LLMs and how to use them

The same model does different jobs in the same firm, depending on who uses it and on what work. Two things decide the job, and both can be placed on a chart. The chart’s third dimension is the answer.

Step 0

Build the matrix

Start with one person and one domain. Their skill sits on one of five levels, from Hubert and Stuart Dreyfus’s study of how people learn a skill.

Expert knows what to do without working it out
Proficient sees at a glance what matters, then works out what to do
Competent plans the work and owns the result
Advanced beginner starts to see situations the rules don’t mention
Novice follows the rules they were given

For work that carries money or reputation and that people do themselves, someone proficient can sign it. Below that dashed line, someone else has to.

Build the matrix one axis at a time. The views open at stage 3.

Step 1

Expertise belongs to a person in a domain

Now put real people on the front of the matrix. Take a firm of four people and four domains. Mary is an expert in property and a novice in securities. John is an expert in tax and securities and a novice in employment. Each piece of work below belongs to one domain, and each person stands at their own level in that domain. The fifth piece of work needs all four.

Every piece of work here carries money or reputation. A test can carry the work to an agent when its two kinds of mistake stay within the tolerances set with the criteria, below. Then the test decides when the agent’s work is done. Work people do themselves needs someone proficient to sign it: on the evaluation loop’s 100-point scale, we set what these stakes demand at 60. Below that line, where Amplify or delegate says the model amplifies you, here it teaches you, and someone who can sign checks the work.

Three terms. The firm’s experts write down what acceptable work looks like as criteria, kept in a knowledge map (Engineering the Context Window). An acceptance test is built from those criteria. And the model teaches from them too: the knowledge map is what goes into its context window when someone is learning.

Tolerances, set with the criteriaIn Deming’s terms an acceptance test is an operational definition, a criterion, a test against it and a yes-or-no decision. How many of each mistake the firm will accept is a separate choice, its risk appetite, made by whoever owns the risk. The defaults are those of Amplify or delegate.

Acceptable work rejected, up to this many in 100

Unacceptable work passed, up to this many in 100

Follow one person

Mary is an expert in property, competent in tax, a novice in securities and an advanced beginner in employment. Of the five pieces of work, Mary creates an agent for 1, deploys someone else’s for 1, is amplified on 1 and is taught on 2, and learns on a sample of 1 of the agent work.

Creates an agent on their own testDeploys an agent someone else createdDeploys someone else’s agent and learns on a sample of its workAmplifier, with the work staying with youTeacher, with someone who can sign checking the work

Pick a dot to see its numbers. The verdicts come from the same model as Amplify or delegate; all numbers are illustrative.

Mary, novice in securities, reviewing a share issue against the articles and the shareholders’ agreementTeacherAny test anyone here could write would reject too much acceptable work, even John’s, so the work stays with people. Mary is below the signing line, so Mary does the work with the model explaining, teaches it back to John, and John signs it. The model teaches from the firm’s criteria in its knowledge map. In the example job of Engineering the Context Window, the map puts 18 of the job’s 20 facts in the model’s window; a novice searching alone would give it 7. What nobody wrote down, John adds when Mary teaches the work back. Used without the criteria, the model hands out answers instead of practice. In a field study of school maths, students who had leaned on it that way did worse once it was taken away.

Mary’s own test63.7% acceptable work rejected, outside the tolerance1.1% unacceptable work passed

John’s test (the firm’s best at securities)15.2% acceptable work rejected, outside the tolerance0.7% unacceptable work passed

Tolerances, set with the criteria10% acceptable work rejected2% unacceptable work passedAt these stakes the cheapest pass mark keeps unacceptable work low, so it is acceptable work rejected that usually decides.

Create or deploy an agent

Amplifier

Teacher Mary, this work

Step 2

Keeping the knife sharp

Nobody stays on one row. People get better at work they do, and lose the knack of work they hand off. In step 1 the simple work goes to agents and the complex work stays with people, so the same person can be gaining in one domain and losing in another. Lisanne Bainbridge saw this in automated plants in 1983.

