AI Management

Question: which process should come first?

Lessons: 1.1 Process, task, agent, and outcome; 1.2 Observing how a good employee works; 1.3 Current workflow, inputs, decisions, and exceptions; 1.4 Baseline, target, and prioritization by value, feasibility, and risk.

Lab: interview the process owner, observe five cases, and map the normal path and two exceptions. Measure volume, minutes per case, rework, and wait time. Five cases are for initial exploration, not a final evaluation sample.

Deliverable: current- and future-state process canvas, scoped to the pilot.

Acceptance criterion: named owner; unambiguous input and completion criteria; metric with denominator; baseline observed or marked as a hypothesis; explicit exclusions.

In the tool: register a process and its goal.

Question: what work is each agent responsible for?

Lessons: 2.1 Mission and delivery agreement; 2.2 Inputs, outputs, and quality criteria; 2.3 What it can and can’t do, and when to ask for help; 2.4 Human responsibility and division of work.

Lab: write the qualifier's profile and break down three cases into human and digital activities.

Deliverable: role profile, owner, and responsibility matrix.

Acceptance criterion: specific mission; verifiable output; concrete prohibited actions; owner for quality and exceptions.

In the tool: create an agent linked to the process.

Question: which configuration delivers the required work?

Lessons: 3.1 Model, agent, workflow, and runtime; 3.2 Selection criteria: quality, latency, cost, and data requirements; 3.3 Comparison using the same set of cases; 3.4 Fallback strategies and vendor dependency.

Lab: compare two configurations on the same 20 exploratory cases; record outputs, errors, cost, and time. These cases do not replace the final validation set.

Deliverable: selection assessment with version, limitations, and alternative.

Acceptance criterion: A decision supported by cases and process constraints, without a generic model ranking. Check prices and terms in the provider's sources at the time of the lab.

In the tool: record the model and agent version; actual costs will be added in the future runtime. In the MVP, execution values are simulated.

Question: how can you turn operational experience into reusable instructions?

Lessons: 4.1 Procedures, examples, and rubrics; 4.2 Retrievable knowledge and sources; 4.3 Skills: trigger, inputs, steps, and output; 4.4 Case memory, operational memory, and retention; 4.5 Source quality and conflicts.

Lab: separate ICP, catalog, commercial rules, and examples into instructions, a knowledge base, and two Skills. Create a memory policy specifying what to store, for how long, and who can correct it.

Deliverable: versioned context package, two Skills, and memory policy.

Acceptance criterion: each document has an owner and version; the Skill has entry and exit criteria; unnecessary data is excluded; test data is not used as a training example.

In the tool: record context references and Skills in the form. Files and knowledge retrieval are deferred to the integrated phase.

Question: how does the agent act in the company's systems?

Lessons: 5.1 APIs, connectors, and MCP as means of access; 5.2 Reading versus writing and minimum scope; 5.3 Contracts, validation, idempotency, and retries; 5.4 Server-side credentials, test environment, and evidence of the action.

Lab: design a CRM query and an approved update, including permitted fields, timeout, error handling, and a key to prevent duplication. Run it in a test environment only when integration is available.

Deliverable: contract for two tools and an integration plan.

Acceptance criterion: no secrets in prompts or browser; bounded writing; retries do not duplicate actions; failures are routed with context.

In the tool: describe authorized tools. The MVP does not connect to CRM, email, ERP, APIs, or MCPs.

Question: what can happen without approval?

Lessons: 6.1 Levels 0 to 4; 6.2 Autonomy by action; 6.3 Promotion and demotion criteria; 6.4 Approval tied to the proposal and deadline; 6.5 Pause and rollback.

Lab: apply levels to five actions; build an action × risk × approval × owner matrix. Simulate a prohibited attempt and a change after approval.

Deliverable: autonomy policy and promotion checklist.

Acceptance criterion: critical actions preserve human decision-making; policy does not rely on the prompt alone; promotion requires evidence and an owner.

In the tool: start at a low level, evaluate, and record promotion one level at a time. The local guardrails are instructional.

Question: when does separating responsibilities improve the process?

Lessons: 7.1 When a workflow or a single agent is enough; 7.2 Specialization and handoff contracts; 7.3 Supervisor, dependencies, and budget; 7.4 Stop conditions and loop prevention.

Lab: represent researcher → qualifier → CRM; define each stage's output, shared evidence, attempt limit, and final owner.

Deliverable: team structure and contracts between roles.

Acceptance criterion: each agent has a reason to exist; the supervisor does not expand permissions; the task has cost, attempt, and deadline limits.

In the tool: organize roles by process. Real multi-agent orchestration is part of the roadmap, not the simulator.

Question: how does the manager act when the normal flow is not enough?

Lessons: 8.1 Normal, uncertain, and critical cases; 8.2 Queue, priority, owner, and deadline; 8.3 Evidence package and decision; 8.4 Incidents and interruption; 8.5 Daily supervision routine.

Lab: resolve cases involving incomplete data, requests outside policy, and critical risk; justify approval or return. Measure wait time and review workload.

Deliverable: exception matrix and incident response playbook.

Acceptance criterion: no exception is left without an owner; approving, returning, and interrupting are distinct actions; decisions leave a rationale.

In the tool: simulate execution, review the pending item, and record the human decision.

Question: does the operation deliver enough value to move forward?

Lessons: 9.1 Representative set, answer key, and rubric; 9.2 Correct, acceptable, incorrect, and critical; 9.3 Versions, regressions, and separation of training and testing; 9.4 Accuracy, autonomy, intervention, and time; 9.5 Total cost, capacity freed up, and estimated ROI.

Lab: prepare 100 reference cases with diverse and difficult cases; evaluate the candidate version; calculate cost per case and capacity freed up. Justify sample size and limitations according to the risk.

Deliverable: evaluation report and economic assessment, with a calculation spreadsheet.

Acceptance criterion: clear denominators; critical errors reported separately; human costs included; simulation is not presented as revenue or realized savings; the assessment records uncertainties.

In the tool: record expected/produced cases, human classification, and calculated metrics. The assessment is not generated by AI in the MVP.

Question: how can you improve without losing control of the operation?

Lessons: 10.1 Run, observe, and measure; 10.2 Cause, hypothesis, and change; 10.3 Comparative testing and validation; 10.4 Gradual rollout and rollback; 10.5 Process portfolio management.

Lab: use an exception to propose an improvement, record the candidate and evidence, compare it with the baseline, and deliberate. Design monitoring and rollback.

Deliverable: LOOP-R dossier and the operation's 30-day plan.

Acceptance criterion: changes are not applied based on isolated feedback; testing precedes validation; a human authorizes promotion; the previous version remains identified.

In the tool: record the problem, hypothesis, evidence, and experiment progress; update the agent separately and reevaluate the new version.