This meeting could have been a prompt

If your company rolled out AI, the adoption numbers probably look great. They do almost everywhere. Weekly use at work among US workers went from 28% to 39% in under two years ^[1], and people really are faster at individual tasks. Support agents with an AI assistant resolved 15% more tickets an hour ^[2].

The return is harder to find. When researchers in Denmark matched AI use to payroll records for 25,000 workers, they found no effect on earnings or hours bigger than 2% ^[3]. Most people said the time they saved just went into other work.

So where did it go? I think most of it went into coordination.

Fred Brooks spotted the problem in 1975: adding people to a late software project makes it later ^[4]. Everyone who has to stay in sync with someone else adds a line of communication, and those lines multiply much faster than the team grows. Meetings, status updates, tickets, decks, and the email about the deck are how people keep them working. They aren't the work itself.

And that's exactly where AI went: into the inbox, the slide tool, the meeting and the ticket queue. A six-month trial across 66 companies measured the result ^[5].

Why didn't meetings budge? The researchers' answer: changing how meetings work "requires coordinating with colleagues and agreeing on new norms." One person's assistant can't do that.

Everyone drives faster between stops now. But the stops are still there, and because the paperwork of coordination got so cheap to produce, there are more of them.

You've been in this meeting. Thirty minutes, two decisions. Ten minutes later the notes land: two pages an AI wrote from the recording, longer than anything a person would have bothered to write. Somewhere in the middle there's an action item with your name on it that you don't remember agreeing to, because the AI "heard" it. So you read all of it, just in case.

Writing got cheap. Reading didn't. The evidence on how much that costs is still mixed. That same trial found people read fewer emails. But in a large study of code written by AI, 61% of it was never reviewed by a person, against 35% of code people wrote ^[6]. When there's too much to read, people stop reading.

The artifacts of coordination got automated. The coordination didn't.

A model on its own doesn't know your business, remember last week, or touch your systems. Everything that makes it useful has to be built around it: memory, knowledge, tools and access. When a company just hands out access to a model, every employee ends up building that part for themselves, a hundred different ways.

Most people aren't experts at this, and they shouldn't have to be. So the quality varies wildly, and the failures are quiet. A homemade memory serves up a fact that stopped being true last month. The assistant fills a gap with something that sounds right. It updates a system based on something nobody checked. In one test, memory systems lost 18 to 30 points once they were penalized for leaning on outdated facts ^[7].

And people usually think it went great. In a randomized trial, experienced developers using AI were 19% slower and believed they'd been 20% faster ^[8]. Agents often report success on work they failed ^[9]. A convincing answer feels like a return.

I built one of these for myself, which is exactly why I don't think everyone should. My agent has curated knowledge, memory, tools, hard limits and a log of everything it does. Its knowledge grew from 35 files in June to 112 in September ^[10], and on questions it had never been tuned on, it went from finding the right answer 65% of the time to every time ^[11]. Getting there took a lot of deliberate tuning, and most of the gain came from fixing how it searched, not from adding more. A company should build that once, for everyone. Not buy fifteen products from a vendor with a nice slide and a slick sales pitch.

Picture an engineer chasing a failing simulation. They read the ticket, open the code, search the docs, pull the logs, ask another team what changed, rerun it, update the ticket, write it up, and then sit in a meeting to tell everyone. An AI assistant can help with each of those steps, but the engineer is still the one carrying information from system to system.

Now flip it. The engineer asks why the simulation is failing. An agent that already knows the project checks the ticket, the code, the docs and the logs, reruns what it's allowed to run, and comes back with what it found. The engineer decides what to do. Once they approve, the agent updates the ticket and saves what it learned for next time.

Documentation stops being a separate chore. Today it's the step that lags, or gets dropped when the next urgent thing shows up. When an agent does the work, the record of the work comes out of it for free.

None of this means turning an agent loose. Agents still fail a lot of long jobs ^[12], so the connection between the agent and your systems needs limits. In one security test, just restricting which tools an agent could reach cut successful attacks from 58% to 7%, and it still did its job ^[13].