“A formerly experienced operator who has been monitoring an automated process may now be an inexperienced one.”Lisanne Bainbridge, “Ironies of automation”, Automatica 19 (1983), p. 776

The agents were created at the start, and their tests keep the quality they were written with. What wears a test down is the world moving. Its criteria are re-assessed on a cycle that matches how fast the domain moves, and only an expert in the domain can do it. While nobody does, the two curves of Amplify or delegate’s step 1 slide together. With nobody to move the pass mark, the test goes on rejecting the same acceptable work, so its checks look as they always did, while more unacceptable work gets through. The model teaches from the same criteria, so it goes on teaching the domain as it was.

Each person’s year is shared across the four domains, and the share of the work they do by hand sets the level they head for. The firm chooses how much of the agents’ work people still do themselves. Here it is eight years on.

Of the work agents do, how much do people still do themselves?

Year

Unacceptable work let through by the agents0.4% property
29.8% tax

Domains with an expert2 of 4

Criteria in date8 of 16

Criteria drifting8

Property. The agent runs on Mary’s test, created at the start. Nobody here is an expert any more, so its criteria are 5 years past re-assessment. It lets 0.4% of unacceptable work through, still within the tolerance of 2 in 100, and still rejects 2.3% of acceptable work, so its checks look as they always did. The learners here are taught from the same out-of-date criteria. Mary has gone from an expert to proficient. John, Jane and Pat, who do 1 in 10 of the agent’s work by hand, have not moved up a level.
Tax. The agent runs on John’s test, created at the start. Nobody here is an expert any more, so its criteria are 7 years past re-assessment. It lets 29.8% of unacceptable work through, against a tolerance of 2 in 100, and still rejects 8.6% of acceptable work, so its checks look as they always did. The learners here are taught from the same out-of-date criteria. Mary has gone from competent to an advanced beginner. John has gone from an expert to proficient. Jane and Pat, who do 1 in 10 of the agent’s work by hand, have not moved up a level.
Securities. The work stays with people. Mary has gone from a novice to proficient by doing the work. Jane has gone from an advanced beginner to proficient by doing the work. Pat has gone from proficient to an expert by doing the work.
Employment. The work stays with people. Mary has gone from an advanced beginner to proficient by doing the work. John has gone from a novice to proficient by doing the work. Pat has gone from competent to an expert by doing the work.

Illustrative. Someone who spends 30% of their year doing a domain’s work by hand goes from novice to expert in about nine years, close to the ten years of practice Ericsson and colleagues found behind expert performance. An expert who hands nine-tenths of their main domain to agents drops below expert in about four; Arthur and colleagues found skill loss growing with time out of use, faster for thinking tasks than physical ones. Criteria are re-assessed in quarters: tax every quarter, securities and employment every two, property every year. Skill changes over years. How fast criteria go stale is our assumption, and so is its effect on teaching, which is described here but not counted.

The verdicts: the model behind Amplify or delegate (/approach/amplify-or-delegate), with two levels added between its novice, competent and expert. Signing: the evaluation loop’s rule that you can sign off when your skill meets what the stakes demand.
Operational definitions: W. E. Deming, Out of the Crisis (1986). Tolerances are a separate choice, the firm’s risk appetite. The line from Shewhart’s statistical control as traced by Z. S. Chin, M. Chiodo, D. Müller and C. Snell, “Reframing AI loss of control” (arXiv 2606.12442, 2026).
H. L. Dreyfus and S. E. Dreyfus, Mind over Machine (1986); S. E. Dreyfus, “The five-stage model of adult skill acquisition”, Bulletin of Science, Technology & Society 24 (3), 2004, pp. 177–181.
L. Bainbridge, “Ironies of automation”, Automatica 19 (6), 1983, pp. 775–779.
Teach-back: G. Pask, Conversation Theory (1975), as set out in A recursive intelligence-amplifier. The knowledge map and the 18 and 7 of 20 facts: Engineering the Context Window.
H. Bastani, O. Bastani, A. Sungu, H. Ge, Ö. Kabakcı and R. Mariman, “Generative AI without guardrails can harm learning: evidence from high school mathematics”, PNAS 122 (2025).
K. A. Ericsson, R. T. Krampe and C. Tesch-Römer, “The role of deliberate practice in the acquisition of expert performance”, Psychological Review 100 (3), 1993, pp. 363–406.
W. Arthur Jr., W. Bennett Jr., P. L. Stanush and T. L. McNelly, “Factors that influence skill decay and retention”, Human Performance 11 (1), 1998, pp. 57–101.

Talk to us about your practice