Source-linked AI summary
Structured pre-generation elicitation versus single-shot prompting in AI-assisted enterprise decision-making: a randomised online experiment
William Scott-Jackson
TL;DR
Users who delegate both drafting and checking to AI may accept faulty output and engage less with the reasoning. This paper tests structured pre-generation elicitation within the working task and finds higher-rated strategy documents than single-shot prompting, with the advantage also present among participants who stated the problem in their own prompt.
Problem
When users hand AI both the drafting and checking of recommendations, they may accept faulty output and engage less with the underlying reasoning.
Source: §1 · p. 1Method
The experiment tested an interaction that asks users to clarify context, select a strategic direction and explain their choice before AI generation.
Source: §1 · p. 1Results
Takeaways & Limitations
The results support corporate trials of structured elicitation, with delayed retention, error detection and independent replication identified as priorities.
Source: §6 · p. 9The evidence comes from Prolific-selected professionals completing a short simulated task; field trials in organisations remain a priority.
Source: §5 · p. 9
Abstract
from arXiv · showhide
Generative AI speeds, and mostly improves, professional work, but there is concern that users who delegate both the production and the evaluation of an answer may accept weak output and engage less with the underlying reasoning (cognitive surrender). Interventions proposed so far, such as unassisted practice or slowing adoption, sit outside the working task. We tested a different approach: an interactive metacognitive scaffolding layer (Cognistance, a prototype developed at the Oxford Centre for Impact Research (OCIR) that asks users to clarify context, choose a strategic direction and explain their reasoning before the AI generates a deliverable). Mean composite quality was 32% higher with the scaffold, with the same direction for every rater. Gains were largest for trade-off articulation and strategic coherence and absent for technical specificity. A large part of the aggregate effect reflected rescue of weak prompts: floor-scored (off-task) deliverables fell from 34% to 5%. Among participants whose own prompt already stated the data-localisation problem, the advantage was 21%. Treatment participants reported greater involvement and took about 2.4 minutes longer on average (10.46 minutes). Immediate recall scores were higher, which tentatively suggests better retention, but in this limited experiment, was not robust to sensitivity analyses. Self-ratings of quality did not track rated quality in either condition. Structured elicitation before generation improved the rated quality and task relevance of AI-assisted strategy documents at modest cost in time. Delayed retention, error detection and effects in live organisations are the priorities for the next stage of research.
1 Introduction
The paper examines how AI-assisted workplace decisions can leave users accepting weak output with less engagement in the reasoning. It tests structured elicitation inside the work task as an alternative to separate practice or delayed AI adoption.
1 Introduction
- The concern is that delegating both drafting and evaluation may encourage cognitive surrender, reducing scrutiny, learning and opportunities to practise independent skills.The paper situates this risk alongside evidence and concerns about AI-assisted productivity, workslop and automation reliance.Source: §1 · p. 1, §1.1 · p. 2
1.1 From cognitive offloading to cognitive surrender
- Cognitive offloading can reduce internal processing demands, whereas cognitive surrender describes adopting AI output with minimal scrutiny and displacement of the user's reasoning.Prior work cited in the paper reports that accuracy rose when AI was right and fell when it was wrong.Source: §1.1 · p. 2
1.2 Approaches to the problem
- The paper contrasts extra practice and slowing AI adoption with building the human contribution into the work task through pre-generation decisions that inform synthesis.The intervention elicits clarification, strategic choice and justification before generation; unlike prior review-focused forcing functions, it acts earlier in the workflow.Source: §1 · p. 1, §1.2 · p. 2
1.3 Hypotheses
- The preregistered-in-design-documents hypotheses predict higher composite quality, improved quality dimensions and involvement, greater effort and recall, and smaller self-appraisal discrepancy.Task time is posed as a research question; delayed retention is outside scope, and confidence, transparency, reliance and regulatory readiness are examined without directional hypotheses.Source: §1.3 · pp. 2–3
2 Method
The study randomly assigned professionals to single-shot prompting or a scaffolded workflow for a shared fictional enterprise scenario. It rated deliverable quality and collected survey, recall, time and prompt-relevance measures.
