ena pragma
·13 min read·ai-agents · ai-verification · ai-adoption · research-method

Personal AI speed is not company speed: few rules, checked on real work

A founder's talk says one person can get very fast with AI while a company does not. The 2026 data agree on the pattern. Here is what it means for anyone running agents at work.

Contents

The short version. A talk posted on November 3, 2026 by Mitch Troyanovsky, a co-founder of the accounting-agent company Basis, is titled "Why AI Makes You 100x Faster but Not Your Company." His point: people who use AI every day get much faster, and the average company does not. He names two problems. The first is a competency gap: agents built like coding tools "outsource the complexity to the user," so each person has to learn how to run them. The second is a process gap: a company needs its agents to be reliable, consistent and trustworthy, and he says that has to be set from the top. The pattern he describes shows up in the 2026 data, at smaller numbers. In a large survey of executives, AI "raised annual labor productivity growth by 0.2% over the last 3 years." Our reading of his argument, and of the data around it, is this: personal speed does not become company speed until the agent's written opinions are simple for the person using it and are checked against real work. We also saw a version of his warning in our own work, and that part is below. One caution first: he co-founded Basis, which raised $100 million at a $1.15 billion valuation and sells the kind of agent he argues for. We found no study that tests several of his claims, and we list them.

#What he said, and what holds

We read the talk from its YouTube auto-captions (audio and captions only; we did not see his slides). His argument, in our words: companies already know how to scale people, by trusting that the people at every level will do the work in a way that gives good results, consistently. An agent that a company can use the same way needs a written set of opinions (what to do, and just as important what never to do), and a way to check that the agent follows them.

The 2026 sources back the main shape of it:

  • Individuals report gains; firms see less. The NBER survey of nearly 9,000 CFOs, CEOs and senior executives says AI "raised annual labor productivity growth by 0.2% over the last 3 years." The abstract does not say whether 0.2% is per year or cumulative, so we quote it as written. The executives also expect "a substantially larger 0.6% per annum boost over the next 3 years." These are executives' reports, and a second NBER paper says "perceived productivity gains are larger than measured productivity gains." McKinsey's State of AI, August 25, 2026, finds that about 6% of respondents are AI high performers (at least 5% of EBIT, which is earnings before interest and taxes, attributed to AI and impact described as significant), a share it says has stayed flat, while "Eighty percent of respondents report that AI has improved their individual productivity." BCG, June 3, 2026: "But most organizations haven’t figured out how to convert the time into value." McKinsey and BCG sell advice, so read both as informed parties.
  • Reliability is the weak axis. A paper accepted at ICML 2026 finds "recent capability gains have only yielded small improvements in reliability." A separate April 2026 paper says "an agent that succeeds once may fail on a repeated execution of the same task." That supports his point that what a company needs is consistency, including an agent that is "consistently able to not do something."
  • Reading real work is a recognized practice. He says you can "look at the trajectories to decide if the behavior, the opinion that you have, is actually happening" (the trajectories are the recorded runs). Anthropic's engineering team says something close about its own agent tests (called evals): "Reading transcripts is how you verify that your eval is measuring what actually matters."
  • Trust grows with use. He describes agents that start narrow and earn wider trust. Anthropic's autonomy data fit the shape: "Among new users, roughly 20% of sessions use full auto-approve, which increases to over 40% as users gain experience." (Auto-approve means the tool acts without asking each time.) That is trust following the person's experience with the tool, which is not the same as a measured record of the agent.
  • Skill with the tool matters. Anthropic's Economic Index (March 24, 2026) found that people who had signed up at least six months earlier had a 10% higher success rate, and says "Facility with these platforms may be a key determinant of success that appears to scale with experience." It adds that early adopters could simply be more sophisticated, so this is an association.

He also says, about himself, that his own productivity has gone up a great deal. It is an aside, not a claim he builds on, and we do not grade it.

#What we found no study of

"We found no study" is a statement about our search (web and paper searches, and the pages we read, in November 2026), not proof that none exists.

His claimWhat we found
Process should come "from senior leadership down, not from any individual employee"We found no 2026 study that tests top-down against employee-set process on a firm's productivity. Surveys link gains to strategy and redesign (BCG: clear strategy lifts measurable business impact by 25 percentage points, against about 5 for better tools alone), but that is a self-reported correlation. A Census Bureau working paper finds use spreading both from the top and from workers.
"for the wide majority of knowledge workers, they have no idea how to do any of this"We found no study that measures it, meaning managing skills, memory and what the agent knows. BCG finds 36% feel adequately upskilled, and Gallup finds 54% of US employees said they had not used AI in their role in early 2026. Those show a skills and usage gap in general, not what the gap is made of.
You get "far larger" signal from checking behavior in recorded runs than from adding outcome-based evals (tests of an agent's final results). He calls outcome evals "really important" too.We found no study that tests it. Anthropic's guidance leans the other way on what to grade ("it’s often better to grade what the agent produced, not the path it took"), though a rule written about a behavior is a different thing from a fixed sequence of steps.
A coding agent's role-played copy of your production agent is a good testWe found no independent test. The nearest preprints disagree: a September 2026 paper finds "Direct self-report is weak (r = +0.04)" and that first-person framing "shifts reports in the flattering direction," while a March 2026 paper finds "frontier models exhibit privileged access to their own policies, outperforming peer models in predicting their own behavior."