Give the agent to the job, not the person. The support lead's agent knows the queue, the escalation history and the customers who always call on Fridays. The finance agent knows every step of closing the month and every exception from the last two years. When the person in the role changes, the agent stays.

You know Joe. He built the spreadsheet everyone looks at, and he's the only one who really knows how it works. When Joe takes the offer from ACME, that knowledge walks out the door with him. With an agent that belongs to the role, it stays with the company, and whoever steps in inherits it on day one.

"Hey, how's the billing migration going?" Today that question interrupts a lead, who interrupts an engineer, who checks the ticket, and the answer comes back an hour later or in a meeting. When agents are doing the work, the answer is already in the record.

That isn't surveillance. The agent isn't watching people. It's reading the work that already exists: the tickets, the results, the changes. Managers get a real view of what their team delivered, and one of the most common questions in any company stops costing anyone an interruption.

Most AI budgets are spent per seat, which buys faster individuals. Agents are worth building per workflow, and they should be measured by what gets done. Ask how an AI rollout is going and you'll usually hear about seats, prompts, tokens and spend. That's the cost. An agent doing the work leaves a record of the work.

That's one real task from my own agent ^[14]. Cost per result is a number a license dashboard can't give you, because it never sees a result.

1. 1
   Build the foundation once.

   Identity, permissions, shared knowledge, memory, governed access to data and tools, and a log of every action. One foundation the whole company runs on, not a stack of products that don't talk to each other.

2. 2
   Pick one workflow that hurts, and measure it first.

   Look for something repeated that crosses several systems and at least one handoff: investigating an outage, onboarding a customer, closing the month. How long does it take, how many people touch it, and how much of that time goes into just gathering information?

3. 3
   Provision an agent for each role in it, read-only at first.

   Let it look at every system the workflow touches before it can change anything. It will map how the work really happens faster than anyone could write it down, and you'll find out what it gets wrong while that's still harmless. More on that in Trust is a schedule, not a verdict.

4. 4
   Give the shared knowledge an owner.

   Decide who reviews what the agents learn before it becomes something every agent relies on. Knowledge nobody curates goes stale and starts to hurt ^[15].

5. 5
   Keep people deciding, and put the limits in code.

   A person approves anything that changes a system, and every action is logged. A policy document doesn't stop anything. Permissions do ^[13].

6. 6
   Measure the outcome, then widen it.

   Track the workflow against your step-two baseline: time to finish, handoffs, rework and cost per completed task. Check the agent's work independently, because agents grade themselves generously ^[9]. When the numbers hold, let it handle the routine cases on its own and move to the next workflow.

Most leaders are still asking how to give their people AI. That question has been answered, and the answer was faster email. The better question is what your company looks like when every role has an agent that's connected, remembers, and works within limits. What's a meeting for when nobody has to carry context into it? What does a manager need to ask a person, once the agent can answer the status question? Who approves what?

Stop asking how to fit AI into the way work gets done today. Start asking what work should look like now that agents exist.

I wrote this with the agent that helps me run this site. It searched this year's AI research on arXiv, a free public archive of research papers, and read the original of every study cited here. The numbers about my own agent come straight from its records.

