AI Copilot Productivity Statistics: Human-in-the-Loop Results for 2026

AI copilot productivity statistics do not produce one universal estimate. Results range from a 55.8% reduction in completion time on a contained programming exercise to a 19% increase in completion time for experienced developers working in repositories they knew well. Both findings came from randomized studies. They measured different people, tools, and tasks.

The best evidence supports a conditional conclusion. A copilot can raise output and quality when its suggestions fit the task and a worker can check them efficiently. It can also add review work or steer a person toward a wrong answer. Human involvement is part of the system, but the presence of a person does not guarantee a better result.

This 2026 review reports the newest traceable evidence available for customer support, consulting, professional writing, and software development. The studies were conducted and published in different years. The year in the title describes this evidence review, not the date of every experiment.

AI copilot productivity results at a glance

The studies below use different outcomes. Time saved on one writing assignment cannot be averaged with customer issues resolved per hour or software tasks completed.

These results answer a more useful question than whether copilots work in general. They show where assistance changed finished output under a defined test.

Customer support: 13.8% more issues resolved per hour

Brynjolfsson, Li, and Raymond studied the staggered introduction of a conversational assistant at a Fortune 500 enterprise software company. Their data covered 5,179 customer support agents, including workers who received access at different times and workers who did not receive it during the observation period. The assistant suggested responses, but agents could edit or ignore them.

Access increased issues resolved per hour by 13.8%. The components of that change were also reported: agents spent about 9% less time per chat, handled roughly 14% more chats per hour, and increased their overall resolution rate by 1.3%. Customer sentiment improved, and requests to speak to a manager declined. NBER, 2023

The average hides a large experience effect. Productivity gains reached about 35% for the least skilled and least experienced workers, while the most experienced workers saw small gains or none. The authors' evidence suggests that the tool made successful response patterns easier for newer agents to use.

This was a human-in-the-loop workflow. The model did not close a support case on its own. The agent remained responsible for the conversation and could reject a suggestion.

Consulting: faster and better until the task crossed the frontier

The BCG experiment randomly assigned 758 consultants to no AI, GPT-4, or GPT-4 plus a prompt overview. Participants completed realistic consulting tasks after the researchers established an individual performance baseline.

On 18 tasks judged to be within GPT-4's capability, AI users completed 12.2% more work and finished 25.1% faster. Independent graders rated their output more than 40% higher than the control group's output. The effect was not confined to weaker performers. Consultants below the baseline performance threshold improved by 43%, while those above it improved by 17%. Harvard Business School, 2023

One task was deliberately chosen because GPT-4 tended to reach the wrong conclusion. On that task, consultants with AI access were 19 percentage points less likely to give the correct answer. The tool produced plausible material, and human judgment did not consistently catch the error.

That reversal is central to human-in-the-loop planning. A reviewer cannot check what they do not recognize as uncertain. Teams need an evaluated task boundary, source access, and an escalation rule. A generic instruction to "review the answer" is not a control.

Professional writing: 40% less time and 18% higher quality

Noy and Zhang recruited 444 college-educated professionals for occupation-specific writing assignments. The preregistered experiment randomly gave half of them access to ChatGPT. Participants wrote documents such as press releases, short reports, and analysis plans, and experienced professionals graded the output.

ChatGPT reduced average completion time by 40% and raised average quality by 18%. The treatment also narrowed the productivity distribution because workers with lower initial scores gained more. Participants shifted effort away from rough drafting and toward idea generation and editing. Science, 2023

This result applies to short, self-contained professional writing tasks. It does not establish the same gain for research that requires original sourcing, confidential records, legal review, or several approval rounds. For those workflows, source checking and correction time belong in the productivity measure.

Software development: field output rose 26.08%

Cui and coauthors combined three company-run randomized field experiments at Microsoft, Accenture, and an anonymous Fortune 100 company. The pooled sample included 4,867 software developers. A random subset received access to an AI coding assistant that suggested code completions during normal work.

Across the three experiments, access increased completed tasks by an estimated 26.08%, with a standard error of 10.3%. The company-level estimates were noisy, but the pooled result was statistically significant. Developers with less tenure showed larger gains and adopted the tool more often. MIT Economics, 2025

The outcome was completed work, not raw lines of generated code. That makes the result more useful for operating decisions, though it still reflects the task definitions and development processes at those three companies.

