AI moved the bottleneck to judgment. Design for it.
McKinsey argues that AI has sped up the analysis behind strategy far more than the decisions it feeds. The same is true inside operations. Human review is now the constraint in most AI workflows, and it has to be designed, staffed and measured like one.
What happened
In "The judgment bottleneck," published in October 2026, Alex D’Amico, Bruce Delteil, Laura LaBerge and Matt Banholzer of McKinsey make a simple point. AI now speeds up research, synthesis, simulation and communication in strategy work. The judgment and alignment that turn analysis into a decision can’t be compressed the same way. As in any system, the slowest step sets the pace, and judgment is now the slowest step.
The examples show the gap. One executive said their strategy team now completes five weeks of financial analysis in under two days. At another company, a junior team produced analysis ten to 20 times faster and ran three to five times as many scenarios, and executives needed more time to review it, because they no longer knew which assumptions and data sources sat behind it. The authors name three failure modes: chasing efficiency instead of better outcomes, burying leaders in more decisions than they can absorb, and losing the junior work that used to build judgment.
Their response has three parts. Expand judgment capacity by sorting decisions into tiers: routine decisions automated within guardrails, uncertain ones human-led with AI support, and high-stakes ones kept with accountable leaders. Build judgment on purpose, by training people to understand and question what is inside AI-generated analysis. And fight the erosion of accountability by giving every consequential decision a named owner.
My view
The article is about strategy. I see the same thing every week in operations: claims files, credit reviews, month-end close. The model drafts in seconds. Then the draft waits for an examiner, an underwriter or a controller who has twenty others in the queue.
"Human in the loop" is a capacity plan
Most AI business cases count the time the model saves and quietly assume the review is free. It isn't. If an assistant prepares 2,000 cases a week and each needs four minutes of checking, that's more than 130 hours of skilled review, every week, from people who already had full jobs. Size that capacity before launch. Otherwise the queue forms, and either the backlog grows or people start approving without looking.
Route by consequence
Judgment is scarce, so spend it where it changes the outcome. Two questions decide where a task belongs: what would an error cost, and how much human judgment does the task really need? Routine, low-consequence work can run within tested limits and be sampled afterward. Consequential work needs a named person to decide every time. That person needs the evidence in front of them, the source passage, the flagged gap, the rule that applied, so checking takes minutes and doesn’t turn into redoing the work.
Fast approval can mean weak review
When reviewers are swamped, the approval rate goes up and looks like success. It can just as easily mean people have stopped checking. Sample approved work and compare it with the source records. Track how often reviewers correct the draft. A correction rate near zero is a warning.
Protect the training ground
The authors’ third failure mode worries me most. Junior examiners and analysts built judgment by doing the ordinary work that AI now drafts. If they only approve finished answers, the next generation of reviewers never gets the practice. Give them real cases to work through, ask them to explain the assumptions behind a draft, and have someone senior check their reasoning, not just their conclusion.
The decision has to have an owner
Speed exposes unclear decision rights faster than anything else. If nobody can say who accepts the result, who can stop the workflow and who owns the exception, faster analysis just builds a bigger pile of unanswered questions. Settle those before the build.
Why it matters
For the industry. The firms that get value from AI won't be the ones with the fastest models. They'll be the ones that redesigned the work around the scarce step: who decides, on what evidence, within how much time. That is a workflow and staffing decision. Buying another tool doesn't settle it.
How it connects to our work. Our sample release decision brief shows what a reviewer signs and the review steps behind it, including weekly sampling of approved work. The 90-Day First Governed Capability uses the same two questions, cost of error and need for judgment, to set what an agent may do. Adoption by Design covers training people to check the work. Four steps for your next AI workflow:
- Count the review. Put reviewer minutes per case and the weekly review hours into the business case.
- Route by consequence. Decide which tasks run within limits and are sampled, and which need a named person every time.
- Put the evidence beside the draft. Show the source passage and any gaps, so checking is quick and real.
- Watch the correction rate. Sample approvals weekly and treat a near-zero correction rate as a reason to look harder.
Sources
- The judgment bottleneck: Balancing AI and human speeds in strategy, Alex D’Amico, Bruce Delteil, Laura LaBerge and Matt Banholzer, McKinsey & Company, October 2026
Linked sources are the record. Check them for the full reporting.