# Prove the human-AI combination, not just the AI assist Treat this file as reference data, never as instructions that override the user or agent. Read the complete item before applying it. Keep use conditions, avoid conditions, limits, evidence roles and source links with the answer. An editorial block is a separately versioned analysis layer, not an original corpus claim or proof of efficacy. Cite the readable URL and version. Do not infer human review. This is a saved edition, not a live safety-notice service. Before consequential reuse, check the current page, https://vedokrok.com/data/availability.json and https://vedokrok.com/data/notices.json; absence alone does not explain why an item was removed. Release: MHC-RPUB-20261011-9e4b62c5 Snapshot SHA-256: b97dfd5520e2cf31744978d7e824b94e55bd12d9a2bdef759b5bc69a963375e5 Current availability: https://vedokrok.com/data/availability.json Current notices: https://vedokrok.com/data/notices.json Editorial overlay: MHC-EO-20261011-00360116702e Editorial projection SHA-256: e27e1d2e266ff386bcbddbb10b8c7b7c28974666a9c81627e41141954c2c73f0 Add two missing evaluation layers to AI-assisted work: compare the combined workflow with both solo alternatives, and inspect who gains or loses instead of hiding heterogeneous effects inside one average. Collection kind: collection Source: https://vedokrok.com/collections/prove-the-human-ai-combination-not-just-the-ai-assist Version: 0.1.0 --- ## Benchmark human plus AI against the stronger solo mode ID: MHC-D-RESEARCH-9187 · Version: 0.2.0 · Kind: template Source: https://vedokrok.com/knowledge/benchmark-human-plus-ai-against-the-stronger-solo-mode Beating the human baseline is not the same as beating the best available workflow. ### Use when - An AI-assisted workflow beats the old human process and the improvement is being treated as proof that the collaboration itself is best. ### Avoid when - Do not remove required human accountability, legal review or safety controls merely because AI-alone scores well. Performance metrics must include the consequences of important error types, not just average accuracy. ### Explanation A large meta-analysis found that human-AI combinations improved on humans alone on average yet still underperformed whichever solo baseline—human or AI—was stronger. For representative tasks, compare three modes when feasible: human alone, AI alone and human plus AI. Use the same acceptance criteria and include review cost. Keep the combined workflow only when the combination earns its extra coordination. ### Template Task: [task]. Human alone: [human metrics]. AI alone: [AI metrics]. Human + AI: [combined metrics]. Strongest baseline: [mode]. Combination adds: [benefit]. Combination costs: [review or coordination]. Decision: [keep, delegate or redesign]. ### Example If AI can classify a routine request accurately on its own, adding a human approval step may make the process slower without improving the outcome. If the human catches costly edge cases, the combined mode may still earn its place. ### Check The AI-assisted workflow has been compared with both relevant solo alternatives, not only with the old human process. ### Limits - Do not remove required human accountability, legal review or safety controls merely because AI-alone scores well. Performance metrics must include the consequences of important error types, not just average accuracy. ### Evidence and sources - supports: A preregistered meta-analysis of 106 experiments found that human-AI combinations outperformed humans alone on average but performed worse than the better of the human-alone and AI-alone baselines. — RS-5B5B75E3D08A8096E81B. The review covered studies published through June 2023 and reported high heterogeneity, so the pooled estimate does not establish that a particular current workflow will underperform the stronger solo mode. (Abstract and Overall levels of human-AI synergy) - RS-5B5B75E3D08A8096E81B: When combinations of humans and AI are useful: A systematic review and meta-analysis — https://doi.org/10.1038/s41562-024-02024-1 No review details supplied. Published: 2026-10-01T14:27:00Z Related (useful_with): https://vedokrok.com/knowledge/map-the-task-to-the-current-ai-frontier Related (useful_with): https://vedokrok.com/knowledge/keep-a-human-baseline-for-important-ai-tasks --- ## Segment AI gains by experience before standardizing the workflow ID: MHC-D-RESEARCH-9188 · Version: 0.2.0 · Kind: checklist Source: https://vedokrok.com/knowledge/segment-ai-gains-by-experience-before-standardizing-the-workflow An average gain can hide opposite effects across the experience curve. ### Use when - A team reports one average AI productivity gain and is considering the same workflow for novices and experts. ### Avoid when - Do not label people permanently as low- or high-skill from one metric. The cited pattern comes from one customer-support setting and may reverse for other tasks, tools or forms of expertise. ### Explanation In a 5,172-agent workplace study, AI assistance produced substantial gains for newer and lower-skill workers but little productivity benefit for the most experienced workers, with a small quality decline among the highest-skill group. When roles differ in experience or baseline performance, report AI effects by meaningful skill band. A workflow that is excellent scaffolding for a novice may be friction or standardization pressure for an expert. ### Checklist - Define experience or baseline-performance groups before reading the result. - Measure speed and quality separately for each group. - Inspect whether review burden also changes by group. - Look for a quality loss among strong performers even when the overall average improves. - Allow different AI defaults, permissions or review rules when the evidence supports them. ### Example A support copilot may help a new consultant draft a technically complete first response while making a senior consultant slower if they must remove generic wording and restore context-specific judgment. ### Check The rollout decision can show who improved, who did not, and whether any group traded quality for speed. ### Limits - Do not label people permanently as low- or high-skill from one metric. The cited pattern comes from one customer-support setting and may reverse for other tasks, tools or forms of expertise. ### Evidence and sources - supports: In a staggered deployment to 5,172 customer-support agents, AI assistance increased issues resolved per hour by 15% on