For consequential use, the organisation needs appropriate evaluation, representative cases and continuing oversight. The customer-support study in the reading register reports differing effects across workers; it is not a legal-training trial or proof of benefit for this workflow. NBER
Measure the checked result
Count setup, generation, source verification, correction and handover. Compare with a similar manual task of acceptable quality, if measured. Record critical errors and unresolved gaps alongside time. Do not advertise a saving based only on fast text generation.
If AI takes longer but teaches you a useful process, say that. If a review requirement makes the approach impractical, redesign or stop. Adoption targets should not encourage staff to hide errors or choose unsafe tasks.
Worked example
In a fictional timing exercise, the manual task takes 18 minutes. AI setup takes 4 minutes, generation 1, checking 8, correction 6 and handover 0 — not applicable. Total assisted time is 19 minutes: one minute longer. If handover were required but unmeasured, 19 would be a subtotal, not a measured end-to-end total. The generation step was fast; this completed example was not faster.
A better next experiment might use a clearer source list or a tested template. It should not remove the necessary checks to make the numbers look better.
Practise 20 planned minutes
Guided practice
Run a paper walkthrough of your Cedar workflow on C01–C05. Then add unsigned C07, and separately test C06's embedded instruction. Use S09 if needed. Record each expected stop or limitation before reviewing the supplied candidate behaviour.
Feedback — compare after your attempt
C07 cannot silently replace the signed price: the draft's status and the authority question remain unresolved. C06 cannot approve the supplier or authorise sending. Correct dependent tables, calculations and drafts if an erroneous value or status has propagated. A note buried at the end is not enough if the headline still presents the wrong result.
Practice assignment
Complete A09 with a five-step workflow, three short test records and a time/quality log. You may use the supplied fictional timing example, clearly labelled; do not present it as your own measured saving. Include the reviewer, stop conditions and manual fallback. No automation software is required.
Participation
Ask a peer to remove one input and observe where the paper process stops. Solo alternative: walk through the missing-source case and identify every downstream output that must remain provisional.
Check 6 planned minutes
Two formative checks. Choose an answer, then check it or open the comparison. Retry freely; this is not an award or an observed-skill record.
Source study and completion check
Optional pointer: NBER's abstract, with its customer-support scope. Explain why its findings cannot guarantee a percentage saving for this workflow. Complete AI09's two checks.
Before moving on, compare your artifact with the module feedback, correct material errors and keep its actual route label. This page does not record or certify completion.