(21) How to Brief an Agent: The Lost Art of Structured Delegation in Knowledge Work

July 26, 2026/15 min read

The models have converged. The gap has not.

As of late 2025, frontier language models, GPT-5.2 Pro, Gemini 3 Pro, Claude Opus 4.5, perform within a narrow band of each other on reasoning benchmarks. Stanford HAI's AI Index 2026 quantifies how narrow: the top four models on the Arena Leaderboard sit within 25 Elo points of each other, the US-China gap has closed to 2.7%, and in professional domains, tax, legal, corporate finance, mortgage processing, the top fifteen models are separated by only three to five percentage points. The Index's own conclusion is unambiguous: "with capability no longer a clear differentiator, competitive pressure is shifting toward cost, reliability, and real-world usefulness." Model selection is approaching a secondary decision. What it reveals, once model capability is set aside, is the variable that actually explains performance variance in production: the quality of the task specification the human provides.

An MIT Sloan study published in August 2025, involving nearly 1,900 participants across three versions of OpenAI's DALL-E, found approximately half of all performance gains from upgrading to a more capable model came from the model itself, the other half came from users adapting their prompts, who wrote 24% longer, more descriptive prompts on the newer version. More troubling: automatic GPT-4 prompt rewriting in place of user adaptation degraded performance by 58%. The model could not improve on what the user meant to say. Dell'Acqua et al.'s pre-registered RCT with 758 BCG consultants reinforced this: AI-assisted consultants completed 12.2% more tasks, finished 25.1% faster, and produced outputs rated 40% higher quality inside the technological frontier, but on one task deliberately placed outside it, AI users were 19 points less likely to produce correct solutions than controls, unaware they had left the model's reliable range. A newer data point sharpens this: a Context Engineering study of 200 documented interactions over four months found tasks with explicit, file-based "Authority" context achieved 89% first-pass acceptance versus 29% where authority was absent or inferred, a 60-point gap attributed to specification structure rather than model choice, though the dataset is observational and single-operator. Specification quality and frontier awareness together determine the outcome. The model is the constant.

Structured delegation is not prompt engineering.

Most enterprise AI deployments embed a category confusion that needs resolving before the quality problem can be addressed. Prompt engineering is a technical sub-discipline of syntactic optimisation, few-shot example selection, and chain-of-thought formatting. Structured delegation is categorically different: it is the complete contract between a human principal and an agentic system, a task specification defining objective, scope, non-goals, and constraints; a decomposition into subtasks with dependency order and stop conditions; acceptance criteria constituting the definition of done; an evaluation design with test sets and regression suites; and escalation protocols specifying when to defer to a human. A Google DeepMind paper (arXiv 2602.11865, February 2026) formalises one core requirement: tasks must be recursively decomposed until their outputs are inherently verifiable. A companion paper, Agent Behavioral Contracts (arXiv 2602.22302), defines the contract as preconditions, invariants, governance rules, and recovery procedures: across nearly 2,000 sessions, contracted agents caught 5.2 to 6.8 soft constraint violations per session that uncontracted baselines missed entirely, in a three-stage loan-processing benchmark analogous to financial services delegation chains. Its framing is blunt: "prompts are not contracts; trust is not governance."

Anthropic's engineering guidance on tool design illustrates the precision required at the boundary layer: precise refinements to tool descriptions, without model changes, enabled Claude Sonnet 3.5 to achieve state-of-the-art performance on SWE-bench Verified, "dramatically reducing error rates and improving task completion." Morgan Stanley's deployment is the canonical enterprise implementation: before reaching 16,000 advisors, advisors and prompt engineers graded outputs for accuracy and coherence, followed by a daily regression suite, with answers generated exclusively from internal content and linked to sources. The specification artefacts were the product. The model was the infrastructure.

The more expert you are, the harder briefing becomes.

The specification quality problem is not evenly distributed across a firm's workforce. A OneStream survey of 2,504 finance professionals found 89% of finance students reported sufficient AI skills for workplace use, versus only 54% of professionals with ten or more years of experience. The practitioners with the deepest domain knowledge have the lowest confidence translating that knowledge into AI-ready form. Cognitive science explains why: expertise is substantially constituted by automaticity and chunking, the consolidation of procedural sequences into intuitive pattern responses that bypass working memory. A credit analyst who has reviewed hundreds of leveraged buyouts does not consciously walk through each valuation step; judgment emerges from pattern recognition built over years. This is precisely what Michael Polanyi meant by "we can know more than we can tell." Nonaka and Takeuchi's SECI model identifies externalisation (converting tacit knowledge into explicit form) as both the most valuable and most difficult of the four knowledge conversion modes.

