---
title: The Balancing Item: The AI Oversight Cost Your Business Case Never Priced
date: 2026-07-26
description: AI business cases count the hours saved but not the hours added by oversight. That unpriced labour is surfacing as fatigue, and it is a Board control cost.
author: map[email:mario@mariothomas.com name:Mario Thomas]
canonical: https://mariothomas.com/blog/unpriced-cost-ai-oversight/
---

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 people absorb it silently on top of existing roles. BCG research published in March 2026 shows the consequence: a distinct mental fatigue attaching to heavy AI oversight loads, while workers using AI only to replace routine work report less burnout, not more. I argue this fatigue is a control cost Boards must price, and the remedy is governance, not resilience training.
<!--more-->
{{< image3 src="unpriced-cost-ai-oversight" type="photo" alt="A modern open-plan office swallowed by dense fog, where a lone figure sits upright at a desk in the right of the frame, still working, his screens of charts and messages crisp at arm's length while the desks, monitors, and lamps of colleagues dissolve into white haze behind him, a visual rendering of the mental fog the research describes, where the person supervising the machine's output keeps working as clarity recedes, and the cost no business case recorded is paid in human attention (Image generated by ChatGPT 5.6)" width="735" height="413">}}

{{< audio2 src="mp3/unpriced-cost-ai-oversight.mp3" >}}

Every AI business case celebrates the hours saved. Automation of drafting, summarising, analysis, and correspondence arrives at the Board with a figure attached: hours recovered, multiplied by loaded cost, presented as return. The other side of the ledger almost never arrives with it. The hours added by AI, the reviewing, correcting, supervising, and deciding-whether-to-trust that its outputs demand before anyone can rely on them, appear in no business case I have seen.

This cost was foreseeable. When I set out the [True Investment Profile](/blog/ai-business-case-foundation/), I named governance overhead as a cost category that AI business cases systematically undercount. What was then a warning about accounting has become a finding about people. When a real cost is not budgeted, it does not disappear; something absorbs it. Research published this year shows what that something is: the finite cognitive capacity of the people asked to supervise the machines. The balancing item on the AI ledger is human attention, and it is being spent without being counted.

## The cost you already predicted

Traditional software economics are familiar to any Board that has approved a technology investment: licence, implementation, maintenance, support. Consumption-based pricing complicated the cost curve, but the categories themselves stayed on the ledger, priced and visible. AI economics differ in kind. Alongside the subscription and the integration sits a category of recurring human cost that never disappears: supervision of outputs, verification of claims, and governance of the decisions those outputs inform. The True Investment Profile treats this in full, and I will not rebuild the argument here. What matters for this article is the shape of the cost, not its existence.

The oversight labour is specific and daily. Someone checks the output before relying on it, corrects the errors it states with confidence, and stays accountable for decisions they did not draft, which means understanding those decisions well enough to defend them. I have described this validation work as the [Verification Premium](/blog/vibe-coding-vs-classical-training/); the point here is that it is a recurring operating cost, not a one-off implementation effort. It grows with usage, it rises with the stakes, and for some categories of decision the law already requires it.

Yet in most organisations this work has no line of its own. There is no role redesign, no headcount adjustment, no budget entry. The work lands on existing people, on top of existing jobs. Any accountant would recognise the pattern: when a ledger must balance and a cost has been omitted, the difference is absorbed somewhere unrecorded. **Oversight labour is a budget line, not a rounding error.** Treating it as the latter does not make the cost smaller; it makes the cost invisible, and invisible costs surface as symptoms rather than variances.

## The evidence arrives

The symptom now has a name. Research by the BCG Henderson Institute, [published in Harvard Business Review in March 2026](https://hbr.org/2026/03/when-using-ai-leads-to-brain-fry), surveyed **1,488** full-time US workers at large companies and identified what the authors call AI brain fry: mental fatigue from excessive use of, interaction with, or oversight of AI tools beyond a person's cognitive capacity. **14%** of workers using AI reported the strain, describing mental fog, slower decision-making, and difficulty focusing.