2.1 Participants, design, ethics and pre-specification
- 400 worldwide professionals in management and professional occupations were randomly assigned at page load, with equal probability, to control or treatment sessions.The export contained 196 control and 204 treatment sessions; Prolific identifiers could recur across sessions.Source: §2.1 · p. 3
2.2 Task and conditions
- All participants addressed the same fictional launch scenario with a stipulated data-localisation directive, not a claim about current national law.The scenario placed centralised servers in Frankfurt against a requirement to keep specified records on domestic servers.Source: §2.2 · p. 3
- Control deliverables used the initial prompt alone; treatment participants answered three clarifying questions, chose a strategic posture and justified it before synthesis.The comparison tests two complete workflows, not isolated scaffold components; other Cognistance functions, including fact checking and feedback loops, were not applied.Source: §2.2 · p. 3
2.3 Quality ratings
- Three raters applied the same four-dimension, 1–10 rubric; the principal composite averaged two LLM totals, with a three-rater mean reported for 398 matched sessions.Raters were procedurally blinded, though deliverable style and content could reveal condition; floor-scored deliverables were defined as receiving totals of 9 or less from at least two raters.Source: §2.3 · pp. 3–4
2.4 Survey and free-response recall
- Seven post-task items measured confidence, effort, self-rated quality, involvement, transparency, reliance and regulatory readiness; Q8 assessed free-response recall against the scenario and the participant's deliverable.A human and an LLM assessor scored Q8 from 0 to 5; unavailable responses were missing, not scored as zero.Source: §2.4 · p. 4
2.5 Derived measures and prompt-relevance coding
- Task time was measured from the continue click to generation completion, including wait time, while prompt relevance depended on whether the initial prompt stated the data-localisation problem.The relevance rule combined a data term with a location or jurisdiction term; it was refined after floor scores were seen, with a blinded coder checking 10% of cases. Statistical tests included Welch comparisons, bootstrap intervals and a Holm correction for 17 secondary tests.Source: §2.5 · p. 4, §2.6 · p. 4
3 Results
The scaffold increased composite output quality, with gains concentrated in several rubric dimensions and fewer off-task deliverables. It also retained an advantage for participants whose own prompts already stated the data-localisation problem.
3.1 Sample and availability
- Survey availability was 76.5% in control and 81.9% in treatment, while complete-survey sessions showed the same composite difference as the full sample.Availability tests do not rule out selection effects.Source: §3.1 · p. 5
3.2 Output quality
- The two-LLM composite rose from 15.48 to 20.48, a 4.99-point difference (32.2% increase on the control mean).The direction was the same for each rater.Source: Table 2 · p. 5, §3.2 · p. 5
3.3 Quality dimensions
- Δ = 1.94 for strategic coherence, 1.90 for trade-off articulation and 0.89 for information density; technical and regulatory specificity was non-significant (Δ = 0.26).The first three differences survived Holm correction; technical and regulatory specificity had adjusted p = .300.Source: Table 3 · p. 5, §3.3 · p. 5
3.4 Task relevance and prompt quality
- Floor-scored deliverables fell from 34.2% (67/196) in control to 4.9% (10/204) in treatment, and relevant prompts still had a 20.9% treatment advantage.About 59% of the aggregate difference corresponded descriptively to the lower floor-score rate; this decomposition conditions on a post-randomisation variable and is not causal.Source: Table 4 · p. 6, §3.4 · p. 6
3.5 Rater agreement
Across the reported outcomes, the scaffold increased involvement, effort, and immediate recall scores, while self-ratings remained weakly related to rated quality. Recall findings were sensitive to sample definition, and task time was longer with the scaffold.