This pattern has concrete financial consequences. The banking literature on soft information establishes that a meaningful share of underwriting quality is not captured in documented records. Liberti and Petersen's 2019 review provides the conceptual framework; separate empirical work on loan officer absence and turnover shows the consequence directly: borrowers whose loan officers are unexpectedly absent experience reduced credit access and worsened delinquency, evidence relationship-specific knowledge does not transfer in codified form. FinRegLab's 2021 research documents how traditional lending relies on face-to-face assessment and cumulative employee knowledge that resists codification. In structured finance, the equivalent knowledge, how to read management credibility in a distressed restructuring, cannot be transferred to an agent by filling out a template. I encountered this directly at Jaja Finance: deploying an agent for credit analysis, the hardest part was articulating evaluation criteria for outputs my team would recognise as good without being able to fully specify what "good" looks like. A banking-sector practitioner outside academia put the same point plainly: judgment "lives in emails, Slack threads, and often in people's heads" rather than in any documented process.

This problem is forty years old. We ignored it twice.

The knowledge acquisition bottleneck is not a discovery of the LLM era. Edward Feigenbaum at Stanford, in the first commercial wave of expert systems, identified knowledge acquisition, extracting expert judgment into computable rule sets, as the primary constraint on expert system performance. Expert systems succeeded in narrow, codifiable domains (MYCIN for bacterial infection diagnosis, XCON for computer configuration) and failed at open-ended judgment. The 1988 literature identifies three consistent failure modes, inadequate expert time allocation, incompleteness of any single expert's domain knowledge, and irresolvable conflicts among experts, each mapping directly onto the current agentic briefing problem. The crucial difference is that LLMs are far more forgiving of underspecification: they produce plausible-sounding outputs from sparse instructions. Expert system failures were immediate and obvious; LLM failures are plausible and invisible until downstream consequences emerge.

The surgical checklist offers a parallel to the resistance financial services will experience. The WHO surgical safety checklist study in the New England Journal of Medicine (2009) documented a 47% reduction in inpatient deaths and a 36% reduction in complications, yet only around 10% of US hospitals had implemented it by 2010, resisted by senior surgeons who viewed checklists as an affront to professional judgment. The data showed the opposite: externalising procedural knowledge improved performance precisely because complexity and cognitive load create blind spots for automaticity. Senior credit analysts reviewing AI outputs under time pressure are in structurally identical territory. The specification investment feels like a demotion. It is the opposite.

The enterprise failure data is consistent and consistently underweighted.

RAND Corporation's 2024 study, based on interviews with 65 experienced data scientists and engineers, found more than 80% of AI projects fail, roughly twice the failure rate of traditional IT projects, and identified misunderstood problem definition as the primary root cause. Gartner predicts more than 40% of agentic AI projects will be cancelled by end of 2027. Deloitte's December 2025 analysis found only 11% of organisations are actively using agentic AI in production, with a dominant failure mode Deloitte terms "workslop": "attempting to automate current processes rather than reimagine workflows for an agentic environment." McKinsey's State of AI 2025 report finds AI high performers are 2.8 times more likely to redesign workflows before deploying AI. Smarsh's April 2026 analysis reports financial services generative AI failure rates between 50% and 95%, citing firms that "pursued use cases without defining measurable outcomes." Stanford HAI's documentation of 847 agentic AI failures across 23 enterprises found that in 67% of cases the team could not reconstruct the agent's decision chain after the fact; the cost became concrete in November 2025, when a European bank's mortgage pre-qualification agent used unscoped data retrieval to access prohibited fields including ethnicity and marital status. The breach went undiscovered until a January 2026 audit and resulted in a €12 million fine.

The strongest counterargument is that AI raises the floor more than the ceiling. The BCG study found below-median performers gained 43% while above-median performers gained only 17%, and Brynjolfsson's customer service research found novice workers gained approximately 34%, with lower gains for experienced workers. If AI distributes encoded expertise to junior staff such that their outputs approach senior-staff quality, demand for high-quality expert briefing may be lower than the thesis implies. This holds for tasks well within the model's capability frontier, but fails where it matters most: for novel deal structures, qualitative restructuring assessments, and credit decisions at the edge of the model's training distribution, the BCG outside-frontier finding reverses entirely. Junior staff with inadequate specifications produce outputs that appear competent and are wrong, and neither they nor their supervisors reliably detect the failure. The levelling narrative holds for within-frontier work; the thesis applies to frontier and beyond-frontier work, which in financial services is also the most consequential.

