How to Evaluate Human-AI Collaboration Before You Scale: Eight Checks That Show Whether Decisions Improve

Last updated Nov 8th, 2026

A meta-analysis of 106 experiments, published by Michelle Vaccaro and colleagues in Nature Human Behaviour, found that, on average, a person working with AI did worse than whichever of the two performed better on its own. Most AI rollouts never run that comparison, because they track usage and satisfaction, and neither tells you whether decisions got better.

At The Decision Lab, we evaluate human-AI collaboration by testing whether the pair beats both the person and the AI working separately, against a control group, before a rollout scales. The pair also has to get there without wearing down people's own judgment or widening gaps between employees. We test that with eight checks drawn from our workforce AI adoption work and our experiments on how people make decisions with technology.

1. Task suitability. The task is one where human and AI judgment combined can beat both. Measure: the pair's score against each partner working alone on the same tasks.
2. Override behavior. People overrule the AI when it is wrong and accept it when it is right. Measure: correct overrides and wrongly accepted AI errors.
3. Error patterns. You know which side introduced each mistake. Measure: a log that tags every error by where it started.
4. Decision quality. Outcomes improve against a baseline, beyond speed or volume. Measure: a scored outcome such as forecast accuracy or rework rate.
5. Reliance. Trust in the AI matches how often it is right, and people's own skills hold up. Measure: acceptance rates against AI accuracy, plus periodic tasks done without AI.
6. Equity. Gains and harms are tracked for each group the workflow affects. Measure: every result broken down by group.
7. Accountability. Someone with the authority and information to intervene owns each AI-assisted decision. Measure: a named owner and a recorded override process for each decision type.
8. Testing before scale. A controlled pilot on real work shows the collaboration helps before it reaches everyone. Measure: the checks above, before and after, against a comparison group.

The first measurement compares the pair with each partner working alone on the same tasks. Without an AI-alone condition, a team can report that AI "improved performance" when the AI would have done better with no one checking its work.

The Vaccaro meta-analysis shows why the comparison matters. On decision tasks, such as classifying cases or forecasting, human-AI pairs lost ground against the better partner working alone. On creation tasks, such as drafting or generating ideas, pairs tended to gain, though the effect was not statistically reliable across studies. When people outperformed the AI on their own, adding AI improved results; when the AI outperformed people, adding a person made results worse.

A field experiment with 758 consultants at Boston Consulting Group, published in Organization Science by Fabrizio Dell'Acqua and colleagues, found the same split inside a single job. On tasks the AI handled well, consultants with AI finished about 12% more tasks and worked about 25% faster than consultants without it. Graders also rated their work about 40% higher in quality. On a task chosen to fall outside the AI's capabilities, consultants using AI were about 19 percentage points less likely to reach the correct answer than those working alone, who got it right about 5 times in 6.

We start every evaluation by sorting a team's work into tasks like these, because the same tool can help with one task and hurt the next one in the same afternoon. Our rule for clients is to hold off scaling a workflow until the human-AI pair has beaten both baselines on the outcome that matters for that task.

A low override rate can mean the AI is excellent or that people have stopped checking it. The useful measure is the match between overrides and AI errors: how often people overrule the AI when it is wrong, and how often they overrule it when it is right.

To measure this, build a test set of cases with known correct answers, including cases where the AI is deliberately wrong, and record each decision. A study of radiologists by Nikhil Agarwal, Alex Moehring, Pranav Rajpurkar and Tobias Salz found that giving them AI predictions did not improve their accuracy on average. The radiologists gave the AI too little weight and combined its predictions poorly with their own information. A study of overreliance on AI advice by Klingbeil and colleagues in Computers in Human Behavior found the opposite failure: participants who knew advice came from AI followed it even when it contradicted the information in front of them and their own assessment.

