The Oversight Cost of AI Agents, and How to Design It Down
Supervising AI agents eats into the gains they promise. How to budget oversight hours and cut them with limits and review by exception.
AI agents need supervision, and supervision costs hours. In a 2026 survey, nearly half of agent buyers said oversight and rework ate more than a quarter of the gains they expected. That cost shrinks by design. Decide which actions need a person, set hard limits on the rest, and move actions out of review as they prove themselves.
The pitch for an AI agent counts the hours it saves. The first month in production counts the hours someone spends checking its work. Plan for those hours up front and they shrink every month. Leave them out of the plan and they eat into the gain.
The cost nobody budgets
Business cases for agents count the hours the agent saves. They rarely count the hours people spend watching it.
G2, a software review marketplace, surveyed 1,019 B2B software decision-makers in July 2026 (G2 2026 AI Agents Insight Report, vendor-sponsored survey). Its findings:
- 31% said employees spend 6 or more hours a week overseeing agent work.
- 45% said overseeing and correcting agent mistakes consumes more than a quarter of the productivity gains they expected.
- 46% reported higher-than-expected operating costs.
Analysts expect the gap to cost projects. Gartner predicted in 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls" (Gartner, 2025). That is a forecast, not a count, but the three reasons it names are the same three that oversight drives.
Some actions should always have a person on them. The work is deciding which ones, putting people only where a mistake is expensive, and keeping that list short.
Why oversight grows
Oversight grows for four reasons, all of them in the design around the agent:
- Every action gets the same review. Drafting an internal note and issuing a customer credit go through the same approval screen. Reviewers spend most of their time on actions where a mistake costs nothing.
- There are no hard limits. Without a cap on dollars or volume, the only safe policy is to check everything.
- Corrections don't feed back. A reviewer fixes the same kind of error every day because the fix never changes the agent's rules or reference data.
- Nobody owns the queue. Review becomes a side task. It piles up, people approve in bulk to clear it, and the review stops protecting anything.
A review step with no named owner follows a familiar path. In the first week, someone checks every item. By the second month, the queue is a backlog, approvals happen in a batch at the end of the day, and nobody can say who approved what. Give the queue an owner, a daily time slot, and a target size, the same way you would staff any other piece of work.
Put oversight in the budget
Estimate oversight before you build, the same way you estimate the time saved:
Weekly oversight hours = actions per week × share reviewed × minutes per review ÷ 60
Illustrative example. Numbers are hypothetical. An agent handles 2,000 actions a week. If a person reviews all of them at 1 minute each, that is about 33 hours a week, close to a full-time job. If only the 10% that carry real risk get a review, and a person spot-checks 1 in 20 of the rest, it falls to about 5 hours.
Same agent, same work. The difference is the review design. Put both numbers in the business case, and decide which one you are building.
Five design moves that shrink it
1. Sort actions by what a mistake costs
List every action the agent can take and put each one in a tier:
| Tier | Example actions | Default |
|---|---|---|
| Internal and easy to undo | Drafting a note, tagging a record, filling a field on a draft | Runs on its own, sampled |
| Internal, harder to undo | Changing a record other systems read, closing a ticket | Runs within limits, sampled more often |
| Leaves the company or moves money | Emailing a customer or vendor, issuing a credit, changing an order or price | A validation rule or a person's approval, every time |
Count how much of your agent's volume sits in the first tier. That is where the hours come back.
2. Set hard limits
Give the agent caps it can't exceed: a dollar amount per action, a number of actions per day, a list of the only tools and records it can touch. Anything over a limit stops and waits. Limits let you trust the routine actions without reviewing each one, because the worst case is bounded. When we build agents, the client sets these dollar and volume limits, and every agent action is logged.
3. Review by exception, sample the rest
People review what the rules flag: low confidence, a failed check, anything over a limit. Everything else runs, and a person reviews a small random sample each week. Sampling catches drift. Reviewing every action wears reviewers down until they approve in bulk.
4. Turn every correction into a fix
When a reviewer corrects the agent, record why. Once a week, look at the top reasons and fix them at the source: a missing cross-reference, a rule nobody wrote down, an instruction the agent reads wrong. A review queue should get smaller every month. Our email order intake article shows what that looks like for orders.
5. Graduate actions out of review, and back in on an error
Set a rule up front. For example: once a type of action has passed review a set number of times in a row without an edit, it moves from approval to sampling. One real error sends it back to approval. Pick the number with the business owner, based on what a mistake costs. The rule turns "when can we trust it?" from an argument into a policy.
What stays with a person
Some actions keep a person on them no matter how well the agent performs: anything that commits money, changes terms with a customer or vendor, or can't be undone. Our rule is simple. Nothing reaches a customer or changes a core system without a validation rule or a person's approval. If an automation fails, the work falls back to your manual process.
You would put the same control on a new employee with signing authority.
Questions to ask whoever builds your agent
- Which actions run on their own, which need approval, and who decided?
- What are the dollar and volume limits, and who can change them?
- Is every action logged, and can we see the log?
- How many hours a week of review should we expect in month 1 and in month 6?
- What happens to the work if the agent stops?
If the answers are vague, the oversight cost will land on your people. Our security page covers how we handle access and logging, and Dooder Operate covers who watches the queue after launch.
Next step
If you're planning an agent, or have one that needs more watching than it should, book a 30-minute call. We'll look at the actions, the limits, and where people belong in the loop. Our AI and agentic automation work is a fixed price for a written scope, quoted after the call. How we work explains the steps.
Get new articles by email.
Practical notes on making AI pay off in mid-market operations. No spam. Reply to any email to stop.
Related reading
- Build or Buy Software Now That Coding Agents ExistCoding agents cut the cost of writing custom software. Running it costs the same. How to compare build and buy over two years, and what to ask first.
- How PE Operating Teams Should Assess AI Readiness in a Portfolio CompanyA five-question framework PE operating teams use to segment portfolio companies by AI maturity and find where a fixed-price build creates the clearest value.
- Why AI Pilots Don't Move EBITDA (and What Does)Most AI pilots add spend without adding margin. Production systems move EBITDA when adoption is tied to one workflow metric. Here is how to set it up.
