01 · Exposure is not authorityThe dashboard knows who clicked complete, not who can do the work

The learning dashboard shows thousands of completions, badges and hours. A line manager still cannot answer a simpler question: which advisers can use the approved assistant on a customer case, detect a fabricated policy citation, verify the answer, abstain when evidence is missing and escalate before harm? The dashboard measures contact with content. The operating system needs a capability claim.

Course completion can establish vocabulary, policy awareness and prerequisite knowledge. It cannot by itself prove that a person can frame a real task, choose or reject AI, protect data, test evidence, use the correct authority, recover from failure and produce a work outcome. Those behaviours depend on the tool, population, interface, queue, manager, reviewer capacity and consequence. A generic prompt course cannot stand in for them.

The unit of upskilling is therefore a bounded work event. Capability is specific to a role, task, tool version, population and authority boundary. A finance analyst may be independent when drafting commentary from synthetic data yet remain supervised when generating statements tied to a live ledger. A recruiter may draft outreach but may not rank applicants. Evidence for one event does not create informal permission for another.

02 · The work eventWork-Embedded AI Capability Contract

The Work-Embedded AI Capability Contract is this article's first synthesis. It ties a business outcome, task, decision and consequence to the learner role, prior experience and accessibility needs. It names the AI contribution, approved tool and version, purpose, data boundary, human authority, prohibited delegation, work standard, quality threshold and time condition.

The contract also defines evidence and provenance, plausible failure modes, warning signals and recovery actions. It fixes the practice environment, supervision, affected-party protection, coach, domain reviewer, line manager and risk support. It states what evidence can widen authority and which model, prompt, data, policy, interface or workflow changes trigger revalidation. Capability expires because the work system changes.

A contract protects the learner as well as the organisation. Accessibility and language support are design inputs, not concessions after assessment. Developmental evidence is separated from punitive performance management unless standards and uses were agreed transparently. Customer, employee and confidential data do not enter a learning portfolio simply because the work is real.

03 · Practice follows consequenceThe Real-Work Learning Loop

The Real-Work Learning Loop is this article's second synthesis: select a bounded work event, record an unaided baseline, model expert practice, rehearse safely, perform in shadow mode, perform supervised live work, review evidence and failures, grant or narrow authority, improve the workflow, then refresh after material change. The loop is not a licence to experiment on customers or employees.

An unaided baseline shows which difficulty belongs to the task, the interface or prior knowledge. Expert modelling makes normally tacit judgement visible, including abstention and no-AI decisions. Simulation introduces adverse, ambiguous and missing-evidence cases. Shadow mode uses live-like inputs without affecting the decision. Supervised work becomes appropriate only when a qualified reviewer can intervene before consequence.

Review examines outcomes and control behaviour rather than polished output alone. The learner must choose sources, verify claims, show provenance, recognise uncertainty and recover from failure. Repeated errors may require more practice, but they may also reveal a broken interface, unrealistic queue, missing reviewer capacity or incentive that rewards speed over checking. The workflow returns to design inside the learning loop.

04 · Evidence changes authorityThe Proficiency Evidence Ladder

The Proficiency Evidence Ladder is this article's third synthesis. Exposure records attendance but makes no capability claim. Explanation proves that the learner can describe the task, failure modes, authority and boundary with a realistic example. Simulation requires representative, adverse and no-AI cases. Shadow performance compares work on live-like inputs without creating consequence.

Supervised live performance allows bounded real work only when a qualified reviewer can intervene. Independent bounded performance requires repeated standards, failure detection and escalation within explicit authority. Coaching and improvement require the person to support others, analyse population errors and improve the work system without widening authority informally.

Progress is not automatic or universal. A claims reviewer may be independent for routine reversible cases and remain at simulation for rare irreversible exceptions. Evidence decays after a model, prompt, data, policy, interface, workflow or population change. Returning temporarily to shadow sampling is a control, not a demotion.

05 · Protect practiceThe Safe Learning Envelope

The Safe Learning Envelope is this article's fourth synthesis. It names the practice population and excluded cases, maximum consequence and irreversibility, approved data and environment, AI and learner authority, supervisor competence and intervention point, review sample and latency, stop condition, rollback, affected-party protection, practice volume, threshold, expiry and responsible owner.

