The distrust tax: how to cut the hours your team spends babysitting agents

G2's 2026 AI Agents report, published on 14 October and based on a survey of more than 1,000 B2B software buyers, found that 66% of organizations now have agents in limited or scaled production. It also named a cost: the "distrust tax," defined as "the ongoing cost of running agents you don't fully trust." Oversight is not a phase you pass through; it is a recurring cost, and it only shrinks if you design it to. This article shows how.

What the survey measured

Three findings from the report matter for operations:

• 31% of buyers say employees spend six or more hours a week "bot-sitting," which G2 defines as overseeing agent work.
• 45% say the time spent overseeing and correcting agent mistakes eats more than a quarter of their expected productivity gains.
• 52% require human quality review before production, and 77% re-evaluate agent performance at least monthly.

These are survey answers from software buyers, not measurements of any one operation. Read them as a signal: most teams pay for oversight, and few have designed it.

Why oversight stays expensive

Oversight stays at full cost when it has no structure. A person reads everything because nothing says what can be skipped. Three causes repeat:

1. No definition of a good result, so every output needs a human judgment call.
2. No record of past errors, so the same mistake is caught again and again by hand.
3. No difference between low-risk and high-risk work, so a draft note gets the same review as a refund.

Step 1: write the check before the work

For each task type, write a pass/fail check a machine can run: required fields present, amounts within limits, the right customer record matched, sources attached. Anything that fails goes to a person. Anything that passes is not reviewed one by one.

Step 2: route by risk, not by volume

Split the agent's actions into three lanes:

• Free: reads and internal drafts. No review.
• Sampled: low-risk actions that passed their check. A person reviews a small random sample each week.
• Gated: irreversible or external actions (sending to a customer, paying, deleting). Always approved, or covered by a rule written in advance.

Step 3: turn every correction into a check

The error log is what lowers the tax. When a person fixes an output, record what was wrong and write the rule that would have caught it. Next week that rule runs on its own. Without this step, review effort stays flat forever; with it, the share of work that needs a person falls, and you can show it.

Step 4: measure the tax itself

Track four numbers per workflow, from your own records:

• Hours of human review per week.
• Share of outputs that fail the automatic check.
• Share of sampled outputs a person had to correct.
• Corrections that became new checks.

If review hours do not fall while the other numbers improve, you are still reading everything. If the correction rate rises after you reduce sampling, widen the sample again. Oversight should move with evidence, not with optimism.

Where a visible path helps

The same logic applies to conversations. In the Rodo case, WhatsApp, social and web enquiries were brought into a service operation with a measurable resolution path. In the first measured window, 82% of 604 live-created enquiries reached Resolved, with a median of about one hour. Those figures describe pipeline records, not all conversations, and there is no comparable earlier baseline. The point for this article is narrower: when every item has a visible stage, you can see where people need to look, and where they do not.

What to do this week

Pick one agent that a person currently reviews in full. Write its pass/fail check on one page, assign each of its actions to a lane, and log every correction for two weeks. Then count how many corrections could have been caught by a rule. That number is the first measure of your distrust tax, and the first lever to lower it.