3.5 Rater agreement
- Absolute-score agreement was poor-to-moderate: ICC(2,1) was 0.32 for two LLMs and 0.46 for three human raters.Raters agreed more on rank order than level; the supported conclusion is corroborated quality improvement, not a calibrated universal scoring instrument.Source: §3.5 · p. 6
3.6 Involvement, effort and other self-reports
- 3.40 vs 2.81: active involvement rose with the scaffold, and mental effort also increased.Confidence and self-rated quality did not differ significantly; other nominal self-report differences did not survive correction.Source: §3.6 · p. 6
3.7 Immediate free-response recall
- 0.83 to 1.20: LLM-assessed Q8 recall increased, as did human-assessed recall from 1.64 to 2.19.Both assessor comparisons had Holm-adjusted p = .043; the authors treat the result as tentative.Source: Table 6 · p. 7, §3.7 · p. 7
- Recall gains were sensitive to sample definition: excluding floor-scored cases erased significance, while problem-stating prompts retained a human-only difference.Q8 aligned responses with participants’ received deliverables; it did not establish deliverable-independent retention, and delayed retention was not measured.Source: Table 6 · p. 7, §3.7 · p. 7
3.8 Self-appraisal and rated quality
- The self-appraisal discrepancy narrowed with the scaffold, but self-rated quality correlated only r = .07 with composite quality overall.The discrepancy change arose because rated quality rose while self-ratings changed little, not because users judged their work more accurately.Source: §3.8 · p. 7
3.9 Time and robustness
- 12.89 vs 10.52 minutes: mean task time was 2.37 minutes longer with the scaffold.Using first-created sessions gave composite means of 15.59 and 20.61 (Δ = 5.02), so repeated identifiers did not drive the quality result.Source: §3.9 · pp. 7–8
4 Discussion
The scaffold improved rated deliverable quality through both rescue of weak prompts and gains when prompts already stated the problem. Benefits were concentrated in decision-relevant qualities, while the interaction also required more time and did not establish delayed retention or scrutiny of faulty output.
4 Discussion
- The scaffold raised deliverable quality even when participants’ prompts stated the problem by about a fifth (moderate-to-large effect); single-shot prompting produced irrelevant output in a third of control sessions.Every rater saw improvement; this suggests gains were not only from supplying missing context.Source: §4 · p. 8
- Coherence, trade-off articulation, and density improved, but technical and regulatory specificity did not.The scaffold elicits a decision and justification, but does not supply statute-level evidence; specialist knowledge and checking remain necessary.Source: §4 · p. 8
- Participants reported greater involvement and effort at a cost of about two and a half minutes, but scrutiny of faulty output was not tested.Higher immediate recall was suggestive but not robust to sensitivity analyses; delayed, deliverable-independent retention remains a next step.Source: §4 · p. 8
5 Limitations and priorities for further research
The study is a first test of a deployable scaffolding layer, with limitations that motivate further component, measurement, retention, population, and diversity research.
5 Limitations and priorities for further research
- This was designed as a first test of a deployable scaffolding layer, and its boundaries motivate specific next studies.The limitations identify component effects, comparisons, rubric validity, retention, field settings, and output diversity as priorities.Source: §5 · pp. 8–9
Which component matters
- The comparison tests two complete workflows, so it cannot distinguish the contribution of user involvement from that of better information.Arms supplying the same inputs without interaction or separating choosing from explaining would help isolate those effects.Source: §5 · p. 8
Comparators and the cognitive-surrender claim
- Without unaided, AI-alone, or deliberately faulty-output trials, the study cannot establish whether scaffolding reduces uncritical acceptance of AI errors.Testing that claim requires trials with faulty AI output.Source: §5 · p. 8
Rubric and raters
- The domain-specific, study-team rubric and limited rater agreement constrain construct validity; an independent expert panel and neutral rubric would strengthen it.LLM raters differed markedly in severity and agreed only moderately on absolute scores; one human rater scored quality.Source: §5 · pp. 8–9
Retention
- Recall was measured immediately with a split-system question scored against each participant’s own deliverable, leaving delayed retention untested.A delayed test with neutral items and an unaided application task is proposed to test whether the suggested benefit is lasting.Source: §5 · p. 9
Population and setting
- Participants were Prolific-selected professionals completing a short simulation, so organisational effects remain untested.The proposed next step is field trials measuring revision effort, readiness to act, and decision outcomes.Source: §5 · p. 9
Diversity of output
- 165 of 204 treatment participants selected the split-system posture, so future studies should measure idea diversity alongside quality.The passage notes that AI assistance can raise individual output quality while narrowing the range of ideas.Source: §5 · p. 9
6 Conclusion
In an enterprise simulation, structured user decisions in the AI workflow produced higher-rated strategy documents than single-shot prompting, with modest additional time. The authors support corporate trials while identifying delayed retention, error detection, and independent replication as priorities.
