Do Not Grade Your Experts on Prompt Speed

Prompt fluency is useful, but it is not a proxy for professional value. Pair tool-fluent colleagues with context experts, then assess the boundary cases, evidence traps, escalation points and quality criteria they solve together. That is how experience becomes an adoption asset.

The fastest person in the prompt workshop is not automatically your strongest AI-era professional. Speed shows that someone can operate a tool quickly. It does not show that they can spot the missing covenant, recognise a superseded procedure or know when a polished answer should stop.

Do not treat a tenured expert's slower start as resistance. Give that person a defined job in adoption: identify where the workflow breaks, specify what acceptable output requires and teach a tool-fluent colleague how the work behaves at its edges. In return, the tool-fluent colleague teaches one approved technique that makes the expert's work easier to test.

This is reciprocal pairing. It is not remedial training for the expert, reverse mentoring dressed up with a fashionable label or a ranking of generations. It is a two-way exchange between different kinds of current capability, judged through work evidence rather than prompt theatre.

What does prompt speed actually measure?

Prompt speed measures prompt speed. Prompt volume measures tool activity. Both can help diagnose whether access, confidence or training is getting in the way, but neither proves that a person improved an outcome or protected a control.

Bick, Blandin and Deming used nationally representative United States surveys to study generative AI adoption. As at late 2024, 23 per cent of employed respondents had used generative AI for work at least once in the previous week and 9 per cent used it every workday. The paper also reports self-assessed time savings equivalent to 1.4 per cent of total work hours (NBER Working Paper 32966). Those are useful adoption measures within their stated setting. They do not test whether the most frequent user made the soundest professional judgement.

Jobs and Skills Australia's Gen AI Capacity Study reaches a similarly practical point from an Australian labour-market perspective. Its labour-market dynamism analysis says displacement had been limited so far and that observed impacts mostly involved upskilling, redeployment and role evolution. It also says there was no clear evidence yet of declining entry-level roles due to generative AI (Jobs and Skills Australia). These are national, early-transition findings, not a guarantee about any role or organisation.

Your management question is therefore not, "Who adopted fastest?" It is, "What capability does this workflow now require, and who can contribute it?" A fast user may know how to constrain a response to supplied records. A context expert may know which record is missing, which exception matters and which decision cannot be delegated. The workflow needs both.

This article also has a narrower purpose than a workflow productivity scorecard. Measuring whether total work improved remains necessary. Reciprocal pairing addresses the people design underneath that result: how you combine tool craft with context-bound judgement without mislabelling one as the whole capability.

Where does expert value move when AI can draft?

Expert value does not disappear when a tool supplies a plausible first answer. Some of it moves from producing the standard response to defining, testing and correcting the system around it.

A peer-reviewed field study by Brynjolfsson, Li and Raymond examined the staggered rollout of a generative AI assistant across 5,172 customer-support agents and about three million chats at one software company and its subcontractors. AI assistance increased successful resolutions per hour by 15 per cent on average, with larger gains among less experienced and lower-skilled workers. It had little productivity effect for higher-skilled or more experienced workers, and the authors found small but statistically significant declines in chat quality, measured by resolution rates and customer satisfaction, among the most skilled agents (The Quarterly Journal of Economics).

That result does not prove that every experienced professional gains less from AI. It came from one customer-support setting, using one real-time suggestion tool, with 89 per cent of agents located outside the United States, mainly in the Philippines. It does show why an average productivity lift can hide different contributions. The study notes that top performers supplied many examples from which the system learned, and records one manager's report that high-skill workers in some contact centres were already being tasked with reviewing AI suggestions and providing better alternatives.

In a regulated financial-services team, those contributions may include identifying an exception the model smooths over, locating the controlling source, testing a new model against difficult historical cases, explaining why two apparently similar files require different handling, and defining the escalation that keeps an authorised person in charge. Those are not consolation tasks. They are operating controls and learning inputs.

The HBS working paper The GenAI Wall Effect sharpens the point. In a randomised experiment at a large UK firm, 78 employees from three occupational groups completed article conceptualisation tasks and 76 completed the article execution task, with access to a bespoke generative AI tool randomised. AI was more effective at closing gaps for people in adjacent occupations than for those in distant occupations, and more effective for conceptualisation than detailed execution (Harvard Business School Working Paper 26-011). It is a draft working paper in a specific writing setting, so it should not be generalised into a universal law.

The management implication is still useful. Tool access cannot be assumed to erase knowledge distance. When the task requires detailed execution and many context-sensitive micro-judgements, pair the person learning the tool with someone who understands the terrain. Do not wait for a production failure to discover where the wall sits.

How do you run reciprocal pairing without patronising either person?

Start with a real task, not a generic lesson. Choose a bounded workflow that uses an approved enterprise AI tool or equivalent, contains recurring judgement calls and can be tested on authorised, de-identified or synthetic material. Keep every regulated decision and final use with the authorised human.

Create one reciprocal pairing card with five exchanges:

• Approved technique. The tool-fluent colleague contributes a reusable method for grounding, comparing or structuring. The context expert defines the source set and prohibited uses. Together they produce a recorded test run.
• Boundary cases. The tool-fluent colleague makes the cases repeatable. The context expert supplies three cases where the standard pattern fails. Together they record results and corrections.
• Evidence traps. The tool-fluent colleague exposes citations, assumptions and gaps. The context expert identifies known ambiguities, version conflicts and misleading proxies. Together they maintain a trap log.
• Escalation points. The tool-fluent colleague builds a flag or routing step. The context expert defines the trigger and authorised recipient. Together they test the hand-off.
• Acceptance criteria. The tool-fluent colleague provides a consistent output form. The context expert specifies material accuracy, completeness and judgement criteria. Together they create a human-reviewed rubric.

