--- title: date: 0001-01-01 canonical: https://mariothomas.com/blog/unpriced-cost-ai-oversight/briefing/ --- # Article Briefing Note The following article briefing note has been prepared to provide enough information to ghost author a first draft of the article. ## The Cost That Came Due: Why AI Oversight Is Exhausting Your People * **Working Title**: The Cost That Came Due: Why AI Oversight Is Exhausting Your People (67 characters) * **Alternative Title**: The Balancing Item: Who Absorbs the AI Costs Your Business Case Left Out (73 characters) * **seoTitle candidate**: The Unpriced Cost of AI Oversight (33 characters — verify with `echo -n | wc -c`) * **Publication Date**: Sunday 26 July 2026 * **Word Count**: 1,750–2,250 words * **Audience**: Board members, directors, senior executives, line-of-business leaders. UK English. * **Author**: Mario Thomas — Chartered Director and Head of Applied AI & Emerging Technology Strategy, with expertise in AI, cloud, board, and emerging technology governance. * **Category**: AI * **Tag candidates (from approved register)**: governance, board-governance, human-impact, future-of-work, ai-in-business ### Connection to Previous Work This article is the True Investment Profile's warning coming due. The AI business case foundation piece established governance overhead as a cost category that rarely appears in traditional technology business cases; the Appreciating Ledger repeated that the True Investment Profile systematically undercounts it. This article shows what happens when the undercounted cost goes unpaid: human cognitive capacity becomes the balancing item. It publishes one week after "Time Saved Is Not Value Created" (19 July) and one week before the EU AI Act Annex III piece (2 August), and should set up the latter's oversight obligations with a single forward reference. **Key cross-references (linked text, existing corpus):** - True Investment Profile (Building Block 3, `/blog/ai-business-case-foundation/`) — the spine of the argument: this cost was predicted - The Verification Premium (`/blog/vibe-coding-vs-classical-training/`) — verification is labour AI cannot supply - The Redeployment Dividend (`/blog/ai-redeployment-dividend/`) — where freed capacity should flow; Bainbridge's ironies of automation already introduced there - Time Saved Is Not Value Created (19 July slug) — the dividend leaking; this piece escalates: sometimes the dividend is negative - The Reasoning Gap (`/blog/the-reasoning-gap/`) — DUAA safeguards mandate human review of automated decisions - Six Board Concerns (`/blog/board-ai-governance-priorities/`) — sits across Financial and Operational Impact and Stakeholder Confidence - Minimum Lovable Governance (`/toolkit/minimum-lovable-governance/`) — the remedy register: governance designed to enable, not smother **Differentiation discipline (critical):** this article must not restate the True Investment Profile (the cost exists), the Verification Premium (verification is unpriced labour), or Time Saved (the dividend leaks). Its unique contribution is that the predicted cost is now measurably surfacing in people, and that human cognitive capacity is a finite, depletable control resource that Boards must govern like any other. ### Core Premise Every AI business case counts the hours saved; almost none counts the hours added. The reviewing, correcting, supervising, and deciding-whether-to-trust that AI outputs demand is real labour, increasingly mandated by regulation, and in most organisations absorbed silently by people on top of their existing roles. BCG Henderson Institute research published in Harvard Business Review in March 2026 (1,488 full-time US workers at large companies) gives the consequence a name — "AI brain fry" — and, crucially, isolates the mechanism: the strain attaches to oversight load, not to AI use itself. Workers using AI to replace routine tasks reported lower burnout. The burden is therefore a work-design choice, and work design is governable. Burnout is not the article's thesis; it is the first observable symptom that the True Investment Profile's undercounted governance overhead has gone unpaid. ### Article Outline #### I. Summary (~114 words) **Draft Summary**: Every AI business case counts the hours saved. Almost none counts the hours added: the reviewing, correcting, and supervising that AI outputs demand before anyone can rely on them. That oversight labour is real, regulation increasingly mandates it, and in most organisations it is absorbed silently by people on top of their existing roles. BCG research published in March 2026 shows the consequence: workers carrying heavy AI oversight loads report a distinct mental fatigue their colleagues do not, while those using AI only to replace routine work report less burnout, not more. In this article, I argue that burnout is a control cost Boards have failed to price, and that the remedy is governance, not resilience training. *(118 words — trim in editorial pass. No em dashes in summary.)