Two-Thirds of Buyers Now Run AI Agents in Production. Almost Half Say Human Oversight Still Eats a Quarter of the Gain

Two surveys published weeks apart in the autumn of 2026 describe the same gap from opposite ends. Buyers say agents have reached production; governance data says they have not reached autonomy. 49% of surveyed leaders now define high-risk use cases in which agents are barred from deciding on their own, and the share who have built controls directly into their agents fell from 43% two quarters ago to 30%. Oversight is tightening while capability is being marketed — and that is not a paradox to be argued away, it is the pricing mechanism.

The 45% figure is the one that breaks the standard agent business case. If supervision consumes a quarter of the gain, the residual saving is roughly half of what the vendor's model promised — before the wages of the people doing the supervising. And 52% of buyers require human quality review before an agent is allowed into production at all, which turns the human layer from an operating cost into a gate on the entire deployment.

The cost of supervising agents does not appear on enterprise balance sheets as a separate line. It appears as services revenue. Accenture booked $12.8 billion of Managed Services orders in its fiscal fourth quarter — a record, at a book-to-bill of 1.4 — and managed services revenue grew 7% in local currency. Consultants are paid to build the stack; operators are paid to run it while the agents are still unreliable.

The second line is safety and evaluation, and it is the most explicit pricing of human judgment anywhere in the market. On September 18, 2026, Accenture and Anthropic agreed to establish a team of embedded evaluators inside the model lab, with each side committing at least $1 billion over five years. Those evaluators red-team models, run alignment assessments and test safeguards — a standing team of professionals whose job is to find what the models get wrong before customers do.

The volume evidence is specific. Accenture added nearly 100 clients to its advanced AI book in a single quarter and ended fiscal 2026 with more than 400, while its eight emerging AI and data partners tripled bookings and more than doubled revenue versus fiscal 2025. More than 60% of firm revenue came through its top ten ecosystem partners, growing 6% against5% for the company overall. What a client books as supervision cost, the vendor books as contract value that scales with agent activity.

Everything above assumes the vendor can charge for the human layer. It cannot, yet. Clients have begun treating agent savings as a contractual discount rather than a shared gain: the Persistent Systems chief executive says customers now demand the same work for 25% to 30% less, and TCS has roughly 80% of its finance, HR and business services contracts priced on outcome measures, double the share when AI went mainstream in late 2023. When the contract pays on outcomes, the client keeps the savings and the vendor keeps the work.

The repricing is visible in the tape. India's Nifty IT index fell by a fifth over the past year, with ten constituents shedding a combined $73 billion of market value, while TCS, Infosys, Wipro and HCLTech grew revenue at 1% to 3%. The human layer was not a growth offset for anyone.

Concentrix is the clearest case of the squeeze. Revenue fell 1.2% as reported and 0.5% in constant currency in the quarter ended August 31, 2026, and the company then took a $1.05 billion non-cash goodwill impairment against a market value that no longer supports the old growth story, guiding Q4 revenue down 3% to 5% in constant currency. Yet management says roughly half of revenue now comes from business won or deployed in the three years since AI arrived. Winning the AI work has not replaced the legacy base, and the vendor still carries the human cost of serving it.

Here is the chain most agent models miss. First, more agents in production means more exception volume — every workflow an agent touches produces cases that a person must review, correct or escalate, and that volume grows with activity rather than with headcount. Second, that exception work is increasingly priced as an outcome rather than as hours, so vendor revenue tracks the number of outcomes handled rather than the number of people deployed. Third, the price per outcome is falling faster than the volume is rising, because the same clients demanding automation are demanding 25% to 30% concessions on identical scope.

The consequence is not obvious. A services firm's revenue now grows with the number of agent-driven outcomes it touches rather than with the labor hours it eliminates, so agent adoption is a volume tailwind and a price headwind at the same time. The firms that grow through this phase convert exception volume into priced outcomes while holding margin: Cognizant added10,100 people year on year to reach 359,900 and still posted an adjusted operating margin of 16.3%, a four-year high, while Genpact grew Advanced Technology Solutions revenue 24.1% to $363 million and lifted gross margin 60 basis points to 36.5%. Inside Genpact, the contrast is stark: the AI-facing segment grew 24.1% while legacy Core Business Services grew 1.9%. Agent adoption is a volume tailwind and a price headwind at the same time.

The clearest physical illustration of the human layer is autonomous trucking, where the human is not a worker but a safety operator and a remote assist team. Aurora Innovation said at its September 23, 2026 investor day that it is fully allocated to exit 2026 with 200 driverless trucks in operation and is targeting more than 30,000 by 2030 under a driver-as-a-service model. In its fiscal Q3 2026 results on October 28, it reported $3 million of revenue, a $234 million net loss, 20 driverless trucks built and heading into the fleet, and a production rate heading to 20 trucks a week in early December. It lost more than $800 million in 2025.

The revenue frame is equally conservative. Aurora has described the 200-truck deployment as a potential multi-year revenue stream in the hundreds of millions, and its trucks average more than 225,000 miles a year on a driverless basis, more than double the utilization of a conventional truck, having completed over 500,000 driverless miles since commercial launch. But autonomy here is a partial claim: as of September 2026, most of Aurora's driverless runs still carried a human safety operator behind the wheel in case the technology failed, and remote assist and on-site support are the operating cost that scales with the fleet.

Order matters. Volume news arrives first — book-to-bill, bookings, truck counts — because it is published sooner and is easier to spin. Price news arrives later, in the guide, and it decides whether the human layer is a business or a rounding error. The guide is where the human layer either shows up as revenue or quietly disappears, which is why the next two quarters of guidance matter more than the next two quarters of headlines.

The second-order point is that oversight is becoming a compliance cost rather than a productivity tax, and compliance costs do not disappear when models improve. In KPMG's Q3 2026 survey, 74% of organizations had added cost reviews to AI approval processes, up from 61% a quarter earlier, and 43% had put usage or token budgets in place. In banking, the share of leaders refusing to let agents touch sensitive data without human oversight rose from 51% to 75% in a single quarter. The direction of governance travel makes the human layer more durable than the agent-efficiency bulls assume.

The bear case for the whole thesis deserves equal weight. If agent reliability crosses the point where oversight drops below roughly a tenth of activity, the exception-handling revenue pool collapses faster than agent volume expands, and the services industry becomes the casualty rather than the beneficiary. Current data does not support that: oversight is still measured in hours per week per team, 52% of buyers still require human review before production, and the largest vendors were still reporting record managed services bookings in the same quarter their forward guides came down.

Fact: agents are in limited or scaled production at two-thirds of buyers, oversight consumes more than a quarter of expected gains at 45% of them, and the vendors selling the supervision layer posted record or near-record managed services bookings in the same quarter that their forward growth guides came down. Inference: the human layer is not transitional overhead waiting to be automated away — it is the price of trust, and trust is now bought as a service with a contract, a book-to-bill ratio and a safety lab. Speculation worth naming: the two-sided repricing resolves one way or the other. Either vendors capture the exception volume through outcome-priced contracts, or clients keep the savings and the entire services revenue pool deflates along with the models.

That produces a clean way to sort the sector. Buy the contract holders who staff the oversight layer while expanding margin — Genpact and Cognizant did both in the September quarter — and stay cautious on the ones whose revenue requires the client to keep believing the old math, where the guidance cut has already started. In industrials, the same logic says autonomy scales in the hundreds before it scales in the thousands, and every unit in the interim still carries a human being, either behind the wheel or on the remote assist line.