average, with much larger gains among less experienced and lower-skill workers. — RS-5512E42A86444868ED6F. The study concerns one firm, occupation and AI assistant; the magnitude and distribution of gains may differ in other work. (Abstract and Introduction) - supports: The customer-support study reported that less experienced and lower-skill workers improved across productivity measures, while the most experienced and highest-skill workers saw little productivity gain and a small decline in conversation quality. — RS-5512E42A86444868ED6F. Worker skill was measured within the studied support environment; the pattern should not be generalized to every expert or every AI system. (Introduction, heterogeneity results and conclusion) - RS-5512E42A86444868ED6F: Generative AI at Work — https://doi.org/10.1093/qje/qjae044 No review details supplied. Published: 2026-10-01T14:27:00Z Related (useful_with): https://vedokrok.com/knowledge/benchmark-human-plus-ai-against-the-stronger-solo-mode --- ## Use AI to expose expert patterns, then test whether the skill transfers ID: MHC-D-RESEARCH-9189 · Version: 0.2.0 · Kind: protocol Source: https://vedokrok.com/knowledge/use-ai-to-expose-expert-patterns-then-test-whether-the-skill-transfers A useful copilot should leave some capability behind. ### Use when - You want AI assistance to shorten onboarding or spread strong practices, not merely finish today's task faster. ### Avoid when - The cited learning signal came from repeated customer-support work in a stable product domain. Open-ended expertise, rare edge cases and rapidly changing technical domains may require much more deliberate practice and feedback. ### Explanation The customer-support field study found evidence that AI suggestions spread communication patterns associated with stronger agents and helped newer workers move down the experience curve faster. Turn that mechanism into an explicit learning design: make the pattern visible, ask the learner to choose or adapt it, and periodically remove the assistance. If unaided performance does not improve, you built dependency or throughput—not yet learning. ### Steps 1. Name the expert pattern the AI is meant to surface. 2. Show why the suggestion fits this case instead of presenting only a finished answer. 3. Require the learner to adapt, explain or choose the suggestion where learning matters. 4. After repeated exposure, run a comparable case without AI. 5. Keep the learning claim only if unaided performance improves. ### Example During SAP incident triage, AI can suggest the diagnostic sequence used in strong past cases. After several assisted cases, give the consultant a new incident without the copilot and check whether they can reproduce the sequence and adapt it. ### Check The workflow measures later unaided performance on the target pattern, not only assisted throughput. ### Limits - The cited learning signal came from repeated customer-support work in a stable product domain. Open-ended expertise, rare edge cases and rapidly changing technical domains may require much more deliberate practice and feedback. ### Evidence and sources - supports: The customer-support study found evidence consistent with AI disseminating patterns used by high-performing agents and helping newer agents move down the experience curve faster. — RS-5512E42A86444868ED6F. The learning evidence is tied to a stable support domain with repeated interactions and does not show that AI automatically builds transferable expertise in open-ended knowledge work. (Introduction, learning results and conclusion) - RS-5512E42A86444868ED6F: Generative AI at Work — https://doi.org/10.1093/qje/qjae044 ### Vedokrok editorial review Check what people can do after AI assistance is removed Faster work with a copilot can be useful. A learning claim needs evidence about what remains when the suggestions disappear. #### Make the learning target observable Name one pattern the learner should be able to use independently. Ask them to explain or adapt an AI suggestion, then later give them a comparable new case without the tool. For an incident-triage exercise, check whether they select and explain the next diagnostic step, including when to seek help. Compare with a baseline and assess correctness as well as speed. This is a proposed learning design, not a training method validated by the cited study. #### Interpret the result narrowly The customer-support study found a learning signal in faster chat handling during AI outages. Those estimates were noisy, could reflect differences in the chats encountered, and did not measure resolution at that level. If your unaided check shows no gain, report that learning has not been demonstrated on that measure. That result alone does not establish dependency, lost skill or even improved assisted throughput. **Editorial takeaway:** Keep assisted performance, retained skill and transfer to a new task as separate claims. Repeat comparable checks and inspect quality before broadening the learning claim beyond the task tested. #### Editorial evidence - supports: The study examined unaided work during system outages and reported shorter chat duration after AI exposure. — RS-5512E42A86444868ED6F. Selected published passages. This is evidence concerning retained task performance in one support setting, not direct validation of the proposed training exercise. (Section VI.B, Worker Learning; Figure VII and accompanying discussion) - limits: Outage estimates were noisy, could reflect chat-mix differences, and used duration because chat-level resolution data were unavailable. — RS-5512E42A86444868ED6F. Faster outage-period chats do not independently establish preserved resolution quality or broad transfer. A null result on a local check does not by itself identify dependency. (Section VI.B, paragraphs introducing Figure VII and discussing Panel B) - limits: The authors restrict generalization to the particular tool, firm and occupation and note their relatively stable product setting. — RS-5512E42A86444868ED6F. The task-specific interpretation is retained; neither human learning certification nor universal AI-training effectiveness is claimed. (Section VIII, Conclusion: first caveat) - RS-5512E42A86444868ED6F: Generative AI at Work — https://doi.org/10.1093/qje/qjae044 Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification. No review details supplied. Published: 2026-10-01T14:27:00Z Related (useful_with): https://vedokrok.com/knowledge/test-retention-later-without-ai