Vijayanta Gupta

Essay

The supervision cost of AI adoption

. 6 min read

The first growth curve of workplace software was underwritten on labour substitution. A tool was worth buying when it removed work, and an AI copilot is the purest version of that pitch: hold headcount flat, add the assistant, collect the productivity. The 2026 budget cycle priced that promise into plans across the software economy. The break is that the work does not disappear. It converts. AI adoption turns doing into reviewing, and the review load lands on the same people, unbudgeted, unmeasured, and absent from every ROI deck. The decision this piece is about is whether supervision becomes a managed cost line or stays an invisible tax. Companies that get it wrong will report AI productivity in the board pack while their margin and their best people quietly absorb the difference.

The visible symptom

BCG’s study of 1,488 full-time US workers, published in March 2026, gave the symptom a name: AI brain fry, the mental fatigue of overseeing AI beyond cognitive capacity. The headline finding is not which function uses AI most, but which juggles the most tools: fatigue concentrates where workers run four or more AI tools at once, and BCG links it to increased errors, decision overload, and intent to quit (BCG, 5 March 2026). One self-reported, US-only survey is an exhibit, not a thesis. It earns its place here because two independent evidence layers now say the same thing in operating and financial terms.

The operating signal: output per head is not what it claims

The operating data shows the conversion of work directly. In the Upwork Research Institute’s survey of 2,500 workers, 96 percent of C-suite leaders expected AI to lift productivity, while 77 percent of employees using AI said the tools had added to their workload, with the largest single driver being time spent reviewing or moderating AI-generated content, at 39 percent (Upwork, 23 July 2024). That gap between the buyer’s expectation and the operator’s experience is the supervision cost, observed at survey scale.

The sharper evidence is experimental. METR’s randomized controlled trial of experienced open-source developers found that allowing state-of-the-art AI tools made them 19 percent slower to complete tasks. The detail that matters for a board is the perception gap: developers forecast AI would speed them up by 24 percent, and after experiencing the slowdown still believed it had sped them up by 20 percent (METR, 10 July 2025).

Supervision cost is invisible to the people paying it, which is why it never reaches the P&L review.

If the people doing the work cannot feel the cost, self-reported productivity dashboards will not surface it either. The instrumentation has to be designed for it.

The financial signal: the margin illusion, inside the org

The token side of AI cost is now visible in public numbers. ICONIQ’s 2026 tracking puts average AI product gross margin at 52 percent, against the 80 percent that defined the SaaS decade, with roughly 230,000 dollars of every million in AI revenue consumed by inference before any payroll (ICONIQ via SaaS Mag, 15 May 2026). Companies have learned to see that line. The supervision line is the same phenomenon on the labour side, and almost nobody books it. Review labour sits inside salaried time, so it shows up nowhere except in slipped cycle times and burnout metrics.

MIT’s analysis of enterprise AI deployments found 95 percent of generative AI pilots delivering no measurable P&L impact, attributing the failures to integration, workflow, and governance gaps rather than model capability (Fortune, 18 August 2025). Read alongside the operating data, a plain interpretation is available: a pilot whose output requires uncosted human verification has not removed labour, it has restructured it, and the P&L correctly registers nothing. This is the AI margin illusion observed inside the organisation rather than in the COGS line: activity that reads as productivity while carrying a hidden cost that erodes the gain it appears to create.

The counterargument: better models shrink the review load

The objection writes itself: supervision is a transition cost, and as models improve, review burden falls. Partly true, and worth conceding precisely. Error rates on a fixed task do fall. But deployment does not hold the task fixed. As reliability improves, organisations push AI into higher-stakes work, where the cost of an unreviewed error is larger and the review threshold correspondingly higher. Supervision load is a function of deployment surface times stakes, not of model quality alone. The conditions under which the objection wins are stateable: if review-to-output ratios fall measurably within a fixed workflow over four quarters, supervision is transitional there and should be funded as such. Where the surface keeps expanding, the load compounds instead, the METR perception gap hiding it as it grows.

What the operator does

Take the constrained scale-up: 50 million ARR, two hundred enterprise customers, a feature-roadmap-first culture, copilots now embedded across support, marketing, and engineering. The friction is predictable. No function will volunteer that its AI rollout created work, because the rollout was justified on removing it. Budget owners defend the productivity narrative they sold. So the correction has to be instrumented, not asked for.

This quarter, on existing authority: the CFO adds supervision labour to cost-to-serve instrumentation, measured as review-to-output ratio per function, starting with the two functions carrying the most tools. The COO sets a tool-count ceiling per role until the measured review load justifies the sprawl, since the BCG data points at tool count, not usage depth, as the fatigue driver. Each function names a single owner of the oversight load, with the explicit question attached: what did this team stop doing to make room for reviewing? Next cycle: comp and capacity plans incorporate measured review load at the annual reset. Board mandate: none required. This is instrumentation, the cheapest reversible move available, and the first decision that is hard and political is publishing the first review-to-output numbers internally, because they will contradict the story the AI budget was approved on. Start there.

Sources

Back to writing