Why a 55.8% lab gain and a 19% field slowdown can both be true

GitHub's randomized study asked 95 professional developers to build the same HTTP server in JavaScript. Forty-five participants received Copilot. The Copilot group averaged 1 hour 11 minutes, compared with 2 hours 41 minutes for the control group, a 55.8% reduction in completion time. The reported 95% confidence interval for the speed gain was 21% to 89%, and automated tests checked correctness and completeness. GitHub, 2022

METR tested a different setting. Sixteen experienced developers worked on 246 bug fixes, features, and refactors in mature open-source repositories. They had contributed to those projects for an average of five years. Each task was randomly assigned to allow or disallow AI, and completed work had to satisfy the project's usual standards. With early-2025 tools, developers took 19% longer when AI was allowed. METR, 2025

The developers predicted a 24% speed gain before the trial. After the trial, they still believed AI had made them 20% faster. Recorded completion time showed the opposite. METR does not claim that AI slows most developers. Its result covers experienced contributors, familiar and complex codebases, and the tools available from February through June 2025.

The two experiments differ in task scope, repository context, participant experience, and review burden. Boilerplate in a fresh assignment leaves less project history to recover and fewer integration constraints to check. Work in a mature repository can make reading suggestions, correcting them, and validating changes more expensive than writing the code directly.

What human-in-the-loop should mean in a measured workflow

"Human in the loop" can describe several very different designs. A worker who clicks accept is technically present. So is a domain expert who checks evidence, edits the output, and can stop the process. Only the second design provides meaningful oversight.

A measurable copilot workflow should record:

The denominator matters. Measuring prompts sent, words drafted, or code suggested can reward activity without showing whether the work was accepted. End-to-end time should include source retrieval, review, tests, corrections, and escalation.

How to run a useful internal copilot test

Start with one bounded task and a representative sample. Randomly assign comparable items to the existing workflow and the copilot workflow. Keep the acceptance rubric the same for both groups.

Record worker time and reviewer time separately. Grade outputs without telling evaluators which method produced them when that is practical. Report sample size, confidence intervals, task exclusions, and the version of the tool. Segment results by worker experience because the customer support, consulting, and software studies all found different effects across skill or tenure groups.

Set stopping rules before the trial. A high-severity defect, unsupported regulated claim, or unauthorized disclosure should pause the test even if average completion time improves. Workers also need a clear route to reject suggestions without a productivity penalty.

Teams planning a mixed workforce can use the broader AI and human workers collaboration research to compare operating models. If recurring research, scheduling, documentation, or coordination needs a named human owner, review the virtual assistant service. These internal pages describe staffing options, not evidence for the statistics above.

What the evidence supports in 2026

The strongest positive studies share two features: the task was defined, and the assisted worker remained responsible for the result. Support agents could reject suggested replies. Consultants applied GPT-4 to a known task set. Developers used code suggestions within existing company workflows.

The negative results show where that arrangement breaks down. BCG consultants followed an AI toward a wrong answer outside its capability. Experienced open-source developers spent more time completing real repository work even though they believed the tool saved time.

For workforce planning, use study results as a range of possible outcomes, not a forecast. Test the actual task, count review labor, and score finished quality. Keep a person in control when an error affects a customer, employee, payment, contract, or regulated decision.

Frequently asked questions

How much do AI copilots increase productivity?

Rigorous estimates vary by task. The studies in this review found 13.8% more issues resolved per hour in customer support, 25.1% faster completion on AI-suited consulting tasks, 26.08% more completed tasks across three software companies, and a 19% slowdown for experienced developers in mature open-source projects.

Do AI copilots improve work quality?

They can. BCG consultants produced work rated more than 40% higher in quality on tasks inside GPT-4's capability, and participants in a professional writing experiment improved quality by 18%. On a consulting task outside the model's capability, however, AI users were 19 percentage points less likely to reach the correct answer.

Who gains the most from a workplace copilot?

Several studies found larger gains for workers with less experience or lower baseline performance. That pattern appeared in customer support, consulting, and the three-company software study. It does not mean every novice will improve, or that experienced workers cannot benefit.

Why can AI make experienced workers slower?

Reviewing, correcting, and integrating suggestions takes time. In METR's randomized trial, developers worked in large repositories they knew well, and AI added 19% to completion time. The study does not isolate one cause or generalize the result to all software work.

What is the best KPI for a human-in-the-loop copilot?

Use accepted output per paid hour together with escaped defect rate. The first captures worker and reviewer time. The second prevents faster drafting from hiding quality failures.

Should a company use vendor productivity statistics?

Use them when the vendor discloses the sample, assignment method, task, outcome, and quality checks. Treat the result as evidence for that setting. Reproduce it on your own work before using it in staffing or financial forecasts.