The investment committee spent an hour revising an agent-written memo. Executives changed adjectives, added caveats and challenged the forecast precision. Nobody questioned the first move: the memo framed the acquisition as a choice between rapid entry and lost market share. It excluded a partnership option, treated customer retention as the objective and selected valuation comparables that made the price appear ordinary. The prose received intense review. The decision frame did not.

This is the hidden power of the first draft. It decides what problem is being solved, which facts enter view, what uncertainty deserves a label and which alternatives appear plausible. Later editing occurs inside that representation. An elegant draft can make review feel active while anchoring the reviewer to a narrow map. If agents create every initial map, people may become accomplished editors while losing the habit and capability of drawing an independent one.

01 · Path dependenceThe first draft narrows every later edit

A first draft performs cognitive work before it performs writing. It establishes the unit of analysis, the causal story, the objective and the default. It decides whether the evidence is financial, operational, legal, behavioural or relational. It creates a sequence that makes some objections look central and others look like digressions. Reviewers can replace most sentences without escaping these commitments.

Anchoring is not proof that AI drafts are always harmful. Human drafts anchor too, and good agents can broaden evidence or expose alternatives people miss. The operational issue is whether an independent representation is needed and whether it remains available. A routine follow-up message with a stable purpose, reversible effect and readily checked facts gains little from protected human-first drafting. An investment thesis, incident diagnosis or policy decision may gain a great deal.

The full AI-first rival deserves respect. Let agents prepare everything, teach people to verify, and use outcome monitoring to catch mistakes. This is efficient when output is routine, reversible and cheaply verifiable. It fails when evidence reconstruction is costly, feedback is delayed, critical cases are rare or the reviewer never formed a domain model. A person cannot reliably detect a plausible omission if they no longer know what should have been present.

Verification is often mispriced. Comparing a draft with a known answer is cheap. Determining whether a plausible frame excludes the right evidence can require nearly the original work. The reviewer may need to locate sources, recreate assumptions and imagine missing options before deciding whether the draft is sound. When reconstruction approaches creation cost, AI-first drafting may save expression time while adding anchoring risk. The route should reflect verification complexity, not the apparent ease of reading fluent prose.

The first draft also allocates attention. Material placed early receives more scrutiny; absent material must first be imagined. Headings turn assumptions into categories, and numerical precision can make uncertain inputs look settled. A provenance trace should therefore record excluded evidence and rejected alternatives, not only the sources cited. This creates a surface on which reviewers can ask why the map has its present shape.

02 · Organisational capabilityJudgement can appreciate, move or decay

Cognitive offloading can improve performance by reducing internal demand. Risko and Gilbert describe mechanisms and consequences across several forms of offloading. Sparrow, Liu and Wegner found in their experiments that expected future access changed recall towards where information could be found. These studies do not establish universal deskilling from AI. They show why the distribution of cognition between person and environment matters.

Judgement is an organisational asset when the business depends on people being able to represent unfamiliar situations, compare evidence, detect boundary errors and take over when automation fails. The asset lives in individuals, team routines, shared language, feedback systems and succession coverage. It is perishable because practice frequency, case variety and outcome feedback can change. Approval accuracy on easy recommendations does not prove takeover readiness.

Offloading can free experts for deeper work. An agent can gather routine facts, format records and surface anomalies while a person retains the representation and decision. It can also support learning when interfaces require engagement and provide feedback. Gajos and Mamykina found different learning effects across advice designs in their experimental tasks. The design question is which cognitive contribution remains active, not whether AI is present.

Capability can move rather than simply decline. A professional may practise less recall and drafting while gaining skill in evidence architecture, simulation or adversarial review. The asset register should recognise the new capability and still test whether any old capability remains necessary for recovery. Organisations should avoid nostalgia for effort that technology has safely removed. They should also avoid assuming that a new supervisory skill automatically contains the domain understanding needed to detect a boundary failure.

