A program of EffectiveSC
How We Use AISources checked October 24, 2026
Analysis
Agents and oversight
Today the review can cost far more than the AI's own work. How much agents really save depends on how cheaply their work can be checked.
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
113×
What an expert's review cost in GDPval, against what GPT-5 spent producing the work.
GDPval, derived by The Oversight Ratio · Estimate
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
The model's own cost is a sliver next to the review
Cost per task in GDPval's figures, in US dollars
Cheaper checking, not a better model, raises the ceiling
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
The ceiling, by what the review costs
Ceiling on an agent's cost advantage over the expert, by what the review costs as a share of the task
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 Workday's survey of 3,200 workers, fixing and redoing AI output ate nearly 40% of the time AI saved.
In a CMU and Stanford study of 48 workers on 16 tasks, handing whole workflows to agents made people 17.7% slower, largely from verifying and debugging the agents' work; using AI to assist made them 24.3% faster.
Sources: Workday, AI rework research (Jan 2026); CMU and Stanford, How Do AI Agents Do Human Work? (arXiv 2510.22780)
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
| Kind of work | Oversight as a share of agent spend |
|---|---|
| Checkable, high-volume work | About 5% to 15% |
| Judgment work | About 15% to 30%, plus human review time |
| Irreversible or regulated actions | Possibly more than the work itself |
| Across a portfolio: evals, monitoring and guardrails | 10% to 20% |
Planning ranges, not measured optimums. No study establishes an optimal oversight ratio, for people or for agents.
How often
How often do US workers use AI?
15% of US employees used AI at work daily in May 2026.
Work and productivity
Are companies getting a return on AI?
Only 37% of respondents in McKinsey's August 2026 State of AI report could attribute any EBIT impact to AI.
Analyst note
Why do the AI-use trackers disagree?
Both trackers now put daily use at about 11% to 15% of US workers; they differ on where it started and where it's heading, and neither tracks depth.
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.