The Agents in the Room: Work Before and After AI

3 Nov 2026 · 8 min read

AI agents are no longer a developer's toy. In 2026 they draft contracts, reconcile ledgers, triage tickets and open pull requests, often in the same company, on the same afternoon.

That shift is real, and so is the bill that comes with it. The productivity story everyone tells is only half the story. The other half is rework, review fatigue and burnout, and it is showing up in the data.

This post looks at both halves: what changed between the pre-AI and post-AI workplace, where agents genuinely help, where they quietly create more work, and what teams can do to keep the upside without burning people out.

Before and after: what actually changed

The biggest change is not speed. It is where human effort sits. Before AI, most effort went into producing the first draft. Now the draft is nearly free, and effort has moved to specifying, reviewing and owning the result.

Dimension	Pre-AI era	Post-AI era
The first draft	Slow, manual, the bulk of the work	Seconds to generate; rarely the bottleneck
Where expertise goes	Writing, building, calculating	Framing the task, checking the output, deciding
Coding	Write, search Stack Overflow, debug	Prompt, delegate to an agent, review the diff
Legal	Associates draft and redline from precedent	Agents first-pass contracts; lawyers verify and negotiate
Finance	Analysts build models and close books by hand	Agents reconcile, flag anomalies, draft commentary
Quality risk	Errors are slow and visible	Errors are fast, fluent and easy to miss
Who pays for mistakes	The author	Often the reviewer downstream
Pace expectations	Set by human throughput	Set by what the tool could do in theory
Junior roles	Learn by doing the grunt work	The grunt work is automated; learning paths are unclear

Read the last four rows carefully. That is where most of the hidden cost lives.

Engineering: from writing code to reviewing it

Software teams adopted first, and the numbers show it. In Stack Overflow's 2025 survey of 49,000+ developers, 84% said they use or plan to use AI tools, up from 76% a year earlier (Stack Overflow). The tools have since moved from autocomplete to agents that read a repo, plan a change, run tests and open a pull request.

Adoption and trust moved in opposite directions, though. The same survey found 46% of developers distrust AI output, up from 31%. The most common frustration, cited by 66%, was code that is almost right but not quite; 45% said debugging AI-generated code eats too much time (Tech.co).

The most sobering data point is METR's 2025 randomized trial. Sixteen experienced open-source developers worked 246 real tasks in codebases they knew well. With AI allowed, they took 19% longer, yet afterwards believed they had been about 20% faster (The Register). Time shifted from coding to prompting, waiting and reviewing, and developers accepted under 44% of generated code (Business Insider).

That study is narrow, and the tools have improved a lot since. But its lesson holds: the feeling of speed and actual delivery are different measurements. Agents shine on greenfield work, boilerplate, tests, migrations and unfamiliar codebases. They struggle with deep context, subtle invariants and large legacy systems, which is where senior engineers spend most of their time.

Beyond code: legal, finance and the rest of the business

The newer story is that agents have left the engineering floor. Contracts, ledgers, policies and spreadsheets are text and structure, which is exactly what language models handle well.

Legal. Thomson Reuters' 2026 AI in Professional Services report, based on 1,500+ professionals, found organization-wide AI adoption jumped from 22% to 40%. Agentic AI is earlier: about 15% of organizations use it, with law firms at 16% and corporate legal departments at 18% (Thomson Reuters, ACEDS). Typical uses: first-pass contract review, clause comparison, due-diligence summaries and research memos.

Finance. Financial services is moving faster. Cambridge's 2026 Global AI in Financial Services report puts agentic AI in active adoption at 52% of industry respondents, with fintechs ahead of incumbents at 57% versus 45% (Cambridge Judge). A Cloud Security Alliance survey found 62% of institutions deploying agents, and over a third granting them conditional or higher autonomy (CSA). Inside the CFO's office, agents handle reconciliations, variance commentary, close checklists and FP&A scenarios.

HR, ops, marketing and support. The pattern repeats: drafting job descriptions and policies, summarizing tickets, building reports, answering internal questions from a knowledge base.

The catch is the same everywhere. A wrong clause or a mis-booked journal entry costs more than a failing unit test, and non-technical teams rarely have the equivalent of code review, CI and version control. Deloitte Australia learned this publicly when it apologized for AI-generated errors in a A$440,000 government report (Philstar/The Conversation).

The upside is real

None of this means AI is a net negative. Used well, agents deliver gains that would have sounded absurd five years ago.

