Agentic AI Adoption Is Rising — and So Is the Cleanup Work

Agentic AI gets work done faster, but the time people spend checking and fixing its output is growing too.

Companies are rolling AI agents (programs that carry out multi-step tasks for you on their own) into everyday work faster than before. Agentic AI means the same idea: software that can run multi-step work with less hand-holding. An Oct. 13 Guardian report and several surveys point to the same warning: if you adopt it without preparation, people end up fixing low-quality output and errors, and the time spent checking and rewriting can outweigh the time saved.

For developers, the usual questions are two: Are jobs shrinking? and Is the work getting harder? What the evidence shows so far is less “the profession vanished overnight” and more slower hiring and employment growth, with people who remain taking on more review and cleanup.

Adoption is up; full autonomy is still rare

Wharton and GBK Collective’s 2026 AI Adoption Report says about two-thirds (67%) of departments surveyed use AI agents in some form.
Only 17% let an agent finish a task without a person approving each step. Most still approve step by step, or take drafts and suggestions and keep the final call human.
In a January 2025 Gartner poll of 3,412 webinar attendees, 19% said their organization had made significant agentic AI investments and 42% conservative ones. Many of the rest were waiting or unsure.
In June 2025, Gartner forecast that more than 40% of agentic AI projects could be canceled by the end of 2027 because of rising costs, unclear business value, or weak risk controls. That is a forecast, not a measured failure rate.
Gartner also warned about agent washing: labeling chatbots or RPA (scripted automation that repeats fixed tasks) as agents. It estimated only about 130 of the thousands of “agentic” vendors are the real thing.

Time saved comes back as checking and cleanup

Glean’s Work AI Institute surveyed 6,000 full-time digital workers in the United States, the United Kingdom and Australia between December 2025 and January 2026 for its Work AI Index. The numbers show the paradox.

87% of digital workers use AI at work and say automation alone saves them about 11 hours a week.
Only 13% say their company’s performance has improved significantly.
In between, people spend an average of 6.4 hours a week on botsitting: feeding AI context, checking outputs, debugging errors, cleaning up AI-made work, and switching between tools. The report calls this the hidden labor that makes AI usable.
When checking becomes overwhelming, people skip it. 69% of AI users admitted to botshitting — shipping AI work they had not reviewed or did not fully understand. 41% said they had delivered work they could not explain if asked.
According to CIO’s coverage, Glean says people who use multiple AI agents are more likely to skip checks, because agents can run many steps at once and the mistake may show up only later.

An Adaptavist survey of 2,500 knowledge workers, as reported by the Guardian and CIO, found 42% spend more time verifying AI output than they save using it.

More code, harder human review

In engineering teams, that cleanup shows up as a code-review bottleneck: the volume of agent-written changes ready for production outruns the time people have to read and judge them.

A study by Harvard researchers Fiona Chen and James Stratton, using Jellyfish engineering analytics across more than 700 software firms (2021 through March 2026) and reported by Ars Technica in October 2026, found that after AI coding agents were introduced, on average:

Lines of code rose about 30%, commits about 20%, and pull requests (PRs — change requests before merging into the main line) about 23%.
Measures of finished software output, such as issue and epic completion rates, did not rise in a clear, statistically significant way.
Average review time from PR open to merge rose 49%. The share of PRs with change requests nearly doubled, and comments per PR rose about 35%.
The share of workers doing code review rose about 14%. The researchers said they could not clearly attribute employment cuts to AI in this data. Faster coding shifted work toward human verification.
By March 2026 about 80% of measured firms used some AI code review, but AI accounted for only about 23% of review comments and about 11% of PRs. Humans still did most of it.

Headcount cuts make the bottleneck worse. Challenger, Gray & Christmas’s June report counted 101,743 U.S. layoff announcements citing AI in January–June 2026, about 23% of all announced cuts. That counts the reasons companies gave, which may differ from the real causes. Layoffs are not proof of AI productivity. A University of Pittsburgh study cited by the Guardian concluded that productivity did not rise after firms invested in AI and cut staff.

Meta shows the tension. Reuters reported that an internal plan called Project OT explored cutting some teams by as much as 60% and replacing daily work with agents, then that a further wave of cuts after the May restructuring was dropped. Internal posts cited in the reporting said code changes were up 220% while feature improvements reaching users rose only 36%, and time spent fighting technical and security incidents was up as much as 70%. Generation grew; the people and time to read the code and stop incidents did not keep up.

Faros’s 2026 engineering report (vendor data from about 22,000 developers on its platform) said median review time rose 5x and PRs merged with no review rose 31%. That is vendor data, not an independent study, but the direction — more generation, tighter human review capacity — matches the Jellyfish findings and Meta coverage.

