AI drafts faster than your team can check it

AI tools make first drafts fast. They do not make checking those drafts fast. A large new study of software teams found that AI agents raised the amount of code people wrote, while the amount of finished work barely moved. The extra drafts piled up at review. Most businesses are about to meet the same problem with proposals, reports and emails.

What the study measured

Fiona Chen and James Stratton, economists at Harvard University, studied data from Jellyfish, a company that measures how engineering teams work (Chen and Stratton). The data covers about 300 million work events from roughly 700,000 workers at 718 firms, from January 2021 to March 2026. It includes code changes, task tickets and calendar time. The authors compared firms that adopted AI tools earlier with firms that adopted them later.

Their findings for AI agents, the tools that complete whole tasks with little supervision:

• Lines of code rose 30%. Commits, the small saved changes, rose 20%. Pull requests, the larger changes sent for review, rose 23%.
• Finished work did not rise by a measurable amount. The authors could rule out any gain larger than 12% in completed tasks.
• Headcount did not change by a measurable amount either.
• The average time to review a pull request rose 49%.
• The share of pull requests sent back for changes nearly doubled. Comments per review rose 35%.
• The share of workers doing reviews rose 14%.

The authors call this a review bottleneck. They estimate that engineers spend only 30% to 40% of their time writing code. The rest went to planning, review and testing. Speeding up one step moved the queue to the next. Ars Technica reported the study on October 12, 2026.

productivity gains do not fully pass through to changes in software output or employment.

The same pattern outside software

A survey from Glean, a company that sells AI tools for work, points the same way for office jobs. Glean's Work AI Institute surveyed 6,000 full-time digital workers, people who work mostly on computers, in the United States, the United Kingdom and Australia (Glean). Workers said AI saved them about 11 hours a week. They also said they spent 6.4 hours a week on what Glean calls botsitting: giving the AI missing context, checking its output, fixing errors and rerunning prompts. That is 37% of their AI time. Business Insider covered the report in June.

Two cautions apply. The Glean figures are what workers estimated, not measured time. Glean also sells products that address the problem it describes. Still, the direction matches the Harvard data: drafting got cheaper, and checking took up much of the saving.

Glean also reported that 69% of AI users admit to sending work they cannot fully explain. That is the cost of skipping the check instead of doing it.

Why this matters for a small business

In a 10-person firm, the reviewer is often the owner or one senior person. If AI doubles the drafts that reach that person, their week fills with checking. Work waits in their inbox. Or they start approving without reading, and errors reach clients. Neither outcome shows up in a count of drafts produced.

Step 1: count the checking, not the drafting

Pick one kind of document that AI now helps produce, such as proposals or client reports. For two weeks, track three numbers: how many drafts reach review, how long each review takes, and how many go back for changes. If drafts go up and finished work stays flat, the bottleneck has moved to review.

Step 2: decide what done looks like before anyone drafts

Much review time goes to deciding what the document should have said. Write that down first. Before a proposal is drafted, list the scope, the price, the dates and the three points the client must see. The reviewer then checks the draft against the list, which is faster than judging it from scratch. In the study, the share of work sent back for changes nearly doubled. A clear brief cuts that rework before it starts.

Step 3: make the drafter check first

The person who asked the AI for the draft should be the first reviewer, not the last link in a chain. Before anything reaches the senior reviewer, the drafter confirms every figure, name and date against its source. A draft the drafter cannot explain does not go forward.

Step 4: send smaller pieces

A long AI draft is slow to review because every paragraph might hide an error. Shorter pieces are easier to check. Review the outline before the full draft. Review the pricing section on its own. A reviewer who sees one section at a time can approve most of it quickly and spend attention where it is needed.

Step 5: give review real time

The firms in the study moved people into review: the share doing it rose 14%. A small business can do the same on purpose. Block review time on the calendar. Train a second person to review routine documents, so the senior person sees only the hard ones. Treat review as part of the job, not an interruption to it.

What to do this week

1. Choose the one document type where AI drafts have grown most.
2. Write a one-page checklist of what a finished version must contain.
3. Ask drafters to mark the source of every figure before review.
4. Track drafts, review time and send-backs for two weeks.
5. Name a second reviewer for routine work.

Checking AI work also costs money when the AI runs on its own. For that side, see Why an AI agent costs more with every turn. For a worked example of checking an AI draft line by line, see Check every reference before the client sees it.

Emeri did not run these studies. The figures come from the sources linked above.