Over four days at the end of September, I personally merged about 500 pull requests with a team of agents. Every request drove specific business value. I also made about 165 decisions a day. At that pace, decision fatigue increased the risk of mistakes.
I made mistakes.
Every team in your organization that adopts agents will run into the same problem, and most won’t notice it happening. Every AI safeguard that asks a person for approval spends against that person’s attention, and attention runs out. When it does, people approve by reflex, and the safeguard stops protecting.
I believe the number-one AI risk is the tired human who signs off on the agent. CIOs should weigh each approval by the business outcome it protects and the risk it prevents, spending human attention only where the stakes justify it.
I call that discipline the Attention Budget.
We’ve Seen This Before
Long before AI, hospitals learned what happens when a safeguard asks for attention too often. In 2013, The Joint Commission estimated that 85% to 99% of bedside alarm signals need no clinical intervention and that clinicians “become desensitized or immune to the sounds.” Of 98 alarm-related events reported to the commission from 2009 to mid-2012, 80 ended in death, and alarm fatigue was the most common contributing factor.
Security teams learned the same lesson. In 2022, an attacker with a contractor’s stolen password triggered login approval requests on the contractor’s phone again and again. According to Uber, “Eventually, however, the contractor accepted one, and the attacker successfully logged in.”
In both cases, the control was still in place, but the person had learned that the answer was almost always yes. A control people have learned to dismiss is harder to spot than a missing one because it still looks like a control on the audit report.
AI Multiplies the Prompts
Agents generate approval requests faster than any monitor or login screen. The people approving them experience the same fatigue clinicians did. An August study of 1,053 paid testers using Claude Code found human reviewers caught 13.6% of dangerous commands, while the Claude classifier caught 89%. Early in a session, people blocked about 17% of dangerous commands; after 50 or more prompts, they blocked about 5%. Users approved 97% of the prompts they saw. Anthropic calls this pattern approval fatigue: the more often people are asked, the less closely they look.
Further, GitHub reported in May that more than one in five code reviews on its platform now involve an agent and that human review capacity hasn’t kept pace with volume. A June study of 11,429 reviews by 400 repeat reviewers of AI-agent pull requests found that reviewers’ approval rate rose from 30.1% to 36.8% while they left 22% fewer inline comments.
Reviewers approved more often and looked less closely.
Researchers from Boston Consulting Group found that direct oversight of AI agents was the strongest driver of what they called “AI brain fry” among 1,488 U.S. workers.
Most AI governance advice recommends adding checkpoints, and much of that advice is right. An agent with no checkpoints at all creates exposure, but problems arise when every new checkpoint becomes another prompt for the same few people. The organization feels safer even as its approvers stop reading.
What Agent Scale Looks Like
Anthropic is an outlier. The company scaled agents before almost anyone else, which makes its experience a roadmap for what the rest of us can expect. Anthropic’s engineering team reports that the company went from merging approximately 2,500 changes per day in May 2026 to more than 30,000 per day by September. Nobody can read 30,000 changes a day line by line.
Instead, Anthropic redesigned how its people spend their attention.
Its team members “still review every change request,” but “they now examine the intent of the change as opposed to reading every line of code.” Its review tooling distinguishes change risk, so humans focus on the most impactful changes.
That design is the Attention Budget in practice. Humans remained involved, but their attention was directed toward the decisions that mattered most.
The Attention Budget Belongs to the Organization
Every team that adopts AI adds approvals. Engineers approve agent-written code, finance staff approve AI-drafted invoices and payments, support leads approve AI replies to customers, and managers sign off on the work their teams’ agents produce. In most organizations, nobody adds them up.
The CIO is the one leader positioned to see all usage.
You should treat your organization’s attention the way you treat the IT budget, as a finite resource that every AI initiative draws on and that someone has to allocate intentionally.
In my own organization, we write that allocation down and continue to revise it. Our agents work from a shared rulebook that defines who makes each decision, while our tooling checks routine work automatically. This leaves people with far fewer decisions to make, each with business consequences.
Set Policy for Classes of Decisions
As CIO, work with your leadership team to decide which classes of decisions always need a person and which ones get automated and logged. Irreversible actions, money movement, regulated data, and anything a customer sees are the usual candidates for human-in-the-loop (HITL) approval.
Each organization will draw that line differently.
Any sign-off that a regulator or your auditors require stays in place regardless. Everything else should default to automation with a complete record your auditors can review, and teams can design their own workflows within that policy. The same policy applies to your leadership team’s decisions. The codified decision-making tenets I described in an earlier article allow you to settle entire classes of vendor, model, and build-versus-buy decisions once.
The policy should also define what the person decides. A classifier catches far more bad commands than a tired reviewer does, but people are better at judging intent and business context.
Each class of decision also needs an owner who is accountable for its outcomes, including failures that slip through an automated approval. That owner ensures the decision loop continuously improves.
Measure It Like Spend
For each team, class of decision, and approver, you should track four numbers.
- Approval Rate: The percentage of approvals granted.
- Decision Time: How long people take to make a decision.
- Catch Rate: How often an approver catches an issue.
- Escaped Errors: How many errors slip through, regardless of whether a person or an automated check approved them.
Fatigue builds throughout an individual’s day, so a healthy team average can hide an overloaded approver.
My rule is that a class of decisions approved more than 95% of the time, decided in seconds, with no recent catches, is an error worthy of investigation and discussion. Irreversible actions are the exception. A single catch there can prevent a loss you can’t undo, so you should keep those approvals and make them rare enough that people still read them.
You should review these numbers alongside your other critical business metrics, including your token spend.
In my article on token spend, I argued that token spend on its own is a vanity metric, and approvals behave the same way. A team rewarded for using more tokens will generate more approvals, and each approval will get less attention. Neither number tells you whether the time and tokens delivered business value. Therefore, you should count human attention and the errors that slip past it as costs alongside tokens:
Cost per outcome = (token spend + approver time + cost of escaped errors) ÷ business outcomes delivered
Catches tell you whether the approver time is actually valuable, and escaped errors tell you whether it is enough. A class of decisions with frequent catches is likely returning value, although frequent catches also mean the agents in that class need better guardrails. A class with no catches and no escaped errors is a candidate for automation. A class with no catches and rising escaped errors is being rubber-stamped.
Your Job on Monday
Start by creating an inventory of AI approval points across your organization.Then bring those inventories to your next leadership meeting, agree on the classes of decisions that always need a person, and name an owner for each class.
You should also tell your teams why.
People who have spent years equating approvals with safety will read fewer approvals as less control, so show them the numbers.
Agents will keep getting faster, and the number of decisions they surface will keep growing. Human judgment will not grow with them. The CIOs who budget their organization’s judgment deliberately will move faster with AI, and they will catch the mistakes that matter when they happen.