Regulation is converting specification quality into a compliance obligation.

Regulatory frameworks are converting structured delegation from a quality preference into a legal requirement on an accelerating timetable. Under the FCA's SM&CR, delegating decisions to algorithms does not dilute senior manager liability, and the FCA's Mills Review (July 6, 2026), the first regulator-initiated global agentic AI review in retail financial services, found plainly that firms "have not yet clarified accountability for autonomous decisions made on a customer's behalf." The PRA's SS1/23 applies model risk management principles (documentation, independent challenge, and ongoing monitoring) to AI models; an undocumented system prompt encoding tacit judgment is a failure under those principles. Broader application is arriving faster than the Bank of England anticipated: on April 1, 2026 the BoE and PRA told the Chancellor existing rules were sufficient, yet by June 30 Deputy Governor Sarah Breeden reversed that framing at the ECB Forum, stating "our frameworks were not built to contemplate autonomous agents"; the July 7 Financial Stability Report went further, naming AI a systemic risk for the first time. Regulators who insisted existing frameworks were adequate reversed within a single quarter. The EU AI Act timetable has also moved: the Digital Omnibus on AI, adopted June 29, 2026, defers standalone high-risk obligations, including credit scoring under Annex III, from August 2, 2026 to December 2, 2027, and embedded AI from August 2027 to August 2, 2028; only Article 50's transparency obligations stay on the original date. That runway does not change Article 14: the overseeing human must meaningfully evaluate outputs, which requires specifying what good looks like before the deferred deadline. FINRA's 2026 report added an "Emerging Trends in GenAI: Agents" section naming "Domain Knowledge" gaps as a distinct agent risk, providing regulator-level validation of this article's tacit-knowledge argument. Firms without structured delegation artefacts are approaching enforcement deadlines with undocumented decision logic in production.

The skill does not exist as a formal competency anywhere.

Structured delegation, task decomposition, acceptance criterion design, edge-case enumeration, evaluation loop design, is not prompt engineering, domain expertise, or AI literacy. It is a distinct professional competency combining all three. To our knowledge, no CFA curriculum includes it, no ACA qualification covers it, no MBA programme offers it, and no CRO certification examines it. Accenture's March 2026 workforce analysis states plainly that human roles in financial services are evolving "from monitoring and feedback at early stages to orchestration and strategic oversight." Academic formalisation is advancing: the principal-agent framework applied to multi-agent ML systems (arXiv 2601.23211v1, January 2026) and the "moral crumple zone", coined by M. C. Elish (2019), applied to AI systems in the California Management Review's 2025 analysis, provide theoretical scaffolding. But as of this writing, no major financial services firm has published a validated methodology, no professional body has incorporated it into qualification frameworks, and no enterprise AI readiness programme systematically invests in building it.

The productivity gap between institutions that realise compounding value from agentic AI and those that do not will not be determined by which foundation model they license. It will be determined by whether their senior practitioners, the people whose tacit judgment is worth capturing, develop the capacity to externalise that judgment in forms agents can execute and regulators can audit. That capacity is not a technology problem. It is a professional development problem the industry has not yet decided to take seriously.