* #### II. Opening: the missing line item (150–200 words) * Lead with the ledger asymmetry: hours saved are celebrated in every AI business case; the hours added by oversight appear in none. No opening question; no "fascinating"; straight to the governance gap. * Establish immediately that this cost was foreseeable — one sentence linking to the True Investment Profile, which named governance overhead as a systematically undercounted category. * Close the opening on the consequence: when a real cost is not budgeted, something absorbs it. The research now shows what. #### III. The cost you already predicted (300–350 words) * Traditional software economics: licence, implementation, maintenance. AI economics add recurring human costs that never disappear: supervision, verification, governance. Reference True Investment Profile as the established treatment; do not rebuild it. * The oversight labour is specific and daily: checking outputs before relying on them, correcting confident errors, maintaining accountability for decisions the human did not draft. The Verification Premium established this as labour; here the point is that it is *recurring operating* labour, not one-off implementation effort. * The silent absorption mechanism: no role redesign, no headcount, no budget line — the work lands on existing people on top of existing jobs. This is the balancing item. #### IV. The evidence arrives (400–450 words) * **BCG Henderson Institute / HBR, 5 March 2026** ("When Using AI Leads to 'Brain Fry'", Bedard, Kropp, Hsu, Karaman, Hawes, Rosen Kellerman; survey of **1,488** full-time US workers at large companies). Core findings to deploy: - **14% of AI-using workers** report the strain (mental fog, slower decision-making, difficulty focusing) - Definition: mental fatigue from excessive use, interaction with, or oversight of AI tools beyond one's cognitive capacity - The pivotal distinction: strain attaches to *oversight load*; workers using AI purely to replace routine tasks reported **lower** burnout, and AI users overall reported less burnout than non-users - Incidence higher in marketing, HR, operations, and software engineering than in legal and compliance — real, citable texture in place of invented examples - Downstream effects: more errors, decision overload, higher intent to leave — retention and quality risk, not just welfare * Handle the nuance honestly and prominently: this is not an "AI causes burnout" finding. The variable is how oversight work is designed and resourced. That precision is what keeps the piece out of the doomer register and inside the governance argument. * **Berkeley Haas / HBR, 9 February 2026** (Ranganathan and Ye, eight-month ethnographic study at a US technology company of roughly 200 employees). Frame explicitly as qualitative field evidence of mechanism, not population statistics. Three intensification patterns: task expansion, blurred work-life boundaries, increased multitasking as workers manage parallel AI threads — a self-reinforcing cycle in which acceleration raises expectations. * **Gallup, State of the Global Workplace**: leaders report higher engagement alongside higher daily stress than employees — the most engaged are the most exposed, and the judgement burden concentrates upward. *Exact figures must be verified against the Gallup primary before drafting; do not use secondary characterisations.* #### V. Why oversight exhausts (300–350 words) * Bainbridge's ironies of automation (1983, already in corpus via the Redeployment Dividend): monitoring is the task humans perform worst and find most fatiguing; automation converts doers into vigilant checkers. Forty years of human-factors research made this cost predictable — and predictable costs left unbudgeted are a governance failure, not a surprise. * What competent oversight actually demands, in plain prose (NOT as a named framework, NOT as a list): sustained critical thinking to interrogate confident output, analytical rigour to trace how a conclusion was reached, systems thinking to judge downstream consequences, creativity to spot what the machine did not consider, emotional intelligence and communication to carry the accountability conversations that follow. These capabilities are the entry conditions for working alongside AI — and each draws on the same finite cognitive reserve. * The compounding factor: this work arrives on top of the day job, at the pace of the machine rather than the person, with the accountability remaining fully human. #### VI. The regulatory ratchet (200–250 words) * Oversight is not discretionary. EU AI Act Article 14 mandates effective human oversight for high-risk systems; the Data (Use and Access) Act 2025 safeguards require human review of automated decisions (link to The Reasoning Gap for the established treatment). * The governance point: a control that is mandated but not resourced exists on paper only. Boards attesting to human oversight while the humans doing it are at cognitive capacity are attesting to a control that is quietly