Team coverage matters because independent performance can be concentrated in one person. A high-performing expert who reviews every difficult case may hide weak succession. Record how many people can frame the problem, detect plausible errors and take over within the required time. Coverage should include availability and authority, not merely course completion. If a critical capability has one practitioner, delegation can deepen the operational single point of failure.

03 · Original constructTwo linked records

The draft provenance trace records who or what framed the problem and boundary, selected the objective and affected population, gathered and excluded evidence, generated alternatives and counterfactuals, expressed uncertainty, challenged the draft, made the decision and accepted residual risk. Provenance should identify meaningful contributions rather than log every keystroke. It reveals whether a human review introduced an independent view.

The judgement asset register names the critical capability and accountable role, consequence if it is unavailable, frequency and quality of unassisted practice, time since outcome feedback and error review, individual and team coverage, evidence of plausible-error detection, conditions requiring unaided exercise and a recovery plan. It is not an employee score. It is a team-level operational record with proportionate access and safeguards.

The two records answer different questions. Provenance shows whether a particular decision contained independent cognition. The asset register shows whether the organisation could supply that cognition when needed. A well-traced case may still rely on a capability held by one ageing expert. A well-covered capability may remain unused in a particular decision because advice arrived first. Link the records through the capability, role and consequence, while preserving the distinction.

Record quality should be tested through use. Select a consequential case and ask whether the trace lets another qualified person reconstruct the frame, evidence boundary, alternatives, disagreement and final authority. Select a capability and ask whether the register points to recent task evidence, feedback and a viable recovery owner. Empty fields and generic labels reveal a ceremonial system before a real incident does.

The register must name consequence. Preserving manual skill solely because a role once performed it is false manualism. If independent capability is unnecessary, economically disproportionate and not part of recovery, safe automation should remove it. If the capability protects safety, rights, capital, customer trust or operational recovery, the organisation needs evidence that enough people can still exercise it.

04 · Work routingFive modes for first-draft cognition

The route depends on consequence, novelty, ambiguity, observability, reversibility, feedback delay, learning value, available expertise and the cost of losing capability. It should not follow hierarchy. Novices need guided construction of mental models. Experts need calibration against rare boundary cases. Both may use AI heavily when the task and evidence support it.

Urgency changes the size of the protected contribution, not necessarily its purpose. In an incident, responders may have seconds to record observed facts and competing hypotheses. In a policy decision, blind parallel work may take days. Define the minimum independent representation that can change the decision under the available clock. If even that contribution cannot be supplied for a high-consequence case, the organisation needs a safe default, stronger coverage or a suspension path.

Route changes must be visible. A routine customer message can move from AI-first to human framing when it concerns a vulnerable population or a novel complaint. A recurring expert task can move from human framing to calibration after the objective and evidence boundary stabilise. A failed calibration moves the affected scope to recovery. Bind the route to case facts and capability evidence rather than to a permanent job title.

Sampling completes the AI-first route. Select random outputs, changed contexts and near-boundary cases; connect findings to the people using the draft and the team maintaining the service. Feedback should include customer response, correction, complaint and downstream repair where relevant. If reviewers repeatedly rewrite the objective or add missing evidence, the task is not as routine as classified. Move it to human framing instead of disguising judgement work as editing.

Blind parallel work needs information symmetry. Human and agent should receive the same authorised decision facts unless the difference itself is part of the experiment. Neither should see the other's recommendation before completing a frame. Reconciliation then compares objectives, evidence exclusions, causal assumptions, options and uncertainty. Merging prose too early destroys the independence the route is designed to create.

05 · Capability operationsPractice that resembles the real decision

Practice should target the representation and error modes that matter. A junior analyst must build a customer-economics model from raw evidence, explain exclusions and compare alternatives before receiving a polished draft. An expert reviewer needs rare, plausible boundary cases that test takeover and disagreement, not a quiz on policy wording. A high-consequence team needs rehearsals that include time pressure, incomplete evidence and changing state.

Ericsson, Krampe and Tesch-Römer emphasised purposeful practice with feedback in their expertise account. Their theory should not be treated as a complete explanation of expertise. For this operating model, the practical lesson is narrower: mere repetition is not enough. Cases must stretch the relevant representation, reveal error and provide outcome feedback. A ceremonial monthly manual draft that nobody evaluates preserves little.

