Not all cognitive friction is waste: what should AI automate, and what must humans still practise?

The thirty-minute task

A manager watches an analyst turn a thirty-minute task into a three-minute task with generative AI. The output looks good. The saving is obvious. It is tempting to automate the remaining twenty-seven minutes as though they were one kind of work.

Sometimes that is exactly right. Reformatting a report, copying data between systems or finding a known clause may consume attention without developing anything the organisation values. But the same task may also contain the comparisons through which an analyst learns, the ordinary cases that keep a reviewer alert, and new prompting or verification work that falls to somebody else.

The minutes saved remain real. They are not a complete account of what changed.

The practical choice is not simply automate or do not automate. Managers can use full automation, assisted practice, periodic manual practice or retained human execution. This framework is not a veto on automation. Anyone arguing that manual effort should remain must identify the capability or readiness it protects, who needs it, how much practice is required and how the organisation will know it remains. If those questions cannot be answered and the controls are adequate, the case should lean towards automation.

Four kinds of cognitive friction

The categories can coexist inside one task. They are managerial prompts for investigation, not scores in a validated psychometric scale.

On a narrow screen, each row is shown as a separate card.

A worked example using the monthly report

“Prepare the monthly management report” sounds like one task. At activity level, it produces several different automation decisions.

On a narrow screen, each row is shown as a separate card.

The manager now has a decision for next Tuesday: automate data retrieval and formatting; run assisted anomaly work for a defined trial; retain a human first interpretation of material movements; assign verification to a named role; and review time saved, errors caught, escalation quality and analyst progression.

The report remains one deliverable, but its activities receive different treatment. The answer may also change with the person and the consequence. A trainee may need a first attempt before seeing the machine’s answer. A regulatory filing requires stronger evidence and independent review than a low-stakes internal summary.

Who owns the decision

The manager accountable for the workflow should own the allocation. The practitioner supplies task knowledge; the risk or control owner sets the consequence threshold; and the person responsible for capability development identifies any practice the organisation genuinely needs. The AI team can explain what the system can do, but should not decide what human capability the business can afford to lose.

Where the answer is uncertain, use a time-limited trial with a review date. Retain manual work only against a specific capability or control hypothesis. “People need to keep thinking” is not specific enough, and neither is a successful demonstration sufficient evidence for full automation.

The Cognitive Load Allocation Audit

Use these questions before deployment, then repeat them with evidence from the live workflow. This is an Andrew’s Intel management instrument, not a validated scientific scale.

1. Break the task into activities. Separate retrieval, comparison, judgement, drafting, checking, coordination and exception handling. Do not classify a whole job title or workflow as one block.
2. Remove genuine waste. Which effort can disappear without weakening judgement, learning, awareness or accountability? What evidence shows that automation is reliable enough?
3. Name the capability at stake. Which activity creates or maintains expertise, who needs that practice and how else could the capability be developed?
4. Define the engagement requirement. What must a person continue to observe or do to recognise an exception and intervene effectively? How will intervention be rehearsed?
5. Count the work AI creates. Who will prompt, verify, reconcile, monitor and handle exceptions? Include transferred workload, not only gross minutes removed.
6. Choose the allocation and owner. Select full automation, assisted practice, periodic manual practice or retained human execution. Record who is accountable and what evidence would change the choice.
7. Measure the result after deployment. Track quality and time alongside error detection, exception handling, escalation quality, learning progression and the distribution of checking work.

Four legitimate decisions

Full automation. Use it when the work is predominantly waste friction, outputs and exceptions are observable, and the organisation does not rely on the task to maintain a human capability.

Assisted practice. Let the person perform the consequential reasoning while AI supplies retrieval, examples, feedback or comparison. This can preserve learning while removing avoidable administration.

Periodic manual practice. Automate routine production while requiring realistic independent performance often enough to preserve the skill the organisation expects people to use during an exception.

Retained human execution. Use it when human judgement, responsibility, relationship or intervention capability is central and automation would create more risk or verification work than value. Set a review date rather than declaring the task permanently untouchable.

What the evidence supports

Human-factors research has treated mental workload as a multidimensional design problem for decades. Young, Brookhuis, Wickens and Hancock’s state-of-the-science review warns against treating minimum workload as an automatic optimum: overload and underload can both impair performance, depending on the task and system. ¹

Learning science provides an equally important guardrail. Sweller’s paper showed that conventional means–ends problem solving can consume cognitive capacity needed for schema acquisition, particularly before learners have relevant schemas. Effort is not formative merely because it is difficult. Managers should preserve productive practice, not manual work for its own sake. ²

Theoretical and review work on automation underload and the cliff-edge principle shows the weakness of leaving people as passive monitors and expecting them to recover complex control at the system’s limit. These are not direct tests of generative-AI office workflows, and much of the underlying evidence comes from safety-critical and control environments. Applying the mechanisms to knowledge work is an analogy to test rather than a settled equivalence. The practical question still travels: what evidence shows that a rarely engaged reviewer will recognise an exception and intervene well? ^{3 4}

Generative AI can also move work rather than remove it. Simkute and colleagues synthesise four recurring human-factors mechanisms, including evaluation work, workflow disruption and the tendency to make easy tasks easier while making hard tasks harder; the paper is not a quantified productivity-loss trial. Tankelevitch and colleagues identify the metacognitive work involved in prompting, evaluating outputs and deciding when to rely on them. These mechanisms should appear in the business case. ^{5 6}

A 2025 CHI study surveyed 319 knowledge workers and retained 936 first-hand examples of generative-AI use at work. Higher confidence in AI was associated with less self-reported critical-thinking effort, while higher self-confidence was associated with more. That is useful evidence about perceived work practices, not a causal or longitudinal demonstration of deskilling. ⁷

The evidence supports asking where the thinking moved, who now performs it and whether the required capability remains. It does not supply a universal formula or prove that using AI makes people less capable.

The sentence for the next automation meeting

Before we remove this effort, tell me what capability it creates, what readiness it sustains, and what new work the AI will introduce.

If the answer is nothing material and the controls are adequate, automate it. If the answer is specific, redesign the workflow to preserve only the valuable practice or attention. That is how an organisation removes waste without accidentally removing the way it learns to think.