failing. * One forward sentence to the 2 August Annex III piece — obligations arriving, oversight load rising. #### VII. Governing the oversight load (350–400 words) * Reframe from resilience to resourcing. The wellbeing response (workshops, resilience training) treats the symptom; the governance response prices the cost. * Practical moves, in flowing prose: - Require every AI business case to state who performs oversight, what proportion of their capacity it consumes, and what it costs fully loaded — restoring the missing line to the True Investment Profile - Design oversight into roles rather than on top of them: explicit review responsibilities, adjusted workloads, recognised in objectives - Set span-of-control expectations for human–AI working: how many agents, tools, or output streams one person can competently supervise is a governance parameter, not an emergent accident - Distinguish deployment patterns: replacement of routine work (burnout-reducing per BCG) versus oversight-heavy augmentation (fatigue-generating when unresourced) — portfolio balance is a Board-visible choice - Watch the leading indicators: review backlogs, rubber-stamping rates, error escapes, attrition in oversight-heavy roles — connect to the Four Indicator Types if natural, without forcing it * Minimum Lovable Governance register: the answer is not more process; it is honest costing and deliberate design. #### VIII. Board questions and close (150–200 words) * Close with Board questions (consistent with house practice from the Bifurcation piece), for example: How many people now spend a material share of their week reviewing AI output, and where did the business case account for them? Who owns the oversight load for each deployed system? What is the fully loaded cost of the human supervision our AI estate requires? Which of our attestations of human oversight would survive an honest capacity audit? * Final register: AI value is real and incomplete until the control layer is priced. Not a case against AI; a case against incomplete accounting. ### Proposed Bolded Verdicts (two, in "X is Z, not Y" form) 1. **Burnout is a control failure, not a wellbeing issue.** (Place late, once earned — Section VI or VII, not the opening.) 2. **Oversight labour is a budget line, not a rounding error.** (Section III or VII.) Softer formulation for the body (not a verdict): burnout is the first observable symptom that organisations have failed to account for AI's supervision costs. ### Supporting Evidence to Incorporate - **14% of AI-using workers** report AI brain fry strain (BCG Henderson Institute / HBR, March 2026; n=1,488 full-time US workers at large companies) - Workers using AI only to replace routine tasks reported **lower burnout scores**; strain attaches to oversight load (BCG/HBR 2026) - Strain incidence higher in marketing, HR, operations, and software engineering than legal and compliance (BCG/HBR 2026) - Downstream effects: increased errors, decision overload, intent to leave (BCG/HBR 2026) - Three intensification mechanisms — task expansion, blurred boundaries, multitasking — from an eight-month single-firm ethnography (Ranganathan and Ye, Berkeley Haas / HBR, February 2026; qualitative, frame as field evidence) - Leaders report higher engagement but higher daily stress than employees (Gallup, State of the Global Workplace — **verify exact figures from Gallup primary before use**) - Bainbridge, "Ironies of Automation" (1983) — monitoring as the task humans perform worst (already cited in corpus) - EU AI Act Article 14 human oversight obligations; DUAA 2025 safeguards (via The Reasoning Gap) - Optional, already in corpus: Anthropic research on AI-assisted coding — engineers using AI scored 17% lower on mastery (supports the vigilance/capability point if needed; cap per single-source rule) ### Evidence Excluded (do not use) - **Microsoft Work Trend Index** in any form — excluded at Mario's direction - The "+33% decision fatigue" figure circulating in reviewer feedback — unverified against the BCG primary; drop unless confirmed there - Secondary-circulated percentages attributed to the Berkeley study (83%, 62%, 38%) — not traceable to the HBR primary - BCG mitigation figures from secondary coverage (15% lower fatigue with manager support; 28% with work-life culture) — verify against the HBR primary before use; drop if not present - Kirsten's observation — private motivation only; no anecdote, no reference - Any invented illustrative example of oversight work — use the BCG industry variation instead ### What to Avoid **Content:** - Naming the six skills as a framework, methodology, or "table stakes skills" — plain prose only; the phrase "table stakes" is reserved in the corpus for competitive parity - Restating the True Investment Profile, Verification Premium, or Time Saved arguments — link, do not rebuild - "AI causes burnout" as a claim — the evidence supports "unpriced oversight causes strain" - Wellbeing-article register: no self-care advice, no HR programme design - Fabricated statistics or examples; unverifiable figures are dropped, not softened - Vendor names in body prose; citations via linked text only - Comments construable as an official AWS viewpoint **Style:** - No bullet points in article body; no opening questions; no em dashes in summary, maximum two in body - No contractions in body prose; UK English; "while" not "whilst" - Forbidden words per style guide (including robust, leverage, landscape, transformative, journey, game-changing) - "Boards" capitalised as governance bodies; sentence-case section headers; Title Case headline 65–75 characters - Bold key statistics on first mention; seoTitle ≤45 characters; description ≤160 characters ### Writing Perspective Guide - Urgently practical without panic; never a cheerleader, never a doomer - The register is a director who has read the research and is impatient with incomplete accounting, not alarmed by technology - Analytical sympathy for the people carrying the load; the argument is made on their behalf through governance, not sentiment - Confident without guaranteeing; the BCG nuance (task replacement lowers burnout) is presented prominently as intellectual honesty, not buried as a concession ### Strategic Through-Line The True Investment Profile predicted an undercounted cost → oversight labour is that cost, recurring and mandated → unbudgeted, it is silently absorbed by people → BCG and Berkeley now show the measurable consequence → vigilance was always the hardest human task (Bainbridge) → regulation makes the labour non-optional → therefore Boards must price, design, and govern the oversight load → burnout recedes as the accounting becomes honest. ### References for Cited Sources - **BCG Henderson Institute / Harvard Business Review (2026)**: Bedard, J., Kropp, M., Hsu, M., Karaman, O., Hawes, J., Rosen Kellerman, G., "When Using AI Leads to 'Brain Fry'", 5 March 2026. Available at: https://hbr.org/2026/03/when-using-ai-leads-to-brain-fry - *Used for*: definition and incidence of AI brain fry (14%), oversight-versus-replacement distinction, industry variation, downstream effects. Article sits behind HBR's paywall; verify exact figures against the full text or BCG's companion page (https://www.bcg.com/news/5march2026-when-using-ai-leads-brain-fry) before publication. - **Ranganathan, A. and Ye, X. M. / Harvard Business Review (2026)**: "AI Doesn't Reduce Work — It Intensifies It", 9 February 2026. Available at: https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it - *Used for*: three intensification mechanisms from an eight-month ethnography at a US technology company (~200 employees). Qualitative single-firm study; frame accordingly. - **Gallup**: State of the Global Workplace. Available at: https://www.gallup.com/workplace/349484/state-of-the-global-workplace.aspx - *Used for*: leader engagement and stress differential. **Figures to be pulled and verified from the Gallup primary at drafting; edition year to be confirmed.** - **Bainbridge, L. (1983)**: "Ironies of Automation", Automatica 19(6). Overview: https://en.wikipedia.org/wiki/Ironies_of_Automation - *Used for*: monitoring as the human task most prone to failure and fatigue; continuity with the Redeployment Dividend citation. - **EU AI Act, Article 14** (human oversight): https://artificialintelligenceact.eu/article/14/ - *Used for*: mandated oversight obligations for high-risk systems; forward link to the 2 August Annex III article. - **Data (Use and Access) Act 2025** — via existing corpus treatment in The Reasoning Gap; no new citation work required. ## Appendix I: Draft Summary Variants > Every AI business case counts the hours saved. Almost none counts the hours added: the reviewing, correcting, and supervising that AI outputs demand before anyone can rely on them. That oversight labour is real, regulation increasingly mandates it, and in most organisations it is absorbed silently by people on top of their existing roles. BCG research published in March 2026 shows the consequence: workers carrying heavy AI oversight loads report a distinct mental fatigue their colleagues do not, while those using AI only to replace routine work report less burnout, not more. In this article, I argue that burnout is a control cost Boards have failed to price, and that the remedy is governance, not resilience training. > The True Investment Profile warned that AI business cases systematically undercount governance overhead. New research shows where the unpaid cost has gone: into people. BCG's March 2026 study of 1,488 workers found a distinct mental fatigue attaching specifically to AI oversight load, while workers using AI only to replace routine tasks reported less burnout, not more. The burden, in other words, is a work-design choice. In this article, I argue that the human oversight AI demands is a recurring control cost that Boards must price, resource, and govern, and that burnout is simply the first observable symptom of a control that exists on paper but is failing in practice.