Feedback latency determines what can be learned. If investment outcomes emerge years later, the organisation needs nearer signals: forecast calibration, assumption tracking, post-decision evidence and peer review. Incident response can use exercises and after-action reviews. Customer work can connect advice to retention, complaint and recovery. The register records when meaningful outcome feedback last arrived, not when a manager last complimented a document.

Calibration cases should be protected from contamination. If staff know the answer, see the AI recommendation or repeatedly encounter the same planted error, the exercise measures recall. Maintain a small governed case library drawn from past incidents, near misses and changing policy boundaries. Refresh it, control access and include more than one plausible path. Scoring should examine representation and evidence use alongside the final answer.

Coaching connects failure to recovery. A reviewer who misses a plausible error needs to see which cue mattered, practise the underlying distinction and demonstrate it in a fresh case. Simply lowering a score or increasing surveillance does not rebuild judgement. The register records the capability gap, practice plan, coach, protected time and evidence of return. Individual results should be handled proportionately and used for learning and safe assignment.

Novice practice should progress through support levels. Begin with a worked example that makes representation visible, move to partial scaffolds, then require a fresh frame before advice. Compare the novice's choices with the agent and an expert, and discuss why differences matter. Production deadlines cannot carry the whole apprenticeship obligation. Reserve protected cases and recognise coaching as real work, otherwise polished AI output crowds learning out of the operating calendar.

Expert practice needs variety rather than volume. Select rare combinations, evolving rules and cases where a familiar heuristic fails. Ask the expert to identify missing information, define a stop condition and explain how the decision would change under another assumption. Outcome reviews should include cases where the agent was right and the expert was wrong. Preservation is calibrated competence, not institutionalised distrust of automation.

06 · Negative controlsSpeed, anchoring and false manualism

Incident responders need AI speed, but a causal narrative can suppress competing hypotheses. Require responders to record observed facts, affected state and at least two live hypotheses before revealing the agent's synthesis. The agent can then search, correlate and propose tests. The trace retains where the human and agent frames diverged. Urgency is handled by making the human-first contribution small and decision-useful.

A junior who only revises polished models learns the surface of a good answer without building the underlying representation. Protect a sequence of raw-evidence cases, coached construction and comparison with the agent. Conversely, do not force every routine draft onto a novice. Use AI to remove formatting and retrieval while keeping the objective, assumptions and model choices active.

Expert decay can hide behind strong approval metrics. If recommendations are usually correct, agreeing with them looks accurate. The expert may still fail a rare unassisted case. Scheduled calibration must include plausible AI errors, boundary conditions and recovery. If performance falls below the defined requirement, approval work continues only within a reduced safe scope while capability is rebuilt.

False manualism is visible when the human first draft introduces no independent objective, evidence, alternatives or uncertainty. It adds delay and protects a ritual. Remove it. For standard follow-up messages, let the agent draft, sample outcomes and return feedback. Preserve judgement where consequence and cognitive contribution warrant it, not where a hierarchy wants to see a person's name first.

Explanation placebo is another negative control. A detailed rationale can make an AI draft easier to follow while reinforcing its frame. Require access to primary evidence and a prompt that asks what would change the conclusion. Where framing is consequential, explanations arrive after a provisional human view or alongside a blind draft. Comprehensibility is useful, but it is not evidence of independent review.

Approval volume can hide a learning deficit. A senior person may review hundreds of accurate routine drafts and receive no practice on the rare decision that justifies their authority. Route routine work away from scarce attention, as Paper 058 argued, and use some released capacity for relevant calibration. Attention and capability are linked: meaningful approval needs both time now and judgement preserved over time.

07 · Worked caseReopening the investment frame

Return to the acquisition memo. The provenance trace shows that the agent framed the choice, selected comparables, excluded partnership evidence and generated the alternatives. Humans edited language and approved the final document, but no independent frame existed. Because consequence is high, feedback is delayed and anchoring is hard to observe, the next case uses blind parallel work.

