The Eglen Group — Field Dispatch On Working With Machines No. V — Summer 2026

The Eglen Group

A Short Course in AI Fluency

The Four
Disciplines

A man can generate a thousand words before breakfast. Fluency is knowing which of them are true, which to keep, and which to throw back in the sea. It is the discipline of making every interaction with AI effective, efficient, ethical, and safe. Anthropic Academy now offers a formal AI Fluency course at anthropic.com/learn. The habits, however, remain yours to build.

◎Effective

◎Efficient

◎Ethical

◎Safe

The Theory of the Iceberg

One‑Eighth Above the Water

The polished answer the machine hands you is the part that shows. It only carries weight because of the seven‑eighths beneath it — your judgment, your knowledge, your nerve. That submerged mass is the work. Done right, it makes your interactions effective, efficient, ethical, and safe. It has four parts.

The Four D’s

The Disciplines

The seven‑eighths beneath the water.

Delegation.

“Know which fights are your own.”

The framework: Setting goals and deciding whether, when, and how to engage with AI at all.

In the field: The first discipline is restraint. Hand the machine the rope-work and the drafting; keep the decisions that carry your name. Most failures are not bad prompts — they are work that should never have been delegated.

Chapter I — Delegation

The trouble with the machine is that it will do anything you ask. It does not get tired. It does not refuse. It does not tell you the task was a bad idea to begin with. That is your job. The first discipline is not about what the machine can do — it is about what you choose to give it. And that choice belongs entirely to you.

In enterprise work, most of the wreckage you will find around AI projects was not built from bad outputs. It was built from good outputs applied to the wrong problem. Someone handed the machine a decision that belonged to a human — a hiring judgment, a risk assessment, a strategic recommendation to a board — and when it came back clean and confident, they signed it. The first discipline asks: before you type a single word, is this yours to delegate at all?

There is a test worth running honestly before you begin. Ask yourself two things. First: does the value of this task live in the result, or in the act of doing it yourself? A strong analyst learns to write by writing badly first. A consultant builds judgment by wrestling with the hard case alone. Some work is valuable because of what it costs you. Second: if the machine is wrong by thirty percent, what breaks? Work through both questions without the answer you want. You will find that fewer things belong to the machine than you thought, and the ones that do, you can hand over with a clear conscience.

Field Notes — How to Decide

High-stakes, low-reversibility work stays in your hands. The machine handles volume; you handle consequence.
Delegate research, synthesis, drafting, formatting. Keep judgment, recommendation, accountability.
Set your success criteria before you delegate — so you know what to measure when the output returns.
Ask: if this output is wrong, who finds out and when? The further the blast radius, the less to delegate.
When unsure, trial it small. Delegate a reduced version and evaluate the output against your standard before committing at scale.

Questions for the Trail

What is the cost if this output is wrong by twenty percent — to my client, my team, my reputation?
Am I delegating this because the machine does it better, or because I want it off my desk?
Would I be comfortable explaining this delegation to a client who asked exactly how the work was done?

Description.

“Say the true thing, and say it plain.”

The framework: Effectively describing your goals so the model produces useful behaviors and outputs.

In the field: A prompt is prose. Vague in, vague out. State the goal, the reader, the shape of the thing you want, and what you do not want. Cut every word that does no work.

Chapter II — Description

A prompt is a letter to someone who knows everything and nothing. The machine has read more than you will read in a lifetime, and it does not know your client, your deadline, your audience, or what you mean by “professional but warm.” You have to tell it everything it needs and nothing it does not. The discipline of description is the discipline of being specific about a thing you want.

Most bad outputs are the machine’s honest answer to a question you did not quite ask. When you write vague in, you get fluent out — smooth, plausible, and aimed at a different target. The fix is rarely a better model. It is a better question. Professionals in every field spend years learning to ask the right question of the right source. That skill does not go away in the age of AI. It becomes more valuable.

Good description has four moving parts. The role: who the machine should be in this exchange — a senior partner, a plain-English editor, a skeptical CFO. The goal: what you need, not how you imagine it should be accomplished. The constraint: what you do not want, what is out of scope, what would make the output useless before it arrived. And the reader: who receives what the machine produces, and what they need to walk away with. Write all four, and you have most of a good prompt.

Field Notes — The Four Parts of a Good Prompt

Role: "You are a change management consultant advising a Fortune 500 HR team preparing for a major ERP rollout…"
Goal: Name the deliverable precisely. Not "write something about X" but "draft a one-page executive summary for a CHRO who has twelve minutes and no patience for jargon."
Constraint: Name what you do not want. "No bullet points. No passive voice. Do not use the words leverage, synergy, or best-in-class."
Reader: Describe who will read the output. The machine writes differently for a skeptic than for a convert.
When the output misses, fix the prompt before re-running. The machine did not misunderstand — you did not say enough.

Questions for the Trail

If someone read only my prompt — no context, no follow-up — could they produce roughly the right output without me?
Have I told the machine who it is, what I want, and what I do not want? All three?
Am I asking for a process, or a result? Ask for the result. Let the machine choose its own path there.