The finding that matters for governance is not the incidence but the distribution. The strain attaches to oversight load. Workers who used AI mainly to replace routine or repetitive tasks reported burnout scores **15%** lower than those who did not use AI that way, and AI use overall did not predict increased burnout. This is a cross-sectional survey, so it establishes association rather than causation, and the honest reading is comparative throughout. But the pattern points somewhere specific: not at AI itself, and not at the people, but at how the work of supervising AI has been designed and resourced. That matters for governance because work design is a choice, and choices can be governed.

The functional variation matters. Strain incidence ranged from **26%** in marketing to **6%** in legal and compliance, with HR, operations, and engineering towards the top of the range. I read that contrast as telling: legal and compliance are functions in which review is already priced into the role, while the harder-hit functions have had oversight added to jobs designed for production. The downstream effects complete the governance picture: workers experiencing the strain reported more frequent errors, **33%** more decision fatigue, and higher intention to leave. The exposure is quality and retention, not welfare alone.

Qualitative evidence points at mechanism. An eight-month ethnographic study by Ranganathan and Ye at a US technology company of roughly 200 employees, [published in Harvard Business Review in February 2026](https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it), observed three intensification patterns: tasks expanding as AI made more work feasible, work-life boundaries blurring, and multitasking increasing as people managed parallel AI threads, in a self-reinforcing cycle where acceleration raised expectations. Interview research by Shin and Sucher, [published in Harvard Business Review in June 2026](https://hbr.org/2026/06/ai-adoption-is-overloading-your-middle-managers), locates where that absorption concentrates: across 18 interviews at two major consulting firms, the oversight, coaching, and quality-control demands of AI had been layered onto middle managers' existing work rather than designed into their roles, while the layers above and below them experienced the gains. Both are field studies rather than population statistics, but they describe how the balancing item is absorbed in practice, and by whom.

The burden also concentrates upward. According to research by Gallup in its State of the Global Workplace 2026 report, leaders are the most engaged cohort at **26%**, yet also the most likely to report a lot of daily stress, at **46%**, **seven percentage points** above individual contributors. Gallup does not attribute this difference to AI. What it shows is that leadership responsibility already carries a substantial judgement burden, and adding unpriced AI oversight to that burden should be treated as an exposure, not as free capacity.

## Why oversight exhausts

None of this should surprise anyone who has read the human-factors literature. In 1983, Lisanne Bainbridge published [Ironies of Automation](https://en.wikipedia.org/wiki/Ironies_of_Automation), observing that automation does not remove the human from the system; it changes the human's task from doing to monitoring, and monitoring is the task humans perform worst and find most fatiguing. I introduced Bainbridge's argument in [The Redeployment Dividend](/blog/ai-redeployment-dividend/), and it bears directly here. Forty years of research made the cost of sustained vigilance predictable, and a predictable cost left unbudgeted is a governance failure, not a surprise.

It is worth being concrete about what competent oversight demands, because the demand explains the fatigue. Interrogating a confident output requires sustained *critical thinking*, since the output carries no signal of its own reliability. Tracing how a conclusion was reached requires *analytical rigour*, particularly when the system cannot show its working. Judging what a flawed output would do downstream requires *systems thinking*. Spotting what the machine did not consider requires *creativity*, because the omission is by definition absent from the page. And when review finds a problem, the accountability conversations that follow require *emotional intelligence* and *clear communication*. These capabilities are the entry conditions for working alongside AI, and each draws on the same finite cognitive reserve. None of them can be delegated to the tool being supervised.

What makes the load heavier still is where it sits. This work arrives on top of the day job rather than instead of it, at the pace of the machine rather than the pace of the person, and with the accountability remaining fully human throughout. A reviewer processing a stream of machine-speed output does not experience the productivity gain the business case recorded. They experience the queue.

## The regulatory ratchet

Boards cannot treat this labour as discretionary goodwill, because for defined decisions the law already reaches it. In the UK, the [Data (Use and Access) Act 2025](https://www.legislation.gov.uk/ukpga/2025/18/contents) requires safeguards for decisions based solely on automated processing that produce legal or similarly significant effects for individuals; those safeguards include the person's right to contest the decision and to require human intervention, a boundary I examined in [The Reasoning Gap](/blog/the-reasoning-gap/). A right exercisable on demand requires standing capacity to answer it, which is oversight labour by another name. In the EU, [Article 14 of the AI Act](https://artificialintelligenceact.eu/article/14/) establishes human oversight requirements for high-risk systems as classified under the Act, and while the July 2026 amendment deferred those obligations to December 2027, the deferral moved the deadline, not the direction.