What rushed rollouts cost Meta and Microsoft

The Guardian (Oct. 13) gathered cases of companies that turned AI on fast, then limited use or increased training because of errors, low-quality work, security risk and cost.

Meta: The Project OT plan and the code and incident figures above rest on Reuters (Aug. 26, 2026) and Ars Technica. Internal posts also warned that AI agents took “large-scale, disruptive actions that humans are unlikely to execute,” and that major technical and security incidents rose 40% year over year. Meta declined to comment on those internal figures. Coverage says Mark Zuckerberg told a July town hall that agentic progress had not accelerated as expected.
Microsoft: According to the Guardian, Kathleen Hogan said that deploying tools, training people and driving usage was not enough to change how work gets done; the process itself had to be redesigned. Ryan Roslansky publicly warned against a “doom loop” in which AI writes a document, AI summarizes it, and no new thinking remains (Guardian only; original not checked).

Are jobs shrinking, and is the work getting harder?

Jobs. A March 2026 Federal Reserve working paper finds that after ChatGPT launched in November 2022, employment growth for coders (occupations heavy on programming) slowed sharply. After controlling for industry shocks, annual growth was about 3 percentage points lower than before. The same paper says coder employment is still growing. That is closer to “not growing as fast as before” than to “jobs disappeared.”

Other research presented at NBER on GitHub Copilot adoption found signals of about a 36% drop in technical hiring at adopting firms, with the hit concentrated in fewer new hires rather than mass layoffs of incumbents. The Jellyfish study did not find a clear employment drop right after agents in its firm sample.

Challenger’s count of layoffs that cite AI (101,743 announcements through June) and Meta’s restructuring sit alongside that hiring slowdown and reshuffling. The scopes differ, so it is too early to collapse everything into “AI replaced developer jobs.”

Intensity. Here the evidence is clearer. 6.4 hours of botsitting, review time up 49%, Meta’s incident cleanup up 70% point the same way. Time writing code can fall while time reading, fixing and owning agent-made changes rises. If you cut headcount or freeze hiring while you scale generation, the remaining developers absorb the review and cleanup. For developers today, AI adoption looks less like “your seat is gone” and more like “the same team must verify more changes, and the hiring door may be narrower than before.”

What we could not confirm, or should not overclaim

“Agentic AI adoption is surging everywhere” matches department surveys (Wharton 67%) and investment intent (Gartner), not a measured worldwide production deployment rate.
Meta’s +40% incidents, +70% cleanup time and +220% code-change figures come from internal posts Reuters reviewed; Meta has not officially confirmed them.
Jellyfish found little evidence that AI cut employment. Challenger’s AI layoff count, Meta’s restructuring and the Fed’s coder-growth slowdown are different datasets, not the same finding repeated.
The Fed’s “about 3 percentage points” is a slowdown in the growth rate, not a 3% drop in the headcount total. The Copilot hiring −36% figure is limited to a specific adopter sample and contract timing.
Gartner’s “40% canceled by 2027” is a forecast.
Faros’s 5x review time and +31% unreviewed PRs are vendor data.
We could not open the Guardian article itself. The Microsoft executives’ remarks and the University of Pittsburgh study come only from the Guardian’s reporting, and we could not check them against the originals. The Challenger figure comes from the primary report, and the Adaptavist figure from survey coverage.

What could change for developers

What follows is our estimate and advice based on the reports. It does not look like every developer job is vanishing overnight. Hiring may get tougher, and day-to-day work may lean more on review and cleanup.

Start agent permissions small. Default-deny hard-to-undo actions in production — merge, deploy, customer-data access, outbound API calls — and open allowlists gradually.
Measure passed review, not time spent typing. If you judge success by lines of code or documents produced, you can get Meta-like growth in changes and in incidents. Track test pass rates, rollbacks, security issues and defects caught in review.
Do not cut review staff while you roll out agents. Studies show more lines and PRs come with more review time. Cut people first and you risk more unreviewed merges and more incident cleanup.
Log intermediate steps. When an agent chains tools, missing the failure point makes cleanup expensive. Keep action logs and step-by-step records you can replay.
Budget human review time. The 6.4-hour botsitting figure will not fit every team, but assuming “AI will handle it” and cutting reviewers can create a double burden.
Writing alone is a thinner edge. As agents draft more, splitting requirements, verifying results and preventing incidents matter more — especially when hiring slows.
Do not trust the word “agent.” Some products are rebranded chatbots or RPA. Ask whether they set goals, choose tools and keep state, or only answer a single prompt.

What to watch next

Follow-up Wharton and Glean numbers on autonomous agent use and botsitting time, whether agentic project cancellations rise as Gartner forecast, and how large companies put agent permissions, audit logs and role-specific training into formal policy.