The blank page is gone. Even METR's participants kept using AI after the study, describing it as closer to editing an essay than facing an empty page (CO/AI). Lower activation energy matters for morale.
Small teams punch above their weight. A solo founder or a five-person startup can now ship, research, draft legal templates and run finance ops that once needed specialists. Lenny's 2026 survey found founders and small-company employees report the highest satisfaction (Podwise).
Drudgery gets offloaded. In UKG's survey of 8,200 front-line workers, AI users reported lower burnout (41%) than non-users (54%) (Newsweek). Removing tedious tasks works when it actually removes them.
Breadth expands. Engineers write passable SQL, finance analysts script automations, lawyers prototype workflows. The line between technical and non-technical work is blurring.
Agents help people who feel helped. Among developers using agents at work, 69% agree productivity is improving (LeadDev).

The hidden costs: rework, workslop and burnout

The cost of AI rarely shows up where the work is produced. It shows up one step downstream, on someone else's desk.

Workslop: effort moved, not saved

Stanford's Social Media Lab and BetterUp Labs coined "workslop" for AI output that looks finished but doesn't actually move the task forward. In their survey of 1,150 US desk workers, about 41% had received some in the past month, and each instance took nearly two hours to untangle (Techdirt, CO/AI). The researchers estimate that as roughly $186 per worker per month, or more than $9 million a year for a 10,000-person company (Techdirt via PieFed).

The social cost is worse. About half of respondents saw colleagues who send workslop as less capable and reliable, and 42% trusted them less (Shelly Palmer).

Rework eats the savings

Workday's 2026 research found employees lose about 37% of the time AI saves them to correcting, rewriting or clarifying low-quality output (Larridin). Global research cited by The Conversation found 66% of AI users had relied on output without evaluating it, and 55% had passed off AI material as their own (Philstar/The Conversation). Hidden use makes errors harder to catch.

Burnout: the pace ratchet

Lenny Rachitsky and Noam Segal's 2026 survey of nearly 6,000 tech workers found 55.7% experiencing significant burnout, up about 11 points from 44.7% a year earlier (36Kr). The top worry was not losing their job to AI. It was being expected to do more for the same pay (StationLM). A February 2026 ResumeTemplates survey similarly found 3 in 10 workers say AI increased their workload (PR Newswire via MEXC).

The mechanism is simple. Leaders see a demo, assume a 10x gain, and raise targets. Workers absorb the gap between demo and reality with longer hours. Review work is cognitively heavy, poorly recognized and never-ending, because agents generate faster than humans can verify.

Deskilling and the junior gap

The tasks agents automate first are the ones juniors used to learn on: first drafts, simple bugs, basic reconciliations. Seniors can review AI output because they once did the work by hand. If the next generation never does, who reviews the agents in ten years?

Getting the upside without the burnout

The teams that win with agents treat them like a powerful but junior colleague: fast, tireless, and in need of supervision.

Own what you ship. If your name is on it, you have read it, tested it and can defend it. AI-assisted is fine; AI-unreviewed is not.
Bring engineering discipline to every function. Version control for contracts, checklists for close, second-reader review for anything external. Legal and finance need their own equivalent of CI.
Measure outcomes, not usage. Track cycle time, rework, defect and correction rates. Only 18% of professional-services firms track AI's business value at all (Thomson Reuters). High adoption with rising rework is a cost, not progress.
Don't raise targets on faith. Re-baseline workloads only after you have measured real gains for your team, not a vendor's demo.
Match the agent to the task. Delegate boilerplate, first drafts and well-specified work. Keep humans on ambiguous, high-context or high-stakes decisions.
Count review as real work. Budget time for it, recognize it in performance reviews, and stop treating it as invisible overhead.
Be open about AI use. Normalize saying how a draft was made, so reviewers know what to check.
Protect the learning path. Give juniors tasks they do by hand first, then with AI, so they build the judgment to supervise agents later.
Invest in managers. Lenny's survey found manager effectiveness was the strongest single predictor of burnout, stronger than role, company size or attitude to AI (36Kr).

The bottom line

The pre-AI workplace was limited by how fast people could produce. The post-AI workplace is limited by how well people can judge. Agents have made output cheap; judgment, context and accountability are now the scarce resources.

The organizations that thrive will not be the ones that adopt the most AI. They will be the ones that redesign work around that new bottleneck: clear ownership, honest measurement, sustainable pace, and people who still know enough to tell when the agent is wrong.

At BrightSpark, that is the bet we are making: AI that amplifies capable people, not AI that buries them in cleanup.