Sources

Academic

Dell'Acqua, F., McFowland, E., Mollick, E., et al., "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality," Harvard Business School Working Paper 24-013 https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321
Holtz, D., et al., MIT Sloan Management Review, "Study: Generative AI Results Depend on User Prompts as Much as Models," August 2025 https://mitsloan.mit.edu/ideas-made-to-matter/study-generative-ai-results-depend-user-prompts-much-models
Cao, B., et al., "On the Worst Prompt Performance of Large Language Models," arXiv 2406.10248, 2024 https://arxiv.org/html/2406.10248v1
Qiang, Y., et al., "Prompt-Perturbation Consistency Learning for Robust Language Models," Amazon Science, 2024 https://assets.amazon.science/16/e1/b790e6c647aea33749aa5bdf2d51/prompt-perturbation-consistency-learning-for-robust-language-models.pdf
Nonaka, I. and Takeuchi, H., SECI model review, Frontiers in Psychology, December 2019 https://pmc.ncbi.nlm.nih.gov/articles/PMC6914727/
Hemmecke, J., "Tacit Knowledge Externalisation," Warwick Business School OLKC Paper https://warwick.ac.uk/fac/soc/wbs/conf/olkc/archive/oklc5/papers/a-4_hemmecke.pdf
Liberti, J. M. and Petersen, M. A., "Information: Hard and Soft," Review of Corporate Finance Studies, 2019 https://ideas.repec.org/a/oup/rcorpf/v8y2019i1p1-41..html
Haynes, A. B., et al., "A Surgical Safety Checklist to Reduce Morbidity and Mortality in a Global Population," New England Journal of Medicine, 2009 https://www.nejm.org/doi/full/10.1056/NEJMsa0810119
Cullen, J., "The Knowledge Acquisition Bottleneck: Time for a Change," Expert Systems, 1988 https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1468-0394.1988.tb00065.x
Jarrahi, M. H. and Ritala, P., "Rethinking AI Agents: A Principal-Agent Perspective," California Management Review, July 2025 https://cmr.berkeley.edu/2025/07/rethinking-ai-agents-a-principal-agent-perspective/
"Principal-Agent Problems in Multi-Agent ML Systems," arXiv 2601.23211v1, January 2026 https://arxiv.org/html/2601.23211v1
Google DeepMind, "Contract-First Decomposition for Intelligent Delegation," arXiv 2602.11865, February 2026 https://arxiv.org/html/2602.11865v1
"Analysis of LLM Reasoning Benchmark Improvements Since GPT-4," arXiv 2505.19676, May 2025 https://arxiv.org/pdf/2505.19676
"AI Complementarity Across Job Vacancies," arXiv 2412.19754, December 2024 https://arxiv.org/abs/2412.19754
Clark, R. E., et al., "Cognitive Task Analysis," Applied Research in Workplace e-Learning, 2007 https://hpttreasures.wordpress.com/wp-content/uploads/2025/11/cta_chapter_2007-1.pdf
Scientific Reports, "Prompting Strategies vs. Human Experts in Diagnostic Medicine," October 2025 https://pmc.ncbi.nlm.nih.gov/articles/PMC12575769/
Barthélemy and Quélin, Offshoring Research Network, Scholarworks UMB https://scholarworks.umb.edu/cgi/viewcontent.cgi?article=1017&context=management_marketing_faculty_pubs
Stanford, "Explainer: AI Complementarity and LLMs in the Workplace," July 2025 https://news.stanford.edu/stories/2025/07/explainer-ai-complementarity-llms-workplace-jobs
Feigenbaum, E., Expert Systems: Principles and Practice https://imarcrobotics.com/wp-content/uploads/2019/04/Feigenbaum-EXPERT-SYSTEMS-PRINCIPLES-AND-PRACTICE.pdf
Stanford Expert Systems Archive https://stacks.stanford.edu/file/druid:vf069sz9374/vf069sz9374.pdf

Added in July 2026 Refresh

Freshfields, "EU AI Act Unpacked #34: The Final Digital Omnibus on AI — Key Amendments," June 2026 https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-34-the-final-digital-omnibus-on-ai-key-amendments-to-the-a-102nber
Reuters, "Agentic AI May Require Regulatory Reform, BoE's Breeden Says," June 30, 2026 https://www.reuters.com/world/agentic-ai-may-require-regulatory-reform-boes-breeden-says-2026-06-30/
Reuters, "Bank of England Sees Growing Risks to Financial Stability from AI," July 7, 2026 https://www.reuters.com/business/finance/bank-england-sees-growing-risks-financial-stability-ai-2026-07-07/
Polar Insight, "The Mills Review: The FCA Plants a Flag on Agentic AI," July 7, 2026 https://polarinsight.com/insights/the-mills-review-the-fca-plants-a-flag-on-agentic-ai-2026-07-07
Stanford HAI, "AI Index Report 2026 — Chapter 2: Technical Performance" (PDF) https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_2_technical.pdf
"Context Engineering: A Methodology for Structured Human-AI Collaboration," arXiv 2604.04258, April 2026 https://arxiv.org/html/2604.04258v1
"Agent Behavioral Contracts: Formal Specification and Runtime Enforcement for Reliable Autonomous AI Agents," arXiv 2602.22302, February 2026 https://arxiv.org/html/2602.22302v1
Signal, "Agentic AI: Demo to Deployment — What Broke," March 2026 https://www.readsignal.io/article/agentic-ai-demo-to-deployment-what-broke
Constantinople, "Much Ado About Nothing: Why AI Agents Keep Failing in Banking" https://www.cxnpl.com/resources/insights/much-ado-about-nothing-why-ai-agents-keep-failing-in-banking