How much will companies spend checking AI agents' work?

In OpenAI's GDPval study, an expert's review cost about 113 times what GPT-5 spent producing the work. Estimate

Where an expert reviews each output and redoes the ones that fail, the AI's own cost was 0.25% of the total, review 28% and redoing failed outputs 72%. Estimate

Checking people's work costs from under 1% to about 27% of the time it took

When people check people's work, checking costs from under 1% of the time it took to make it (a pass/fail acceptance check: 11.4 minutes on a 28.9-hour project) to about 27% (a full expert review: 109 minutes for work that took 404). Estimate

If people check everything, agents save far less

In GDPval (experts in 44 occupations, 220 tasks), GPT-5 working alone was 90x faster and 474x cheaper than the expert; with the expert reviewing one attempt and redoing failures, the advantage dropped to 1.12x faster and 1.18x cheaper.

Letting the model retry before the expert stepped in raised the speed and cost advantages to 1.39x and 1.63x; with the older GPT-4o, the combination did worse than the expert working alone.

If an expert reviews every output and redoes every failure, GPT-5's 39% acceptance rate gives a 1.18x cost advantage and even 99% acceptance gives only 4.0x; no model can push it past 1 divided by 24%, about 4.2x. Estimate

On judgment calls, experts mostly took the agent's side

In a randomized trial Anthropic published in October 2026, 50 professional analysts used Claude Code to analyze data for a fictional farm cooperative and advise whether to roll out a program. For each of four analytical choices, a hidden coin flip set which of two defensible options their agent argued for. Analysts went with the agent's side 86% of the time, against 50% by chance.

When an analyst had already taken a side and the agent argued for the other, the analyst switched 60% of the time, against 6% when the agent agreed. That comes from only 38 such cases.

Asked afterwards to explain their choices, 77% repeated the agent's arguments, and only 15% mentioned the agent. None suspected it had been given a side.

The agent's lean moved the final recommendation less: by 17% of the distance from discontinuing the program to rolling it out, which was not statistically significant. Analysts working with the agent made serious errors in 3 of 50 analyses, against 6 of 11 in a smaller, less experienced group that chose to work without it.

On calls with no right answer, a person reviewing an agent's analysis may not be an independent second opinion. Anthropic says it made the agents' leans stronger than usual and tested one task, one field and one model.

Source: Anthropic, How AI agents shape experts' judgment calls (Oct 27, 2026); full paper. Anthropic makes Claude and studied its own agent; Claude also helps research this site. Checked October 28, 2026.

In security operations, most CISOs still want a person to approve what agents do

In a survey a16z published in October 2026, 100 chief information security officers (CISOs) at companies with more than 1,000 employees described how far AI already reaches into their systems. Counting limited as well as broad production use, 92% say they have employees' generative AI tools in production, 83% coding assistants or agents, and 55% MCP servers or other agent-to-application integrations, which let agents work inside other software. 34% say they have autonomous agents with tool access or write permissions.

For AI agents in their own security operations, 52% are comfortable letting them act only with a person's approval, and 11% want them only to recommend. 36% are comfortable letting them act on their own on low-risk actions, and 1% on most actions. Within 12 months, 60% expect to be comfortable with autonomy on low-risk actions, and 7% on most.

63% of their responses describe security slowing at least some AI adoption, including reviews, approval requirements and restrictions on particular uses. Asked to pick their five biggest worries from a list of 16, 46% chose agents acting outside their permissions or being compromised.

In security operations, then, a person's approval is today's most common check on an agent, and most of these CISOs expect to drop it for low-risk actions within a year. That is an expectation they state, not a change anyone has measured.

It is one investor's survey of 100 CISOs at large companies. a16z does not say how the 100 were chosen or when they answered; its charts are dated October 18, 2026. It counts limited use as production.

Source: a16z, The New Security Agenda (Oct 28, 2026), text and charts. a16z led Armadin's Series B and invested in Neo, both named in the piece. Checked October 28, 2026.

Forecasts put oversight at roughly 10% to 20% of agent spend by 2030

Analyst forecasts and Effective SC's projection converge on oversight at roughly 10% to 20% of agent spend by 2030, close to the human ratio. Estimate

Analyst forecasts bracket that range: Gartner expects guardian agents to take at least 10% to 15% of agentic AI markets by 2030, IDC finds enterprises planning to put 16.7% of AI spend into AI and agent security and governance, and Forrester sees AI governance software reaching 7% of AI software spend by 2030. Estimate

Effective SC's token projection puts checking at about 5% of enterprise tokens in 2026, rising to 15% to 20% by 2030 as agents take over checking from people. Estimate