High-consequence work may remain simulated or shadowed indefinitely when no qualified supervisor or safe environment exists. That is a valid control decision. Synthetic data can support early finance practice. Shadow mode can reveal fabricated policy citations without exposing customers. A public decision with irreversible consequences must not become a training ground because a programme needs live examples.

The envelope must also protect workers. People need protected learning time, accessible tools, transparent evidence use and a route to appeal assessment. Employee representation is proportionate where monitoring or job change is material. Developmental work samples must not become covert productivity surveillance or stylistic conformity scores.

Practice volume is not a substitute for representativeness. Ten nearly identical easy cases can create fluent repetition while leaving the decisive failure mode untouched. The cell should deliberately compose a small case set across ordinary demand, missing evidence, adversarial or ambiguous input, population edge cases, time pressure, tool outage, correct abstention and recovery after a wrong first move. The mix should reflect the event's plausible consequence, not a generic benchmark. Reviewers need an answer key for control behaviour as well as output quality: what evidence should have been checked, when the learner should have stopped, who could authorise continuation and what remedy would still be possible.

Supervision is real only when intervention is timely and competent. A nominal reviewer who sees a weekly sample after an irreversible decision cannot protect live practice. Before opening the envelope, the accountable manager should calculate how many cases can arrive, how long review takes, which exceptions require specialist judgement and how rapidly an affected decision can be stopped or reversed. When that capacity is absent, the honest learning design is shadow work, simulation or no training yet. Scarce reviewers are an operating constraint to expose, not a gap to disguise with a completion target.

06 · Evidence without surveillanceWork-Sample Evidence Portfolio and Learning Production Cells

The Work-Sample Evidence Portfolio is this article's fifth synthesis. It contains versioned examples of task framing, source choice, instructions where relevant, outputs, edits, verification, abstention, escalation, decisions, outcomes, failures, reflection and reviewer observation. It records the behaviour required for the role, not tool volume or a preferred writing style.

Personal, customer and confidential data are minimised or removed. Access is purpose-limited. Workers know how evidence will be used, how long it is retained and how to challenge an assessment. Multiple reviewers and calibrated rubrics are appropriate when an assessment changes consequential authority. A certificate may summarise the evidence but cannot replace it.

Learning Production Cells are this article's sixth synthesis. Each temporary cell combines learner, domain practitioner, line manager, AI or product specialist and proportionate control or employee representation. The cell selects work, protects the envelope, reviews evidence, fixes workflow barriers and transfers coaching into the operating team. It is not a permanent training committee or expert correction layer.

A useful portfolio separates an observation from the decision made from it. The artefact may show that a learner verified nine of ten citations, recognised the unverified tenth and escalated it. A calibrated reviewer then decides whether this is sufficient for the named event and consequence. Keeping those layers distinct makes disagreement inspectable and appeals possible. It also prevents one attractive output from overpowering repeated evidence about weak source control, or one recoverable mistake from erasing a wider pattern of safe judgement.

The cell closes when three transfers have occurred. The operating manager can run the evidence gate without the programme team; local coaches can support representative failures without becoming an invisible correction queue; and product or process owners accept the workflow changes revealed by practice. Closure criteria matter because a permanent central cell would make apparent learner independence depend on backstage labour. Its final record should name unresolved exclusions, the authority that was granted, the evidence expiry date and the trigger that will reconvene a short revalidation cell.

07 · Learning must survive the courseThe Capability Transfer Test

The Capability Transfer Test is this article's seventh synthesis. Can the learner perform without the trainer? Recognise when the training example no longer fits? Work under real time, queue and interface constraints? Stop and escalate instead of forcing an answer? Can the line manager observe evidence and support the new work? Has the workflow removed duplicated steps and incentives that block application? Does performance persist across cases and after a delay? Does improvement avoid shifting hidden labour, error or risk to others?

A failed transfer test does not automatically mean the learner failed. A manager may preserve duplicate manual steps, punish slower early practice or provide no review capacity. The interface may hide provenance. The policy may contradict the course. Repair the work system before buying another module. Training transfer research has long located outcomes in trainee characteristics, training design and the work environment; AI does not remove that interaction.