Regulation is, in any case, the smaller part of the point. Most organisations already assert human oversight without any statute compelling them: in AI policies, in assurances given to customers and regulators, and in the risk descriptions attached to Board approvals. A control that is asserted but not resourced exists on paper only. A Board attesting to effective human review while the humans providing it are at cognitive capacity is attesting to a control that is quietly failing, with the failure invisible until an error escapes or someone asks how review actually happens. **Oversight fatigue is a control failure, not a wellbeing issue.**

## Governing the oversight load

The instinctive organisational response is a wellbeing response: resilience training, workshops, encouragement to manage workload. That treats the symptom. The governance response prices the cost, and it begins with the business case. Every AI investment proposal should state who performs the oversight it assumes, what proportion of their capacity it will consume, and what that costs fully loaded. This is not new bureaucracy; it is the restoration of a missing line to the True Investment Profile.

From honest costing, design follows. Oversight belongs in roles rather than on top of them: explicit review responsibilities, workloads adjusted to accommodate them, and the work recognised in objectives and appraisal rather than absorbed as invisible goodwill. Executives should also set span-of-control expectations for human and AI working. How many agents, tools, or output streams one person can competently supervise is a governance parameter to be chosen deliberately, not an accident that emerges from adoption.

Capacity should then be tested in operation, not accepted as a planning assumption. Management should be able to show the expected volume of AI output, the review time allowed for each item, the escalation route when confidence is low, and the point at which a growing queue triggers additional capacity or reduced automation. Without those parameters, human oversight describes an aspiration rather than an operating control.

Portfolio balance is a Board-visible choice for the same reason. The BCG research distinguishes deployment patterns with opposite human consequences: using AI to replace routine work was associated with lower burnout, while oversight-heavy augmentation generates fatigue when it is unresourced. A Board reviewing its AI portfolio can ask which pattern dominates and whether the oversight-heavy portion has been resourced honestly, in the same way it would examine any other concentration of operational risk.

Finally, the load can be watched. Review backlogs, rubber-stamping rates, error escapes, and attrition in oversight-heavy roles are leading indicators of a control under strain, in the sense I set out in the [Four Indicator Types](/blog/maximum-fidelity-four-indicators/): signals that precede the failure rather than record it. None of this requires a new committee. The register is [Minimum Lovable Governance](/toolkit/minimum-lovable-governance/): the answer is not more process but honest costing and deliberate design, applied where the exposure is material.

## Questions for the Board

Three questions establish whether the balancing item has been priced. *Who is performing the oversight assumed by each AI business case?* *What capacity has been removed from their existing workload to make that oversight credible?* *What evidence shows that human review is substantive rather than delayed, rushed, or routinely rubber-stamped?* A fourth question sharpens the stakes for any organisation making public or regulatory assertions of human oversight: *which of those attestations would survive an honest capacity audit?*

The value AI creates is real, and it remains incomplete until the control layer that makes it trustworthy is priced. In my view the research arriving this year is not a case against AI. It is a case against incomplete accounting, and the account is one that Boards, not the people silently absorbing the cost, are responsible for settling.

{{< campaign "unpriced-cost-ai-oversight" "hello@mariothomas.com" "Hello" "Let's Continue the Conversation" "Thank you for reading about the oversight cost that AI business cases leave unpriced. I'd welcome hearing about your Board's experience with the balancing item - whether you're finding ways to make oversight labour visible in investment proposals, redesigning roles so that review work is resourced rather than absorbed, or discovering that attestations of human oversight look less certain under an honest capacity audit. I'd also value hearing where the load is landing in your organisation, and what leading indicators are telling you about whether the control layer is holding." "Thank you for submitting your details, here's what you provided:" "Click send to share your input with me." >}}