Design a workflow another operator can use
Review and limits. The course content has completed cross-model editorial review. Domain weights, timings and the pass standard remain provisional: no completed job-task analysis or empirical standard-setting study is claimed.
This blueprint extends the shared CAI-3 framework. CAI-2 is recommended with no admission prerequisite. Agent Fundamentals supplies the common bridge, and the challenge-only path remains available. Course completion does not itself earn a credential.
The whole-task spine develops an account-evidence register, bounded outreach/offer chain, proposed CRM state change and pipeline-quality operating dossier. All work stays text or Markdown in the AI workspace. Tools are written contracts and manual dry-runs; actual assistant responses are separately observed. No external account, upload or live integration is required.
Framework alignment
The five domains retain equal 20% weight: D1 workflow architecture; D2 bounded tools and integration; D3 data boundaries and grounded verification; D4 human authority and recovery; D5 testing and operating oversight. The three exam families remain equally weighted with their shared criterion ids. These are course-practice artifacts, not new exam definitions, timers, cut scores or empirical validity claims.
| Module | Main domains | Dossier contribution | | --------------------------------------- | -------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | 1 Agent Fundamentals | D1–D5 | Shared foundation: bounds, authority, source evidence, observed trials and recovery | | 2 Design account evidence chains | D1, D2, D3, D4 | workflow, field map, source register and owner handoff. Keep organization facts separate from inferred needs and purchase intent | | 3 Build offer and outreach draft chains | D1–D5 | exact prompts/schemas, normal/missing/unavailable actual assistant trials, a specific revision/retest and an exact-payload approval table | | 4 Integrate CRM state and recovery | D1, D2, D4, D5 | written tool contracts, dependency/state table and recovery playbook. Do not infer success from matching current stage | | 5 Oversee pipeline evidence and quality | D2, D3, D4, D5 | a reproducible counts memo, four-defect audit and signals/owners/restart plan | | 6 Assemble and test the sales workflow | D1–D5 | Integrate separate design, build/test and oversight audit sections with an operating handover across all five domains. S601 source-grounded account/outreach draft, S602 commercial term absent, S603 propose_crm_update unavailable. No live send, commercial promise or CRM execution |
Every role module has four lessons: a scored branching scenario, sandbox chain trials, an event-grounded critique and a dossier artifact checkpoint. Checkpoints contain three retrieval questions, with two revisiting earlier concepts, and a module-final capstone-practice rubric. Practice targets 0.8 mastery with corrective feedback and retakes. Each criterion has four independent binary conditions; levels 0–4 equal the number met. Practice scores do not authorize credentials.
Fictional operating context
Tarnfield Software sells Tarnfield Rota. PROD-v2 supports rota drafts, manager shift approval and CSV schedule export; payroll calculation and automatic payroll writes are unavailable. ACCOUNT-v1 states that fictional Hestrel Hotels has 12 sites and managers reconcile changes manually. A product-account link remains a hypothesis unless supported. Nia approves exact outreach/CRM payloads; Omar decides commercial terms. Tessa owns integration repair and Samir owns quality review. OPP-v1 contains six distinct opportunities: three qualified, two discovery, one proposal, with no revenue or probability claim. Shared traces contain only account/opportunity ids, source ids/versions, state, proposal/approval/receipt ids; training traces stay 14 days in the restricted sales-quality store, then are deleted. Contacts and browsing records stay out.
Every exact approval binds the current payload and relevant source versions, expires after 30 minutes and is invalidated by relevant changes. Pending, denied or expired approvals cannot authorize an action. The immutable proposal retains its expected state/version, and the application must compare it atomically again before writing. Only a successful receipt tied to the original action and exact payload establishes completion; matching current state and an assistant claim are insufficient. Explicit temporary failures allow at most two total attempts with the same action key and integration-enforced deduplication. Unknown outcomes block new writes pending receipt/state reconciliation.
Cases, evidence and feedback
Each sandbox supplies independent normal, missing-input and unavailable-tool fixtures. The learner records expectations before each actual assistant trial, checks its response against source/policy, gives a verdict, makes one specific revision and records the relevant retest. Written integration and manual arithmetic checks remain explicitly labeled. Worked prose contains fictional examples; those examples are not a learner's observed practice evidence.
Each critique supplies a separate restricted audit view with exactly four operational defects, policy/event ids and successful controls. The learner cites observation, identifies scoped impact and uncertainty, proposes corrections, retains minimal redacted evidence and names the authority and retest needed for restart. The records concern ordinary source, approval, completion and trace-handling failures. No outside access or live action is required. Restricted approval payloads and shared quality traces have different allowed fields.
Measures state numerator, denominator, window, threshold, owner and response. Useful output is paired with unsupported-output, missed-handover, invalid-input or unapproved-action signals as applicable. One forbidden consequential action can pause its affected path; throughput alone cannot justify restart.
The final practice dossier has three separately identifiable sections: workflow design; configuration plus observed trials, fix and retest; oversight audit plus operating handover. Different cases support design, trials and audit.
How results are decided
- A person reviews every CAI-3 result, pass or fail, and makes the decision. The AI score is only a starting point for that reviewer. No result is released automatically.
- Award rule (provisional): an overall mean of at least 2.5 of 4 across the three equally weighted exam tasks, with every task at least 2.0. There is no separate minimum for any single criterion, including recorded AI collaboration.
- The reviewer's decision is final and may differ from the AI score. You can still appeal any result.