The Supervision Tax: What Your AI Agents Really CostThe Supervision Tax: What Your AI Agents Really Cost

A RevOps lead walked me through an agent his team had shipped in the spring. It reads inbound leads, enriches them, scores them against the ICP, and routes them to the right rep with a summary note attached. It runs on every form fill, all day, without complaint. He was proud of it, and he should have been. It works.

Then, almost as an aside, he mentioned the document.

Two people on his team keep a running Google Doc of every time the agent gets it wrong. Wrong industry, wrong account owner, a summary that misses the part of the form that actually mattered. One of them spends the first twenty minutes of every day scanning the overnight routing queue before the reps get to it. The other does a longer pass on Fridays. They have been doing this since April, and the document is now forty pages long.

I asked him what the agent saved the team each week. He knew that number cold: about fifteen hours. Then I asked what the checking cost. He went quiet, did the math on the spot, and came back with something close to nine.

Nobody had ever put those two numbers on the same page.

That gap has a name, and most companies are paying it without ever writing it down. Call it the supervision tax: the recurring human cost of watching, verifying, and correcting the AI you deployed to save human cost. It is not a rounding error, and right now it is remarkably widespread. In a Harris Poll survey of 300 data, privacy, and AI decision-makers conducted for Collibra, 87% said their teams regularly check that the context available for AI agents remains accurate and current, and more than half said employees spend hours manually reviewing and correcting agent outputs. In the same survey, 72% said that when AI initiatives fall short, the root cause is a poor data foundation.

The Tax Is Real, and It Is Not a Transition Cost

The comfortable read on all this is that supervision is temporary. New system, new process, of course you watch it closely for a while. Give it two quarters and the checking fades.

I do not think that is what is happening, and the shape of the adoption curve is the reason. Gartner's 2026 CIO and Technology Executive Survey found that only 17% of organizations have deployed AI agents so far, while more than 60% expect to within two years, the most aggressive adoption intent of any emerging technology it measured. Most of the supervision bill has not even been issued yet. The companies paying it today are the early ones, and the tax is not shrinking for them. It is holding steady, because the thing that generates it has not changed.

What generates it is simple. A human reading a dashboard notices when a number looks wrong and goes to find out why. An agent does not notice, and it does not go find out. It acts. Collibra's CEO Felix Van de Maele made the point precisely in that CIO Dive piece: when a definition is ambiguous or context is missing, an agent can still produce an answer and act on it with confidence. So somebody has to stand behind it, and that somebody is on your payroll.

Supervision is the interest payment on context your machines cannot reach. It recurs because the debt is still outstanding.

You Are Paying for the Part of Your Business the Agent Cannot See

The size of that debt is now measurable, and it is larger than most leaders assume.

Google Cloud and MIT Technology Review Insights surveyed 300 data and technology executives this summer and found that organizations grant their AI systems access to an average of just 45% of their enterprise data. The organizations the report calls data leaders expose more than 70%. The laggards expose 30% or less. 55% said legacy data systems are actively preventing them from scaling agentic AI, and only 51% trust the accuracy and relevance of their agents' decisions. Among the laggards, that trust figure falls to 22%.

Read that as an operating statement rather than an IT statistic. If your agent can see 45% of what your business knows, then roughly half of every judgment it makes is being made without the relevant material, and a person has to supply the other half after the fact. That is the twenty minutes every morning. That is the forty-page document.

And notice which half is missing. It is almost never the clean, structured, well-governed half. It is the exceptions, the account histories, the reasons behind past decisions, the informal rules everyone in the building knows and nobody has written down. That is exactly the material I argued was the whole game in Context Is the Whole Game, and it is the material an agent most needs and least often has.

The Corrections Are the Asset, and Most Teams Throw Them Away

Here is the part that turns a cost into a trap.

Every entry in that forty-page document is a true statement about the business that existed nowhere else. The agent routed a lead to the wrong owner because it did not know about a territory exception agreed to in a call last year. Somebody fixed it. That fix is worth more than the routing, because it is a piece of the operating model finally getting written down. But it got written down in a Google Doc that no system reads, so the same correction will be made again next month, and the month after. The team is paying full price for the lesson every time and filing the receipt in a drawer.

This is why time saved is the wrong headline number, as I argued in Your Freed Hours Are Not Money. Fifteen hours saved against nine spent checking is a real six-hour gain, and a fine result. But a six-hour gain that never improves is a plateau, and it is being reported internally as a transformation.

Put the Tax on the Page This Quarter

You can measure this in a week, and I would do it before approving another agent.

Pick one agent already in production. Count the hours it saves, honestly. Then count every hour spent reviewing its output, correcting it, re-running it, or explaining it to someone who did not trust it. Subtract. That difference, not the first number, is what the agent returned.

Then track the second number over time, because that is the one that tells you whether you are building anything. A supervision cost that falls quarter over quarter means your corrections are landing somewhere the system can use. A supervision cost that holds flat means you have automated a task and hired a checker, and it will stay at exactly that level until something structural changes.

Finally, ask where the corrections go. If the answer is a document, a Slack thread, or one person's judgment, your most valuable operating knowledge is being generated daily and discarded daily. Route it into the context layer the agents actually read and the tax falls on its own, which is the case for owning an intelligence layer rather than renting a tool that I made in RAG Is Only Half the Story.

The Takeaway

Your agents are probably working. That is not the question. The question is what you are paying to keep them trustworthy, and whether that payment is going down.

Supervision is not a sign of a bad agent. It is a precise readout of how much of your business your agent cannot see. When 45% data access is the average and 87% of teams are manually checking that the context is still current, the supervision tax is not an edge case. It is the standard operating condition of enterprise AI right now, and it is invisible on every ROI slide I have been shown this year.

Put both numbers on the same page. Then go work on the half your agent cannot reach, because that is the only thing that ever lowers the bill.

Tracy Thayne is the founder of Expona, an AI consulting practice for smaller B2B companies in construction, manufacturing and the trades around them. Take the free Decision Line Assessment or subscribe to the blog (below) for plain spoken posts on putting AI to work in businesses that don't call themselves tech.