An AI assistant gives a customer-support agent a confident answer to an unfamiliar account-access problem. The answer arrives in seconds. It sounds plausible. It might even cite the right procedure.
What has improved?
Maybe the agent has better support. Maybe the system has only produced a faster mistake. We cannot tell from response time, output volume, or model confidence.
The same ambiguity appears across many AI-assisted workflows. Most AI evaluations still collapse three different outcomes into one:
- Faster output: the system produces something sooner.
- Better assistance: the system provides useful inputs to a person.
- Better decisions: the accountable person reaches a better outcome.
These are not interchangeable. The first is easy to measure. The third is the one that matters.
I call the difference judgment leverage.
It is not a property of the model. It is a measurable improvement in an accountable human decision process. AI can reduce the cost and delay of retrieving information, comparing options, and testing assumptions. The leverage appears only when those inputs help a person make a better decision—and when the workflow preserves evidence, challenge rights, feedback, and recovery.
This is the optimistic case for AI that I find most credible. AI does not need to replace human judgment to be transformative. It needs to make the inputs to competent judgment more available at the moment of work.
Output is a proxy, not the objective
The easiest AI metrics to report describe what the model does: tokens generated, tasks completed, benchmark scores, time saved, or percentage of a workflow automated.
Those measurements can be useful. They can also hide the outcome.
A system can draft twice as many reports while making important errors harder to notice. It can shorten handling time by pushing rework downstream. It can generate a recommendation that looks complete enough to discourage a second look. None of those systems has demonstrated judgment leverage.
The evaluation unit should be the decision loop, not the model response.
For a consequential workflow, ask:
- What decision is being made?
- Who owns it?
- What information did AI add?
- What evidence can the owner inspect?
- How does the system learn whether the recommendation worked?
- What happens when it is wrong?
If those questions are missing, the AI evaluation is probably measuring production rather than value.
What bounded evidence can—and cannot—show
The evidence I trust is bounded: generative AI can improve performance in specific settings. None of the studies below establishes that AI universally improves judgment.
In Generative AI at Work, Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond studied the staggered introduction of a conversational assistant across 5,179 customer-support agents at a business-software company. Access to the assistant increased issues resolved per hour by about 14% on average. (NBER working paper; October 2023 revision)
The operator interpretation is narrower than the headline number: this was evidence of higher throughput in one bounded workflow, not a general measure of better decisions. The effects also varied by experience and skill; experienced and highly skilled workers saw minimal productivity impact. In the paper’s preferred specification, the average change in customer satisfaction was not economically significant, and the change in resolution rate was small. (NBER working paper)
That is still meaningful. It shows measurable augmentation is possible inside a bounded workflow. It does not prove that more resolved cases always means better decisions, or that the result transfers to every occupation and company.
The interaction can also fail in the other direction. In a controlled experiment with 199 participants, Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos found that people sometimes followed incorrect AI recommendations even when they would have performed better without the assistance. (paper; CHI publication)
The design implication is direct: making a recommendation effortless to accept can weaken the review around it. Interfaces that required participants to think before seeing or accepting the recommendation reduced overreliance, though they did not eliminate it and were rated less favorably. The experiment used a meal-ingredient substitution task rather than a deployed workplace system, so it establishes a design risk—not how common that risk is across work. (CHI publication)
The operator lesson is uncomfortable: removing every point of friction can make the decision worse. The smoothest experience is not always the safest or most valuable one.
The judgment-leverage loop
Consider an illustrative support case: an agent must decide whether to escalate an unfamiliar account-access problem.
A useful AI system does not merely produce an answer. It supports a five-stage loop:
- Decision context. The system receives the relevant product version, permissions, prior troubleshooting steps, and the cost of an incorrect account change.
- Model assistance. It retrieves procedures, compares likely causes, proposes a diagnostic check, and exposes uncertainty and source links.
- Human authority. The agent decides whether to run the check, request more information, or escalate. The model cannot change permissions or close the case.
- Preserved evidence. The workflow records the inputs, cited procedure, recommendation, uncertainty, decision, and reason for an override.
- Feedback and recovery. If the recommendation is wrong, the agent can stop, reverse a reversible action, escalate, and label the failure for review.
Remove context and the answer can be locally plausible but operationally wrong. Remove human authority and assistance quietly becomes automation. Remove evidence and the recommendation cannot be audited. Remove feedback and the same mistake returns. Remove recovery and one bad suggestion can compound into damage.
These are not optional safeguards attached after the product is built. They are part of what makes the assistance valuable.
How to evaluate whether AI improves decisions: the five-part scorecard
Evaluate the complete decision loop across five dimensions: decision quality, access to relevant expertise, time to useful feedback, evidence quality, and recoverability.