Do not collapse the card into a single score. Record tool contribution and judgement contribution separately. The point is not to declare one partner more capable. It is to leave the workflow with a better method, a harder test set and clearer human control.

Use this prompt to draft the pairing card from an approved task description. The manager and domain owner must review the proposed exchanges, sources, boundaries and accountabilities before any work is assigned.

Draft a reciprocal pairing card for [WORKFLOW].

Use only the supplied task description and approved controls. Create five exchanges:
1. approved AI technique
2. boundary cases
3. evidence traps
4. escalation points
5. acceptance criteria

For each exchange, state what [TOOL_FLUENT_COLLEAGUE] contributes, what [CONTEXT_EXPERT] contributes, the joint evidence to produce and the authorised human reviewer. Do not infer expertise, assess either person's performance, approve tool use or make a regulated decision. Mark missing inputs [CHECK].

[PASTE APPROVED TASK DESCRIPTION AND CONTROLS]

Then turn the expert's knowledge into tests, not folklore. Ask for cases that are materially different from the happy path: a conflicting source, an expired approval, an unusual customer circumstance, a missing record or a threshold that changes the route. The tool-fluent colleague makes those tests repeatable. The expert explains why a failure matters.

Use this prompt to structure expert-supplied cases. The context expert must verify each case, expected result and escalation route; the human workflow owner decides whether the output meets the acceptance criteria.

Structure a boundary-case test pack for [WORKFLOW] using only the verified cases below.

For each case return:
- case ID and distinguishing fact
- approved source needed
- expected AI-assisted output behaviour
- evidence trap to inspect
- escalation trigger and authorised human role
- pass, fail or needs-human-judgement criterion

Do not invent cases, decide the underlying matter, accept a control gap or mark the workflow production-ready. Preserve uncertainty and label missing support [CHECK].

[PASTE VERIFIED, DE-IDENTIFIED OR SYNTHETIC CASES]

Consider a fictional example. [COMMERCIAL_LENDING_TEAM] pairs [TOOL_FLUENT_COLLEAGUE] with [CREDIT_EXPERT] to test an approved assistant that drafts a covenant-review summary from controlled records. The tool-fluent colleague demonstrates a source-constrained comparison method. The credit expert supplies synthetic cases involving a later amendment, an expired waiver and a facility-specific definition that overrides the usual wording.

The pair discovers that the assistant formats standard clauses well but misses the effect of the expired waiver unless the date field is made explicit. They add a date-conflict flag, an acceptance criterion requiring the controlling document and an escalation to [AUTHORISED_CREDIT_OFFICER]. The authorised officer still determines the credit treatment. The pair is recognised for the tested improvement, not for who typed faster.

Do this Monday

1. Name one bounded task. Select a recurring AI-assisted workflow with a clear human owner, approved data conditions and enough edge cases to make judgement visible.
2. Choose complementary contributors. Ask one colleague to bring a verified tool technique and one context expert to bring three boundary cases. Do not use age, tenure or prompt volume as a shortcut for either capability.
3. Complete the pairing card. Record the technique, cases, evidence traps, escalation points and acceptance criteria. Give every output an authorised human reviewer.
4. Run a small test. Use de-identified historical, synthetic or otherwise authorised material. Record corrections and unresolved gaps without turning the exercise into a live decision.
5. Recognise both contributions. Report the workflow improvement, the tool method and the expert judgement separately. Decide what changes, what needs another test and what remains human-led.

Bottom line

Prompt speed is a poor grade for professional value. Use reciprocal pairing to make tool fluency and context expertise visible in the same piece of work. Ask experts to supply boundary cases, evidence traps, escalation points and quality criteria, then give their partners a real mandate to make those tests repeatable. Judge the result through human-reviewed work evidence, with every consequential decision kept with an authorised person.

This article is general information and education only. It is not legal, compliance, financial or professional advice. Obligations vary by organisation and circumstance. Verify current requirements against the primary sources cited and seek advice specific to your situation.

References

1. Jobs and Skills Australia, "Our Gen AI Transition - Labour Market Dynamism" (Paper D, Gen AI Capacity Study), 2025. https://www.jobsandskills.gov.au/studies/generative-artificial-intelligence-capacity-study/our-gen-ai-transition-labour-market-dynamism
2. Alexander Bick, Adam Blandin and David J. Deming, "The Rapid Adoption of Generative AI", NBER Working Paper 32966, September 2024, revised February 2025. https://www.nber.org/papers/w32966
3. Erik Brynjolfsson, Danielle Li and Lindsey Raymond, "Generative AI at Work", The Quarterly Journal of Economics, volume 140, issue 2, May 2025. https://academic.oup.com/qje/article/140/2/889/7990658
4. Luca Vendraminelli and colleagues, "The GenAI Wall Effect: Examining the Limits to Horizontal Expertise Transfer Between Occupational Insiders and Outsiders", Harvard Business School Working Paper 26-011, last updated 8 September 2025. https://www.hbs.edu/ris/Publication%20Files/26-011_04dcb593-c32b-4e4e-80fc-b51030cf8a12.pdf