That last row matters for the advice below. He says a copy of the agent is "much cheaper, much faster" and concedes it is worse than a true fork of the session. We found no measurement of the cost or the speed of either.

#How it applies if you run agents at work

These are our recommendations. Each rests on his argument and the sources above, and none is a measured result.

  1. Write few opinions, not everything. He says to take "the most important opinions" because "if you put everything, you'll be duplicating a lot of state." We found no study of how many rules is right, so treat this as design judgment. It fits one finding we can cite: researchers testing AI models that monitor coding-agent transcripts for dangerous actions found that three named models "miss these actions 2× to 30× more often when they occur after 800K tokens of benign activity than when they occur on their own." That is one dataset about monitors reading long transcripts, not about agents acting, but it is a reason to keep what an agent reads short.
  2. Decide what the agent consistently does not do. A list of things the agent never does alone (change a price, close a record, send to a customer) is easier to write down, easier to explain to a person, and easier to check than a list of things it should do well.
  3. Start narrow and widen trust on a record. His own test for how closely to look at a change is how much damage it could do and how competent its author is: "If the least competent person in your company is sending you a PR that's editing a core service, you're going to look at it." (A PR is a proposed code change.) Apply that to agents. The record you need is a count: how often the agent's work was undone, and how often it did something on the never list. A gate that is nearly always approved is a weak gate: Anthropic reports that "in practice users accept 93% of them anyway," and that for its automated safety classifier "The 17% false-negative rate on real overeager actions is the honest number" (a false negative is a miss; the figure is on 52 such actions, so about 9 misses, our arithmetic).
  4. Have someone other than the author read real sessions, for whether the opinions show up. The author knows what the rule meant and will read the agent's work kindly. Pick a few real sessions a week and ask one question per rule: did it show up? Start with checks you can run as a plain test (did this phrase appear, did this action happen without a yes before it). Our own judgment is to avoid an AI monitor for the first checks, given the long-transcript result above.
  5. Do not let a role-played copy clear a change. This is our reasoning, not a finding. A copy that runs inside another agent's wrapper has no real tools or errors, and the model knows it is a test. A role-played run is worth doing to find problems. A failure there counts. A pass proves little. He pitches a role-played copy as a cheap way to iterate on ideas and does not say it clears a change; we go further and say a pass should not clear one. After it, put the change in front of one real agent for a few real uses before it goes to everyone.

#What we found in our own work

The talk's warning is that companies make the person carry the complexity. We saw a version of it in our own work. In a recent training session with a client team, a large share of the time went to explaining how the agent's memory works. A clear explanation did not remove the burden, because the person still had to learn the model.

What is meant to remove the burden is a start-of-session step that loads the current state on its own, so the person does not have to learn the model. We had not yet seen it work in a real session when this was written. We tested that step's mechanism in a throwaway setup, which shows it runs, not that it helps a person in practice. What would tell us it works: the state shows up at the start of real sessions with nobody typing anything, corrections get filed without the person asking for it, and teaching minutes fall in later sessions. What would tell us it does not: people re-explaining context at the start, or most filings needing the person to ask.

The second thing we found is the checking gap. We already check a few rules about specific commands and actions an agent must not take, against real tool calls. For judgment-level rules (what the agent asks first, how it uses memory), we found no standing check against real sessions. We write those rules in many places and have no count of how often they show up. That is the check we are working on, starting with plain tests on real session logs, read by someone other than the author. We are not promising a result or a date. If we publish numbers, we will include the method.

#What we did not verify

  • We watched no slides and read captions, which garble names and words; every quote from the talk is from auto-captions.
  • We found no independent evidence on whether Basis's customers trust it. Basis's own release says it works with about 30% of the top 25 accounting firms; Reuters reports "about seven of the top 25," which is close to 30% (7.5 of 25, our arithmetic), and both figures come from the company. The release also quotes an investor on "20-50% efficiencies across practices," a quote with no method.
  • The NBER, McKinsey and BCG figures are survey reports, and McKinsey and BCG sell advice. The arXiv papers are preprints, and not all are peer reviewed.
  • We have not measured whether our start-of-session step works in real sessions, and we have no number for how often our judgment-level rules show up.
  • One talk is one speaker's view, and he co-founded the company that sells what it argues for.

Related field notes: Agents Did Not Clear the Backlog. They Moved the Bottleneck. covers where agent work piles up once the writing is fast, and AWS says 75,000 AI agents fit in a Raspberry Pi. The number is a division, and the cost to watch is the cold restart, not the waiting shows the same habit of reading what a number measured.

#Sources

All read November 4, 2026, Central time. The talk first, then primary papers, then the rest.

Keep reading