We have seen both failures inside one organization. When we increased AI adoption across a 20,000-person workforce at a Fortune 500 HR technology company, we found two employee segments with opposite patterns. Enthusiasts were already experimenting with AI and ran into friction, while skeptics doubted its value and stopped at first contact, so we read override and usage data separately for each group.

When a human-AI decision goes wrong, record where the error started. A simple log with four categories covers most cases:

• The AI's output was wrong and the person accepted it.
• The AI's output was wrong and the person corrected it.
• The AI's output was right and the person changed it into an error.
• The AI flagged a problem, and the warning was missed or never shown to the person.

Over a few weeks, the counts show whether people are catching more AI errors than they introduce. The pattern also tells you what to fix. Accepted AI errors point to a reliance problem, while correct output turned into errors or missed warnings point to a trust or interface problem. We use the same approach in our behavioral red teaming for large language models, where the goal is to find the conditions under which people accept harmful or flattering output.

Speed and volume are the easiest gains to measure and the easiest to mistake for success. Decision quality needs an outcome you can score: forecast accuracy, rework rate, customer complaints, the share of cases a senior reviewer would have decided differently.

Where the outcome arrives months later, we build a scored set of past cases with known results and run people through them with and without AI. The same principle applies when the decision belongs to a customer rather than an employee. To understand how shoppers decide on Pinterest, we built working replicas of Pinterest and two other platforms where people shop, then ran five behavioral experiments across 2,400 simulated shopping sessions. We tracked save rates, revisit frequency, image-view duration and time to purchase rather than asking shoppers what influenced them. The method carries over to AI-assisted workflows, which can be tested in a controlled copy of the real environment before they touch live decisions.

Good reliance means people lean on the AI about as much as its track record deserves. The aim is calibrated trust rather than maximum trust, and the failure to watch for is automation bias, the habit of accepting automated output without enough scrutiny. Measure it by comparing how often people accept AI output with how often that output turns out to be correct, broken down by task type. Self-reported trust is a useful second signal, but what people do with the output is the primary one.

Reliance has a second dimension that most evaluations skip: what heavy use does to the person's own skills. In a survey of 319 knowledge workers published at CHI 2025, Hao-Ping Lee, Advait Sarkar and colleagues found that workers with more confidence in generative AI reported less critical thinking on the tasks they described. Across the 936 work examples they shared, workers reported spending less effort on understanding information than they would without AI in about 4 in 5 cases.

A team can show better output this quarter while losing the judgment it needs to catch the AI's mistakes next year. We expect that trade-off to become one of the central questions in AI evaluation.

To measure it, give people periodic tasks without AI and compare their scores over time with a group that uses AI less. Design can also reduce overreliance. In an experiment with 199 participants, Zana Buçinca and colleagues found that asking people to commit to their own answer before seeing the AI's recommendation reduced overreliance more than showing explanations did. Participants rated those designs least favorably, so satisfaction scores alone would have pointed teams toward the worse option.

Calibrated reliance matters outside employee workflows too. When we helped Mila change how billions of phones communicate risk, its AI model calculated a graded risk score for a contact tracing app, and we found the more precise score only helped once people understood how to read it and act on it. We ran weekly surveys of Canadians on trust in tracing apps and replaced blunt alerts with graded risk guidance. That work contributed to Apple and Google moving their exposure notification system from yes-or-no alerts to multiple risk levels.

Averages hide who gains and who pays. Erik Brynjolfsson and colleagues studied 5,179 customer support agents and found that an AI assistant raised the number of issues resolved per hour by about 14% on average. Novice and lower-skilled agents gained about a third, while the most experienced agents saw little change. In the consulting experiment, consultants who scored below average on a baseline task improved their performance scores by about 43% with AI, measured against their own baseline, compared with about 17% for those above average.

Report effects separately for the groups that matter in the workflow. For employee tools, that usually means tenure, role, baseline skill and language. For customer-facing systems, add geography, accessibility needs, task complexity and the impact on the customer, and our guide to building customer trust in AI products covers reliance from the customer's side. Narrowing gaps between employees can be a real benefit, as long as it does not come from wearing down the judgment of experienced staff.