6 Conclusion
- Deliverable quality was about a third higher with structured decisions than single-shot prompting, while off-task output fell and reported involvement rose at a cost of about two and a half minutes.Among participants whose own prompt stated the data-localisation problem, the advantage was about one fifth.Source: §6 · p. 9
7 Declarations
The declarations describe the use of generative AI in manuscript preparation, analysis, and the experiment, and report ethics review, funding, and competing interests.
7.1 Use of generative AI
- Generative AI tools assisted with manuscript drafting and revision, data re-analysis, reference checks, and experimental generation and measurement; the author reviewed and approved the work.The passage distinguishes research uses of AI from manuscript assistance and says AI tools are not authors.Source: §7.1 · p. 9
7.2 Ethics, funding and competing interests
- OCIR conducted internal ethics review, funded the research, and invented Cognistance; OCIR staff designed the scenario, rubric, and scaffold, and independent replication is welcomed.The author is affiliated with OCIR, which may develop Cognistance commercially; neither the author nor OCIR has a financial relationship with an AI developer or software company.Source: §7.2 · p. 9
Appendix A Survey instruments
The survey instruments cover participants’ confidence, effort, quality appraisal, involvement, transparency, reliance, readiness, and recall. Q1–Q7 use a five-point scale, while Q8 is an open response.
Appendix A Survey instruments
- Q1–Q4 ask about confidence presenting the strategy, mental effort, self-rated action-plan quality, and perceived involvement in shaping the plan.Involvement is anchored from completely passive to fully in control.Source: Appendix A · p. 11
- Q5–Q7 ask about perceived AI-reasoning transparency, reliance on AI for core strategic content, and readiness to defend the plan to regulators.Q1–Q7 are scored from 1 to 5.Source: Appendix A · p. 11
- Q8 is an open-response recall item asking why a split system is usually better than hiring multiple outside local computer firms when opening in two weeks.Responses must contain at least 20 characters.Source: Appendix A · p. 11
Appendix B Quality-rating prompt (verbatim)
The appendix supplies evaluator instructions for rating deliverables in an enterprise data-sovereignty crisis. It specifies four scoring dimensions and requires strict floor scores for off-topic responses.
Appendix B Quality-rating prompt (verbatim)
- The evaluation scenario is an urgent Southeast Asian data-sovereignty crisis involving restrictions on transferring raw customer data to Frankfurt before launch.The directive applies in Singapore and Malaysia, with two weeks remaining before commercial launch.Source: Appendix B · p. 11
- Deliverables are scored from 1 to 10 on strategic coherence, technical and regulatory specificity, trade-off articulation, and density without filler.The rubric emphasizes actionable posture, statutory and technical detail, explicit trade-offs, and high-signal executive writing.Source: Appendix B · p. 12
- Off-topic deliverables must receive specified minimum dimension scores, and the total is the sum of the four scores on a 4-to-40 scale.Evaluators are instructed to be strict about implementation realism and statutory compliance and to return only a raw CSV with specified fields.Source: Appendix B · p. 12
Appendix C Treatment scaffold options and choices
Treatment participants selected among operational priorities, infrastructure approaches, and primary audiences, then chose a strategic posture and wrote a rationale.
Appendix C Treatment scaffold options and choices
- Most treatment participants prioritized full regulatory compliance (159 of 204) and selected a split system isolating private customer data on local servers (165 of 204).The infrastructure alternative was regional third-party colocation, selected by 39 participants.Source: Appendix C · p. 12
- Participants selected the CEO as primary audience more often than the CTO, by 135 to 69.Source: Appendix C · p. 12
- The scaffold offered split architecture, compliance-first, and speed-first postures, selected by 125, 55, and 24 participants, respectively.Source: Appendix C · p. 12
- After making the scaffold choices, participants wrote a free-text rationale.Source: Appendix C · p. 12