The AI Rework Costs Nobody Owns

Imagine a quarterly operating review. Sales announces that AI has shortened proposal preparation. Legal reports a growing approval queue. Finance is investigating more disputed commitments. All three reports sound credible. But was the company more productive? The effort required to make AI-generated work acceptable has slipped across departmental boundaries.

That disconnect matters more than another adoption target. In Gallup’s October 2026 survey, 63% of US employees who used AI at work said it helped them work faster. Meanwhile, McKinsey’s 2026 global survey found that 80% of respondents reported individual productivity gains, while 37% reported some positive contribution to organizational earnings. These are self-reported figures with several possible explanations. One deserves attention: costs nobody counts.

Why Good Departmental Metrics Can Mislead

The first mechanism is incentives. Managers receive praise for rollout speed, AI usage, and time saved on initial drafts. The employees who must check those drafts have a different performance target: avoid an error reaching a customer. A sales team might improve its drafting metric while pushing an unmeasured burden onto legal or customer service. Each department optimizes the part it sees.

The Workday-commissioned research released in January estimated that nearly 40% of reported AI time savings were lost to correcting, rewriting, or verifying output. It is a survey finding, not a universal benchmark. It illustrates why a speed incentive can reward activity that shifts costs rather than eliminating them.

The second mechanism is fragmented ownership. IT budgets for software and uptime. Operations measures throughput. Finance evaluates labor costs. Legal and compliance carry the responsibility for inaccurate output. When a problem crosses those boundaries, the expense often lands outside the team credited with the supposed gain. Research on AI-generated “workslop” and its burden on recipients describes one way apparently polished output can transfer effort to colleagues.

The third mechanism is timing. A proposal takes minutes to draft; its incorrect service commitment may take weeks to surface. The pilot has closed, and the correction appears as an exception or dispute in another reporting period. MIT Sloan’s analysis of the verification bottleneck explains why producing answers faster does not remove the economic cost of deciding whether they are trustworthy.

Employees may hesitate to surface this mismatch when leaders celebrate AI usage. Gallup also found that 52% of employees felt they had less influence over technology adoption than they wanted. A dashboard that counts suggestions accepted but ignores errors caught gives managers an incentive to overlook the people doing careful checking.

Put One Executive in Charge of the Finished Outcome

For every material AI-assisted workflow, assign a business executive, often the COO or division head, responsibility for the full path from request to accepted result. IT owns technical performance; it should not carry sole accountability for whether a customer receives an accurate answer or a contract clears approval. The accountable owner needs Finance, quality reviewers, and frontline employees at the same table—and authority to change the cross-departmental processes and incentives that create the burden.

Define success as an accepted business outcome: a resolved inquiry that stays resolved, a contract approved without material omissions, or an invoice reconciled correctly. Then track the labor and defects across every handoff. Include preparation, drafting, verification, corrections, escalation, reopened work, and software expense. Report which team performed the work and which team’s dashboard recorded it. It makes transfers visible without assuming every verification minute is waste.

A short diagnostic trial still matters, but its role should be narrow. Compare a baseline set of cases with comparable AI-assisted cases over two weeks, using a common acceptance standard. Record downstream corrections after the trial where feasible; two weeks alone cannot establish durable return. The NIST generative AI risk profile provides a useful framework for evaluating risks within real workflows.

Measurement should also allow genuine gains to survive scrutiny. A study of 5,172 customer-support agents found a 15% average rise in issues resolved per hour with AI assistance, with different effects across workers. The executive question is which employees, tasks, and quality standards yield reliable results, and whether those results hold after handoffs.

Make Accountability Part of the Investment Gate

Before approving expansion, the CEO and CFO should require four answers. Who signs off on the finished output? Which teams bear verification and correction costs? What quality or exception threshold triggers redesign? And what operational outcome improves after all costs are counted? Put the answers beside license spending and usage figures, with a named owner and a review date.

Make the scorecard small enough to use: total labor and cost per accepted outcome, serious errors and reopened cases, end-to-end turnaround, and the actual use of recovered capacity. Where verification depends on scarce specialists, track the capacity it consumes, not just the minutes. Reward people for catching defects early. Do not give a rollout team full credit for faster drafts while a different department absorbs the cleanup.

An experiment involving 7,137 workers at 66 companies found that users spent about two fewer hours weekly on email in the latter half of the trial, without detectable changes in the quantity or mix of their tasks. Turning those savings into organizational value required management choices. St. Louis Fed research on AI use by detailed work task likewise shows why broad adoption figures reveal little about the economics of particular workflows.

At the next investment meeting, bring Sales, Operations, Finance, and the people who verify AI-generated work into one discussion. If a department claims a gain, ask who carried the work after its output left the department. Until someone owns that answer, the company has an AI activity report rather than a credible business case.