A program of EffectiveSC
How We Use AIAs of November 14, 2026
Blog /
Agents
Three models got the same persona and the same three memories. Swapping GLM and Gemma moved compliance by 3.3 points. Swapping GLM and DeepSeek moved it by 0.7 points, inside the variation of reloading the same model.
Holding one persona and three remembered decisions fixed, swapping GLM-5.3-Flash and Gemma4-31B moved post-pressure compliance by 3.3 points, and swapping Gemma4-31B and DeepSeek-V4.1-Flash moved it by 2.5 points. Swapping GLM-5.3-Flash and DeepSeek-V4.1-Flash moved it by 0.7 points, inside the variation of reloading the same model.
Every model, in every condition, stated a strong or absolute commitment to the rules in 100% of runs. After pressure, compliance was 97.1% for Gemma4-31B, 94.5% for DeepSeek-V4.1-Flash and 93.8% for GLM-5.3-Flash. The same persona and memory raised Gemma4-31B's compliance by 5.8 points. Without that file, Gemma4-31B complied on 91.2% of pressured decisions, below GLM-5.3-Flash's 93.8%.
The same file did not significantly change compliance for GLM-5.3-Flash or DeepSeek-V4.1-Flash. Asking first would not have caught the swap. The commitment question came before the scenario, and the runs with the persona removed said the same thing. A line telling the agent its backend had changed, and to stick to the written identity, did not significantly change compliance. Another model scored which option was picked. This is one persona and one multiple-choice compliance test, not a company agent at work.
The smaller of the two significant swaps barely clears the authors' line, and they do not adjust for how many pairs they compared. GLM and DeepSeek, from different families, were interchangeable here. Gemma was not.
The figures are on How well AI remembers and on How much will companies spend checking AI agents' work?.
Follow on X, LinkedIn, and email.
The headroom is not five to ten times as many daily users
On All-In, daily use in corporate America was put at 10% or 20%, with room for five to ten times as many people. Gallup's figure is 15% of US employees. If every industry reached technology's 42%, that would be about 2.8 times today's share, on this site's arithmetic. The larger gaps are in the share of the work, and in work that runs without a person.
Agent context is working memory only
Agents
Working memory without broader context is like asking two completely different people to get you a drink. They have never met you, and they have no idea who you are.
AI keeping us honest
Abstract
Sounding like a professional writer is a small fib. The larger ones are invented experience on one side, and jobs that are not as advertised on the other. AI makes both cheaper to write. It will not detect every lie, but it will do the homework and pick up cues from both sides.
Passing the steps is not finishing the job
Agents
Replit asked nine models to build a product, feature by feature. The strongest passed up to 98% of the individual tests and still delivered a working product in at most 65% of builds.
Throttle by friction
Abstract
Friction used to set how much work a company had to do. AI made asking, note-taking, and rewriting cheap for a person, and left the company with someone still signing the result. For most companies the new work is larger than the work removed. Tokens and that checking are the tax. The early gain lasts when it is spent on automation that clears ordinary work. A buyback, a dividend, or a project that fails to clear work leaves the human throttle where it was.
Daily use, two different questions
Gallup's daily share has climbed. The St. Louis Fed's every-workday share has stayed in a narrow band since 2024.
How often, and how much of the work
Fifteen percent used AI daily in May 2026. Two percent said it did all or most of their work.
Who sells it, and for how much
List prices and the commercial result each company has published. Seats, a run rate, and a price per resolution stay in that company's unit.
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.