1. ExtBick, Blandin and Deming, "The Rapid Adoption of Generative AI," NBER Working Paper 32966, revised Feb 2025, with the Federal Reserve Bank of St. Louis update "Does generative AI save time at work?" (Aug 2026). Nationally representative US survey. Weekly use at work 28.2% (Q3 2024) → 39.2% (Q2 2026); users save 5.4% of work hours, 1.4% across all workers. Self-reported, which the update itself calls inherently approximate. nber.org/papers/w32966
2. ExtBrynjolfsson, Li and Raymond, "Generative AI at Work," Quarterly Journal of Economics 140(2), 2025 (arXiv 2304.11771). Staggered rollout to 5,172 customer-support agents at one firm. +15% issues resolved per hour, about +30% for the least experienced, no gain and small quality declines for the most skilled. arxiv.org/abs/2304.11771
3. ExtHumlum and Vestergaard, "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI," NBER Working Paper 33777, revised Mar 2026 (formerly "Large Language Models, Small Labor Market Effects"). About 25,000 workers in 7,000 Danish workplaces, surveys linked to administrative records. Effects on earnings and hours ruled out above 2% pooled; 85% of users reallocated saved time to other tasks; AI created new work for 12% of users. Denmark only, two-year window, not peer reviewed. nber.org/papers/w33777
4. ExtBrooks, The Mythical Man-Month: Essays on Software Engineering, Addison-Wesley, 1975, chapter 2. The source of Brooks's Law and of the n(n−1)/2 count of pairwise communication paths. Theory, not measurement.
5. ExtDillon, Jaffe, Immorlica and Stanton, "Shifting Work Patterns with Generative AI" (arXiv 2504.11436, v4, Nov 2025). Six-month randomized field experiment, 66 firms, 7,137 workers. Email time −1.37 h/week on an 11.65 h baseline (−12%, significant); emails read down; meeting time +0.10 h on 5.22 h, not significant; document measures not significant. Authors are partly employed by the vendor of the tool tested and state they cannot measure productivity. arxiv.org/abs/2504.11436
6. ExtDuma et al., "These Aren't the Reviews You're Looking For: How Humans Review AI-Generated Pull Requests" (arXiv 2605.02273, May 2026). 33,596 agent-written pull requests in repositories with 100+ stars: 61.38% had no recorded review, against 34.52% of human pull requests in the same repositories. One dataset, GitHub only. arxiv.org/abs/2605.02273
7. Ext"From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents" (arXiv 2604.20006, 2026). Six memory systems and four models; scores drop 17.8–29.5 points once reliance on outdated memories is penalized. Synthetic data, model-judged with 88.3% human agreement. arxiv.org/abs/2604.20006
8. ExtBecker et al. (METR), randomized controlled trial of early-2025 AI tools on experienced open-source developers (arXiv 2507.09089): 16 developers, 246 tasks, 19% slower; developers estimated afterward they had been 20% faster. Authors scope it to experts in mature codebases. arxiv.org/abs/2507.09089
9. ExtStudy of false success claims in agent trajectories (arXiv 2606.09863, 2026): 9,876 τ²-bench and 1,879 AppWorld trajectories. Agents claimed success on 13–79% of their failures depending on model family; with independent verification in the environment, 3%. Preprint. arxiv.org/abs/2606.09863
10. DataMy agent's knowledge files, counted monthly, Jun – Sep 2026: 35 → 53 → 92 → 112 files; individual facts 1,006 (July) → 2,477 → 3,295. That measures size, not usefulness.
11. DataMy agent's memory test: questions about its own systems, scored on whether the right fact comes back in the top five results. Held-out set of 26 questions: 0.654 on Aug 27, 2026 and 1.000 on Sep 24, 2026. Four held-out questions had their fact text edited to match; the other 22 scored 22 of 22. Most of the rise came from a Sep 6 change to how results are ranked, so scores either side of it aren't strictly comparable. It's a small test I built myself; the 95% band on 1.000 runs from 0.87 to 1.00.
12. ExtOSWorld 2.0 (arXiv 2606.29537, 2026): 108 long computer workflows with a median human time of about 1.6 hours. Best model 20.6% complete, 54.8% partial; the same model scores 83.5% on the short-task version. arxiv.org/abs/2606.29537
13. ExtDebenedetti et al., "AgentDojo," NeurIPS 2024 Datasets and Benchmarks (arXiv 2406.13352). 97 tasks, 629 security cases. With one model, targeted attack success 57.69% undefended vs 6.84% with a tool filter, benign utility 73.13%. The filter fails where the tools a task needs are enough to carry out the attack. arxiv.org/abs/2406.13352
14. DataOne traced task from my agent, Sep 24, 2026: a planning session, joined from its telemetry by session and request. 6 requests, 95 model calls, 89 tool calls (5 failed), 1 h 44 min, $7.10 at list price as the client estimates it, ending in one committed change.
15. ExtWang et al., "Not All Skills Help" (arXiv 2606.15390, 2026). AppWorld hardest split: one model 52.5% bare, 49.9% with an uncurated skill library, 66.4% with curated skills; over 90% of skills help on some tasks and hurt on others. Preprint. arxiv.org/abs/2606.15390