Field studies of AI assistance show bounded effects in specific tasks and populations, sometimes with different benefits across experience levels. They do not establish universal competence. The jagged capability frontier reinforces task specificity: success on one family of tasks can coexist with worse performance outside it. Evidence must remain tied to the event and conditions tested.

Transfer should be sampled at more than one moment. An immediate observation tests whether a learner can perform with recent guidance. A later sample tests retention when prompts, coaches and memorable examples have faded. A changed-case sample tests discrimination rather than imitation. A pressure sample tests whether verification and escalation survive the actual queue. A recovery sample tests what happens after an incorrect suggestion has already entered the workflow. None requires continuous surveillance; a small, declared and proportionate set of work samples is usually more informative than a complete activity log.

The transfer decision must also ask where value and burden moved. Faster drafting is not successful transfer if verification time is silently shifted to a senior colleague, accessibility work is moved to the affected person, or downstream teams inherit poorly labelled uncertainty. The manager should compare the whole work event before and after: elapsed time, rework, reviewer load, exceptions, affected-party outcomes and reversibility. A capability grant is warranted only when the learner can control the task and the surrounding system can sustain that control without concealed harm.

08 · Ten worked casesEvidence gates, not blanket progression

Each case writes the capability contract, places the learner on the ladder, sets the envelope, names work samples, runs transfer, protects workers and affected parties, identifies the coach and accountable manager, and defines the next evidence gate.

09 · Practical instrumentWork-Embedded Upskilling Designer

The designer is this article's eighth synthesis. It combines consequence, current evidence, environment, supervision, transfer, worker protection and change. It returns one of five decisions rather than a universal skill score.

10 · Strong rival and failure boundaryCourses are scalable, consistent and sometimes the safer start

The strongest rival is that common courses and standard assessments are cheaper, more consistent and safer than asking every employee to learn through production work. That rival is partly right. Foundational vocabulary, policy and prerequisite knowledge can be taught efficiently before practice. Real-work programmes can privilege workers with attractive assignments, burden managers, expose customers, create inconsistent standards and turn work output into surveillance.

The answer is layered. Use concise common education, then require role-specific demonstration before changing authority. Use synthetic data, simulation and shadow mode where live practice is unsafe. Calibrate rubrics, use multiple reviewers for consequential gates, provide time, coaching, accessibility and appeal, and keep developmental evidence distinct from punitive performance use.

Do not train yet when the work event, affected population, authority boundary, safe environment, qualified supervisor or accountable manager is unclear. Do not widen authority when a material safety failure persists, even if output quality and tool usage are high. Capability belongs to the person, work and system together.

Evidence and limitsSources checked 3 September 2026

UK guidance supports role-sensitive, inclusive and job-relevant capability but does not prove this article's mechanisms. OECD and ILO evidence describes changing tasks, training and worker outcomes across specific surveys and settings. Field and controlled studies show bounded performance effects; they cannot establish universal learning transfer. Training-transfer, deliberate-practice and organisational-learning research informs the mechanism. NIST supports differentiated human roles, workload and oversight boundaries.

The Capability Contract, loop, evidence ladder, envelope, portfolio, transfer test, production cells and designer are this article's synthesis. Their thresholds require local validation by task, population, jurisdiction and consequence.

1. UK Government, AI Skills Tools Package.
2. UK Government, Skills for AI: What Works for AI Upskilling.
3. UK Government, AI Foundation Skills for Work.
4. UK Government, Digital and Data Profession Capability Framework.
5. UK Government, Artificial Intelligence Playbook.
6. OECD, Artificial Intelligence and the Changing Demand for Skills.
7. OECD, AI and Skills.
8. OECD, The Impact of AI on the Workplace.
9. ILO, Artificial Intelligence Adoption and Its Impact on Jobs.
10. Brynjolfsson, Li and Raymond, Generative AI at Work.
11. Dell'Acqua et al., The Cybernetic Teammate.
12. Dell'Acqua et al., Navigating the Jagged Technological Frontier.
13. Noy and Zhang, Experimental Evidence on the Productivity Effects of Generative AI.
14. Baldwin and Ford, Transfer of Training.
15. Ericsson, Krampe and Tesch-Römer, Deliberate Practice.
16. Edmondson, Psychological Safety and Learning Behavior.
17. Argote and Ingram, Knowledge Transfer.
18. NIST AIRC, AI Risk Management and Human-AI Interaction.