The human drafter receives the decision facts and mandate without the agent's recommendation. The agent works from the same authorised evidence boundary. A reconciler compares objectives, exclusions, alternatives, uncertainty and valuation assumptions before any prose merge. The committee decides with the disagreement visible. The cost is justified by consequence and the need for an independent representation, not by a belief that the human draft is automatically better.

The final memo retains a difference appendix. It identifies the partnership option, competing objective, excluded evidence, assumptions that drove valuation and why the committee chose its route. This appendix is not a transcript of brainstorming. It is decision evidence that can be revisited when outcomes arrive. Later review can then update both the investment method and the judgement asset register.

If the committee later finds that no member can build a defensible valuation frame without the agent, high-consequence delegation pauses. Live acquisitions route to qualified coverage or a narrower authorised method. Members complete coached cases using raw evidence, compare results with outcomes and demonstrate boundary detection. Delegation resumes in stages, beginning with human framing before returning to blind parallel or calibrated assistance.

08 · Interactive instrumentJudgement Preservation Router

The Router makes the preservation choice explicit. It does not infer cognitive state or rank employees. Use task outcomes, calibration cases, evidence-use patterns and team coverage with lawful safeguards. Set local thresholds with the accountable business, workforce, risk and learning owners.

09 · Operating disciplineProduce outcomes and the capacity to challenge them

Start with a small inventory of capabilities whose loss would change a consequential decision or recovery. Each entry should state the decision it protects, the credible failure if unavailable and the safe service condition while capability is uncertain. Define the task evidence that proves independent performance. Assign an owner, coverage requirement, practice pattern, feedback source and recovery plan. Avoid a sprawling register of every professional skill. Preservation becomes credible when it is selective and tied to consequence.

Governance should separate workforce development from employment surveillance. Use case results, evidence choices, feedback and team coverage collected for a stated operational purpose. Limit access, retention and secondary use. Do not infer cognition from cameras, keystroke rhythm or hidden behavioural proxies. Employees should understand the capability boundary, exercise and remediation path. Trust is part of the learning environment and the control.

Use the provenance trace on sampled production cases. Measure whether frames, evidence exclusions, alternatives and disagreements are visible. Compare human-first, AI-first and parallel routes on outcome quality, time, confidence, error detection and learning. Treat self-reported effort as perception evidence, not direct proof of skill loss. Lee and colleagues' 2025 survey is useful within that boundary.

Boards and executives need a compact view: which judgement assets protect high-consequence outcomes, where coverage is thin, which routes are in use, when calibration last succeeded and which recovery actions remain open. They do not need employee rankings or an average judgement score. A red condition should name the affected capability, consequence, evidence gap, interim control and owner.

Control owners should review the register after incidents, major model or workflow changes, role redesign and sustained shifts in case mix. A capability can become more important when automation removes routine exposure and leaves people only the difficult tail. It can become less important when a process is made deterministic, reversible and independently testable. Preservation effort follows the current production design and credible consequence, not tradition.

Tankelevitch and colleagues' provocations provide one bounded intervention: short critiques and alternatives can prompt reflection in studied knowledge-work tasks. Buçinca and colleagues found cognitive-forcing designs could reduce overreliance with an experience cost in their experiment. These are design candidates, not universal controls. Test whether they improve decision evidence under real urgency and expertise.

Interface evidence should show whether a reviewer opened primary sources, introduced a new alternative or changed an assumption, but these signals require interpretation. Opening a document does not prove comprehension, and many sound decisions legitimately agree with AI. Combine interaction evidence with sampled reasoning, outcomes and calibration. Use it to improve the work design at team level, not to infer private mental states or automate discipline.

When calibration fails, reduce assistance within the affected scope, route cases to competent coverage, and run coached unassisted practice with outcome feedback. The recovery boundary should be observable: detect planted and natural errors, form an adequate representation, use relevant evidence and make the required decision. Resume delegation gradually and continue sampling. Hiding a failed calibration to protect throughput converts a learning signal into operational risk.