For every AI-assisted decision type, the evaluation should confirm that a named role is responsible for the outcome. That person needs the information to judge the AI's output and the authority to set it aside. The EU AI Act's human oversight rules for high-risk systems, in Article 14, require that the people overseeing a system can stay aware of the tendency to over-rely on it and can decide not to use its output. The NIST AI Risk Management Framework asks organizations to define roles and responsibilities for human-AI configurations and oversight.

To turn that principle into a procedure, write down four things for each decision type: who can override the AI, what information they need to do it, how the override gets recorded and what happens when they disagree with the AI. We then test the written version by asking the people closest to the work who would be blamed if an AI-assisted decision went wrong. When their answers differ from the procedure, people may defer to the AI to avoid blame or avoid it entirely to protect themselves, and either habit shows up in the override data.

The last check pulls the first seven together into a pilot. Choose a small number of teams, give them the AI-assisted workflow, and compare them with similar teams working without it over a fixed period, measuring the outcomes above before and after.

We used this design with the Fortune 500 HR technology company described in check 2. We designed 20 candidate programs, each aimed at a specific barrier for a specific employee segment, and selected eight to pilot. We ran all eight for one month with more than 100 employees against a control group, measuring employees' attitudes and behavior toward AI before and after. Confidence in exploring AI tools rose 41% in the pilot group and fell 26% in the control group over the same period. These are results from one company's pilot, published in our case study, and the programs are now scaling to the company's full 20,000-person workforce. We would expect the effect to differ in other organizations, and that difference is the reason to run a pilot of your own.

Controlled experiments are the strongest evidence, but they are not always possible. When teams cannot be randomly assigned, a staggered rollout, where some teams start later and serve as the comparison, still gives a far better answer than a before-and-after survey.

An evaluation describes one version of the workflow. Re-run the core measures whenever something changes that could shift how people use the AI:

• a new model version
• a rewritten prompt or system instruction
• a new interface or data source
• changed escalation rules or policies

This framework draws on peer-reviewed research on human-AI decision-making and on The Decision Lab's applied experimentation work. It is designed for evaluating workflows where people and AI share decisions, and it does not certify that an AI model is safe or compliant on its own.

In client work, we break the work into tasks and review the research on the decisions involved, then interview and survey the people doing it. From there we build scored test sets and run controlled pilots before recommending what to scale. When a pilot shows the pair doing no better than the stronger partner working alone, we change which parts of the task each side handles and test the new split before anyone scales it.

What is human-AI collaboration evaluation?

Human-AI collaboration evaluation measures whether a person working with AI makes better decisions than either would alone. It goes beyond model accuracy to four human questions: how people accept and override AI output, whether their own skills hold up, who benefits, and who is accountable. At The Decision Lab, we run them as controlled pilots on real work.

Which metrics show whether people and AI work well together?

The most informative metrics compare the pair with the person alone and the AI alone on the same tasks. Add the rate at which people overrule the AI when it is wrong, an error log showing which side introduced each mistake, a scored decision outcome, and results broken down by employee group.

How do you measure overreliance on AI?

Build a test set that includes cases where the AI's answer is known to be wrong, and record how often people accept it anyway. Compare acceptance rates with the AI's actual accuracy for each task type. Periodic tasks completed without AI show whether people's own skills are holding up over time.

When should a human override AI?

A person should override AI when they have information the AI lacks or when the output conflicts with what they know about the case. Evaluation checks whether that happens in practice. The goal is overrides that track AI errors, rather than a low or high override rate in itself.

How long should a human-AI collaboration pilot run?

A pilot should run long enough to capture the decisions that matter and for early enthusiasm to fade. Our Fortune 500 pilot ran for one month against a control group. Tasks with slow outcomes, such as forecasts or hiring, usually need a scored set of past cases alongside the live pilot.