AI agents work, but checking them eats the gains. What G2's new survey means for your office

What's new

On Oct. 14, G2, the business software review site, released its 2026 AI Agents Report. It surveyed 1,019 people who buy or help buy business software, in July 2026, across North America, Europe and Asia-Pacific.

The headline is good news for AI: agents have moved out of pilots. 86% of buyers purchased AI agent software or services in the past year, up from 69%. Two-thirds now run agents in limited or full production.

The catch is in the name G2 gave its main finding. G2 calls it the Distrust Tax: "the ongoing cost of running agents you don't fully trust." Companies get real speed from agents, then give part of it back checking, correcting and handling the cases the agent got wrong.

The numbers

AI agents: the payoff and the overhead

G2 2026 AI Agents Report. Online survey of 1,019 B2B software buyers, July 2026. G2 sells software reviews and buyer data.

1. Bought AI agents in the past year86%
2. Have agents in limited or full production66%
3. Require human quality review before production52%
4. Operating costs higher than expected46%
5. Oversight and rework eat more than a quarter of expected gains45%
6. Staff spend 6+ hours a week overseeing agents31%
7. Reduced headcount14%
8. Comfortable with full autonomy7%

Source: G2 news release, Oct. 14, 2026 (a paid press release).

A few more details from the release:

• The gains are real. Four in five companies using agents on high-volume work report cycle times at least 11% faster, and nearly half see at least 25%.
• Jobs moved more than they disappeared. 35% moved employees into higher-value roles. Just 14% cut headcount.
• Bigger companies pay more tax. 27% of enterprises need significantly more oversight than expected, nearly twice the 15% rate at small and mid-size businesses. G2 counts companies under 250 employees as SMB and 250-999 as mid-market.
• Nobody wants a fully free agent. 48% prefer agents with guardrails, 29% want them limited to specific tasks, and only 7% are comfortable with full autonomy.
• Per-seat pricing is falling out of favor. Just 9% think per-user pricing is the fairest way to pay for agents. Pay-as-you-go models rose from 38% to 57%.

Why the checking costs so much

Picture an agent that drafts sales orders from customer emails. If it's right 90% of the time, someone still has to look at every order, because nobody knows which 10% are wrong. If that person has to open the PDF, find the line, and compare it to the screen field by field, checking takes almost as long as typing. That's the distrust tax.

The answer isn't to skip the check. It's to make the check fast. The best review screens do three things:

• Show the source next to the result. The customer's PO line sits beside the order line it became, so the person compares at a glance.
• Flag what's different. A new ship-to, a price that doesn't match the price list, a part number the system didn't recognize. Those get the attention. The rest gets a quick look.
• Set limits by stakes. ServiceNow showed a version of this on Oct. 15. In a demo, as reported by SiliconANGLE, a customer-service agent could approve refunds under $25 on its own. An $18 refund went through; a $40 one was stopped for a person to review. That feature isn't out yet, but the idea works anywhere.

Review isn't the problem. Slow review is.

What the other side has right

• It's vendor research. G2 sells data and marketing to software companies, and it distributed these findings through a paid press release. G2 also says it used generative AI to help design the survey and analyze the results. We sell AI automation, so weigh our view the same way.
• The sample skews big. 38% of companies with agents in production run more than 100 agents, and half have an agent budget of at least $1 million. That's not most 40-400 person companies.
• Oversight may be worth it. Six hours a week of review can be a bargain if the agent did 40 hours of work. The survey measures how much of the expected gain was lost, not whether the project paid off.
• Some of the tax is temporary. As people learn which cases an agent gets right, they can check less. 77% of buyers re-evaluate their agents at least monthly, which is how trust gets earned.

Our take for a 40-400 person company

This survey is a fair challenge to how we build. We put a person on every order before it reaches the ERP. If that review is slow, we've just moved the work around.

So we design the review first. At a food-equipment manufacturer that supplies national grocery chains, the software reads each customer order and fills it in, and their team approves it. The job is to make that approval quick: the order and the source side by side, with anything unusual called out.

For a mid-size company, the G2 numbers carry one more lesson. You don't need 100 agents. You need one process, like order entry or AP, done well, where you can measure the time before and after, including the time spent checking.

What to do this month

1. If you already use an AI tool or agent, time the checking. For one week, track how long people spend reviewing and fixing its work. Compare that to the time it saves.
2. Ask vendors to show the review screen. Not the demo where everything is right. The screen where something is wrong, and how fast a person spots it.
3. Set limits by dollar amount or risk. Decide what can go through on its own and what always needs a person, the same way you'd set approval limits for a new hire.
4. Watch pricing. Per-seat pricing can charge you for every reviewer. Ask how the price works if the volume doubles.
5. Read the related guides: when AI agents went off script and why most AI pilots fail.