Before deploying an AI system, establish a baseline for each dimension.
1. Decision quality
Baseline: Measure correct resolution, avoidable escalation, rework, reversals, and material error rates before AI enters the workflow.
Evidence of improvement: The assisted process improves one or more outcome measures without degrading customer impact or transferring errors downstream.
Gaming signal: The team reports answers generated, model agreement, or cases handled as decision quality while excluding rework and reversals.
The test is not whether the model produced a plausible answer. It is whether the final decision improved.
2. Access to relevant expertise
Baseline: Measure how often people wait for specialists, abandon a question, or escalate because verified guidance is unavailable when needed.
Evidence of improvement: More people reach relevant procedures or specialist-quality inputs in situations that previously lacked timely support.
Gaming signal: Every model response is counted as “expertise,” regardless of provenance, relevance, or review.
AI can make access to information cheaper. That does not make every response expert advice.
3. Time to useful feedback
Baseline: Measure the elapsed time between an initial decision and the first validated correction, useful insight, or justified escalation.
Evidence of improvement: The workflow reaches meaningful feedback sooner while maintaining or improving outcome quality.
Gaming signal: The team measures time to first output instead of time to validated feedback.
Speed matters most when it accelerates correction, not just production.
4. Evidence quality
Baseline: Audit whether current decisions preserve sources, assumptions, uncertainty, and the reasons for overrides.
Evidence of improvement: A reviewer can reconstruct what supported the recommendation and identify unsupported assumptions.
Gaming signal: Fluent explanations, generated rationales, or unverified links are treated as evidence.
A confident paragraph is not an audit trail. Evidence must be inspectable outside the model's own assertion.
5. Recoverability
Baseline: Document which actions can be paused, challenged, reversed, or escalated—and how long recovery takes.
Evidence of improvement: A decision owner can stop or reverse a bad path before damage compounds, with clear escalation ownership.
Gaming signal: A human approval button is called oversight even though the reviewer lacks time, evidence, authority, or a practical rollback path.
A nominal human in the loop is not enough. The human needs the means to disagree.
The strongest objection: assistance can weaken the operator
The case against judgment leverage is not that AI will never help. It is that fast, plausible recommendations can create false confidence. People may stop practicing diagnosis. Review can become ceremonial. The organization can retain formal accountability while the system shapes decisions nobody has time to challenge.
That risk is real enough to test for, but it should not be promoted into an unsupported universal claim that AI causes de-skilling.
The answer is to design for active judgment:
- Require the decision owner to identify the actual choice before requesting a recommendation.
- Separate evidence from generated interpretation.
- Make uncertainty and missing context visible.
- Sample overrides, reversals, and near misses—not only accepted suggestions.
- Preserve an escalation path that is practical under real time pressure.
- Periodically test whether people can still perform critical parts of the task without assistance.
NIST's voluntary AI Risk Management Framework similarly treats trustworthiness and risk management as concerns across the design, use, and evaluation of AI systems. It does not validate this scorecard or prove a deployment works. It reinforces the operating principle that evaluation belongs around the system, not just the model. (NIST AI RMF)
Use AI where judgment compounds
The best AI opportunities are not necessarily the workflows with the most text or the highest automation percentage. They are places where better framing, wider access to relevant knowledge, earlier correction, and recoverable action improve repeated decisions—whether the operator is resolving a support case, changing a process, or testing a research claim.
Use a stricter selection rule: prioritize repeated decisions where relevant evidence is currently delayed, errors can be observed, and actions remain recoverable. In those settings, AI can become infrastructure around the decision without becoming the decision owner.
This is also why benchmark spectacle is a weak adoption strategy. A model can perform impressively in isolation and still fail inside a workflow with incomplete context, weak evidence, rushed reviewers, and irreversible actions.
The operator question is narrower:
Did this system make an accountable human decision measurably better?
If the answer is yes, document how. If the answer is “it produced more,” keep measuring.
The optimistic claim worth making
Some of the loudest AI optimism imagines autonomous systems replacing the constraints of human expertise. I think the more credible opportunity is already ambitious enough.
AI can reduce the distance between a person, a difficult decision, and the relevant inputs needed to make it well: earlier feedback, wider access to useful knowledge, and a clearer comparison of options that might otherwise be missed.
But assistance earns the name judgment leverage only when the human decision process improves and remains accountable, inspectable, and recoverable.
Before approving the next AI system, do not ask only how much it can produce or how many steps it can perform alone.
Score the judgment loop.
Sources
- Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work”, NBER Working Paper 31161, with the October 2023 revision PDF.
- Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos, “To Trust or to Think”, with CHI 2021 publication record.
- National Institute of Standards and Technology, AI Risk Management Framework.