AI Workforce

◇ Before you start

Four things get fixed before Level 0 and do not change for eight weeks. Everything you build is built against them.

One: the work record.

List the ten most consequential tasks you personally performed last week. Real tasks, with real outputs: the proposal you wrote, the enquiry you answered, the invoice you chased, the figures you checked. Not categories. Tasks. This list is the raw material for Level 0 and the honest baseline the final project is compared against.

Two: the candidate process.

Name one process you run repeatedly that produces an artefact another person acts on. A proposal from an enquiry. A monthly report from raw figures. A booking from a request. It must be recurring, because you will run it through your workforce in week eight and compare it to how you ran it alone. It must have a real recipient, because your own opinion of your own output is not a measurement.

Three: the confidential boundary.

Write down what may never enter a tool your organisation does not control. Client identities. Personal data. Unpublished figures. Anything under agreement or regulation. Take it from the actual policy if one exists. If none exists, write your own and note that you wrote it. Every coworker you build inherits this boundary, so it gets written once, now, not rediscovered per role.

Four: the approval line.

Write down what never leaves your organisation without your sign-off. Prices. Commitments. Public statements. Anything a court, a client or a regulator could hold you to. This line is where your workforce ends and you begin, and the whole of Level 5 is about defending it.

Write those four down in week one, before any of them has been tested. What you keep of them will be empty for a while, and that is the correct state: an empty record is not evidence of a healthy workforce, it is evidence that nothing has been written down yet, and you do not invent a failure to fill it.

§ Non-negotiable

Seven rules. Where a rule rests on published evidence, it is marked. Where it is a design choice of this course, it is marked as that, because a course about supervision should not smuggle its own preferences past you.

One. Every role is built on your real work. Nothing in this course runs on an invented scenario. Design choice. Delegation problems are specific to the work being delegated, and a generic exercise cannot surface yours.

Two. No coworker works without a role card. Purpose, responsibilities, inputs, outputs, knowledge, boundaries, escalation, standard. Design choice, with evidence behind it. In the largest study of why multi-agent AI systems fail, poor specification and system design was the largest single category of failure, present in around half of the failed traces examined (Cemri et al., 2025). The model was rarely the weakest part of the system. The instructions were.

Three. Work is accepted on evidence, never on the worker's account of itself. A coworker reporting a task complete is a claim, not a fact. You check the artefact against the standard. Design choice. The reason is structural: a system that grades itself will pass itself.

Four. Approval is something the workforce cannot produce. If a coworker can generate, forge or trigger the thing that counts as your sign-off, it is not a sign-off. Design choice. The test is simple: could the workforce produce this approval without you? If yes, redesign it.

Five. Nothing automates until it has run supervised. Automation removes the human start from a process. It must first be a process, observed end to end, with its failure modes known. Design choice. Automating a badly designed process produces a faster badly designed process.

Six. Silence is a fault, not a status. A coworker that returns nothing, a workflow step that stalls, a check that never ran: each is recorded as a failure, never displayed as health. Design choice. The characteristic failure of monitoring surfaces is that absence of data renders as absence of problems.

Seven. You remain accountable. Evidenced as the regulatory position. The EU AI Act requires deployers of high-risk systems to assign oversight to people with "the necessary competence, training and authority" (Article 26(2)), and Article 14 requires that the person overseeing can interpret the output, decide not to use it, and intervene or stop the system. Whatever your jurisdiction, no framework examined for this course transfers accountability to the tool, and neither does this course.

◈ Reference

Five components. The weights are a design choice; what each component demands is not.

Work inventory and delegation analysis, 15 per cent. A tracked week of real work, each task classified, with the bottleneck test applied and defended. Marked on honesty of classification, not on how much turned out to be delegable.

Role design, 20 per cent. At least three complete role cards, each commissioned through a bounded trial with a recorded verdict. Marked on whether the boundaries and escalations reflect the actual work rather than the template.

Knowledge and workflow design, 20 per cent. A starter knowledge base with a source-of-truth table, and at least two workflows mapped with handoffs and approval points. Marked on whether a conflict between sources has been found and resolved, because every real organisation contains at least one.

Governance and control matrix, 15 per cent. A control matrix covering every coworker, with approval categories named and at least one unnecessary approval point removed. Marked on whether the matrix would actually catch the failures the course demonstrated.

Final project, 30 per cent. One complete workflow run end to end through at least two coworkers and one approval point, on real work, with the record kept: what was delegated, what came back, what was corrected, what was measured, and what remains yours. Marked on the record, not the polish.

Nothing is assessed on which tools you used.

† Evidence

Every load-bearing figure in this course carries a citation and one of three grades.

Evidenced means a meta-analysis or several independent studies, or one large study whose full text was read.

One study means exactly that, with the sample stated so you can weigh it.

Convention means no adequate test exists. This course says so at the point of use rather than teaching silence.

A structural warning about capability figures. Agent benchmark results move faster than any course can track, and this course was written in August 2026. Where a figure depends on a model generation, the generation is named. Treat every capability number here as a floor that has probably risen, and every reliability caveat as a pattern that has repeatedly survived model improvements.

A second warning, about the oversight evidence. Nearly all published measurement of humans supervising automated systems concerns discrete recommendations: a score, a label, an alert. Almost none of it measures review of long generated artefacts, where the error is diffuse and verification is expensive. The direction of the findings almost certainly transfers. The magnitudes are unknown, and this course says so where it matters.

Claims you will not find in this course, and why.

"AI agents can now do most office work." Removed. On the most realistic office-work benchmark available, the best 2025-generation model completed 30.3 per cent of tasks fully; on structured tool-calling over clean databases, 2026 models reach the high eighties. The spread between those two numbers is the course.

"A human in the loop catches the errors." Removed as a blanket claim. A meta-analysis of 106 experiments found human-AI combinations performing worse than the better of the two alone on decision tasks. The conditions under which review genuinely works are narrower than the phrase suggests, and Level 5 teaches them.

"Explanations help reviewers judge AI output." Removed. In the strongest studies, explanations increased acceptance of recommendations regardless of whether they were right.

"More agents are better than one." Qualified. Under equal compute budgets, single agents have matched or beaten orchestrated teams on reasoning tasks; orchestration earns its cost on parallelisable work and loses it elsewhere. The course teaches roles for control and clarity, not for performance mystique.

"You can train away automation bias." Removed. The review literature finds the bias persists across expertise levels and survives training interventions. The course designs approval points around the bias instead of pretending training removes it.

‡ Worked to standard

Three real working artefacts from an operating one-person smart organisation, reproduced in their working format and annotated. They are the formats this course teaches, shown at full strength rather than as blank templates.

The role card. A complete digital coworker definition: purpose, responsibilities, inputs, outputs, knowledge, boundaries, escalation, standard. Annotated where each field earns its place by catching a specific failure.

The work order. The packet a coworker actually receives: one job, the files owned, the current decisions, what not to touch, what done looks like, the check to run, the report format. Annotated for what is deliberately absent, because what a worker does not receive is a design decision.

The control matrix. The one-page governance surface for a whole workforce: what runs freely, what is prepared for review, what requires approval, what is never performed, what escalates. Annotated with the approval categories and the reason each row exists.