Recovery has service consequences. Reducing assistance may slow delivery, so the owner must narrow demand, invoke qualified backup or communicate a temporary service level. Do not pressure the same weakened team to complete extra manual work at the old automated volume. The recovery plan includes safe capacity, case prioritisation, customer protection and a decision date. If the capability is no longer viable to maintain internally, redesign the authority and sourcing model explicitly.

The system should learn from disagreement. Track whether human and agent frames differ on objective, evidence, alternatives, uncertainty or action. Review outcomes without treating agreement as success by default. Repeated agent advantages may justify more assistance; repeated human advantages may justify protected framing or product repair. Differences clustered around one population or policy edge reveal where the route boundary needs adjustment. This is richer evidence than acceptance rate.

Human judgement is retained neither by banning AI nor by adding a signature after an agent draft. It is retained when people still form independent representations where those representations matter, practise the rare work they may need to take over, receive outcome feedback and hold real authority to disagree. The organisation should be able to show which cases exercised that capacity, what was learned, where coverage remains thin and how a failed boundary changes live work. That capacity must remain usable when conditions change. Routine drafting can move to agents. Consequential cognition must remain evidenced, refreshed and recoverable.

SourcesReferences

1. Bainbridge, L. (1983), “Ironies of Automation”, Automatica. doi.org/10.1016/0005-1098(83)90046-8.
2. Endsley, M. R. and Kiris, E. O. (1995), “The Out-of-the-Loop Performance Problem and Level of Control in Automation”, Human Factors. doi.org/10.1518/001872095779064555.
3. Parasuraman, R. and Riley, V. (1997), “Humans and Automation: Use, Misuse, Disuse, Abuse”, Human Factors. doi.org/10.1518/001872097778543886.
4. Risko, E. F. and Gilbert, S. J. (2016), “Cognitive Offloading”, Trends in Cognitive Sciences, 20(9), 676–688. doi.org/10.1016/j.tics.2016.07.002.
5. Sparrow, B., Liu, J. and Wegner, D. M. (2011), “Google Effects on Memory”, Science, 333(6043), 776–778. doi.org/10.1126/science.1207745.
6. Buçinca, Z., Malaya, M. B. and Gajos, K. Z. (2021), “To Trust or to Think”. arxiv.org/abs/2102.09692.
7. Gajos, K. Z. and Mamykina, L. (2022), “Do People Engage Cognitively with AI?”. arxiv.org/abs/2202.05402.
8. Bansal, G. et al. (2021), “Does the Whole Exceed its Parts?”. arxiv.org/abs/2006.14779.
9. Drosos, I. et al. (2025), “It Makes You Think: Provocations Help Restore Critical Thinking to AI-Assisted Knowledge Work”. arxiv.org/abs/2501.17247.
10. Lee, H.-P. et al. (2025), “The Impact of Generative AI on Critical Thinking”, CHI 2025. doi.org/10.1145/3706598.3713778.
11. Brynjolfsson, E., Li, D. and Raymond, L. R. (2023), “Generative AI at Work”. arxiv.org/abs/2304.11771.
12. Dell'Acqua, F. et al. (2025), “The Cybernetic Teammate”, NBER Working Paper 33641. doi.org/10.3386/w33641.
13. Ericsson, K. A., Krampe, R. T. and Tesch-Römer, C. (1993), “The Role of Deliberate Practice in the Acquisition of Expert Performance”, Psychological Review. doi.org/10.1037/0033-295X.100.3.363.
14. Kellogg, K. C., Valentine, M. A. and Christin, A. (2020), “Algorithms at Work”, Academy of Management Annals. doi.org/10.5465/annals.2018.0174.
15. NIST (2023), “Appendix C: AI Risk Management and Human-AI Interaction”. airc.nist.gov Appendix C.
16. NIST (2023), “AI Risk Management Framework Core”. airc.nist.gov AI RMF Core.
17. European Union (2024), Regulation (EU) 2024/1689, Article 14. eur-lex.europa.eu/eli/reg/2024/1689/oj.