III

Discernment.

“Read the water before you trust it.”

The framework: Accurately assessing the usefulness of AI outputs and behaviors.

In the field: The machine is fluent and confident and sometimes wrong in the same breath. Edit it the way a cold-eyed editor reads a first draft — assume nothing, check the facts, trust the result only after it has earned it.

Chapter III — Discernment

The machine is never wrong in an obvious way. It does not misspell badly. It does not trail off. It writes the way a confident expert writes — which is the problem. Confidence is not correctness. Fluency is not accuracy. The discipline of discernment is reading the output the way a good editor reads a first draft: not to admire it, but to find where it breaks. And it will break somewhere.

There are failure patterns worth knowing by name. The machine hallucinates — it invents specific facts, names, citations, and statistics with the same smooth tone it uses for true things. It reflects your prompt back at you, which means your assumptions travel into its answer and come out dressed as conclusions. And it is calibrated on the average of everything it has ever read, which makes it mediocre on institutional nuance, edge cases, and anything that happened recently. A fluent answer about the wrong company is still wrong.

The test is simple, and it takes nerve to apply honestly. Read the output assuming it is wrong. Look for the claim you cannot verify. Find the number you have not checked. Notice when the reasoning sounds airtight but silently skips a step. This is not cynicism. A good skeptic and a good reader are the same person. You are applying the same discipline to the machine you would apply to a junior analyst who was very good but new. You would not sign their work without reading it. Do not sign this either.

Field Notes — Where Outputs Break

Verify specific claims: names, statistics, citations, dates. These are where hallucinations hide, dressed as fact.
Check the logic separately from the language. Fluent prose is fully capable of carrying broken reasoning end to end.
If the output agrees with everything you already believe, probe harder — the machine is often a mirror.
Use the description-discernment loop: when output is off, revise the prompt and re-run before accepting the gap.
For high-stakes deliverables, have a colleague review independently. The machine’s errors are often invisible to the person who prompted it.

Questions for the Trail

What is the single claim in this output that would hurt most if it turned out to be wrong?
Does the reasoning hold if I pull one core assumption out from under it?
Could I explain how the machine arrived at this conclusion, or am I trusting the finish instead of the structure?

Diligence.

“Sign your name to every word.”

The framework: Taking responsibility for what we do with AI and how we do it.

In the field: The byline is yours. Be honest about what the machine touched, be square about the risks, and own the outcome the way a captain owns the boat. Accountability does not transfer.

Chapter IV — Diligence

At the end of the day, the work is yours. Not the machine’s. Yours. It does not matter how much of the drafting was done by the model, how clean the output was, how little you changed. If your name is on it, you own it. That is not a burden to resent — it is the thing that makes the work mean something. You are not a human in the loop. You are the one who decides what the loop is for.

Diligence has three faces. Transparency: being honest with your clients, your team, and your organization about how AI was used and where the human hand came in. Limitation awareness: knowing where the model fails, where the data runs out, where the answer should not be trusted without a second look. And accountability: understanding that the machine is not a defense. You cannot say the AI told you so. You can only say you should have checked, and you will carry that.

In the change management and consulting work that defines this practice, diligence is the discipline that builds trust over time. Clients do not need to know every tool you reach for. They need to know that what you give them has been thought through, checked, and backed by a professional who stands behind it. The machine can help you do more. Diligence is how you ensure that it remains worth something — to you, and to the people who depend on your judgment.

Field Notes — What Diligence Looks Like

Know your organization’s AI use policy before you need it. Surprises there are never good ones.
Be transparent with clients when AI substantially shaped a deliverable — and have a clear answer about your quality process.
Do not put sensitive client data into a model without understanding where it goes and what retains it.
Know which tasks AI should not touch: final board recommendations, legally consequential decisions, work requiring lived human judgment at high stakes.
Ask: would I be comfortable if my client could see, in detail, exactly how this was made?

Questions for the Trail

Have I been honest — or could I be honest — with my client about AI’s role in this work?
Is there data in this workflow that does not belong there?
If this output turns out to be wrong, can I explain what I did to check it, and why I believed it?

Automation.

“Set the line, bait it well, and trust the current.”

The framework: AI executes specific tasks based on your instruction. You set the course; it runs the line.

In the field: Automation is where volume lives. What takes you three hours, it does in three minutes. The discipline does not loosen here — it tightens. A bad instruction run once produces one bad output. Run it ten thousand times, and the error compounds quietly.

Chapter V — Mode A: Automation

Automation is the simplest of the three modes, and the one where failure most often comes from overconfidence. You write the instruction once, clearly, and the machine runs it. The value is in the volume and the consistency — what you do a hundred times, it does a thousand times without complaint, without fatigue, and without varying the output in ways you did not intend.

But the discipline does not relax in automation. It concentrates. The work of the human here is entirely in the setup: the delegation judgment made before you begin, the sharp description written before you run, and the discernment applied to the early outputs before you commit to the batch. You check the first three results with more care, not less. The machine amplifies your instruction — good or bad — at scale.

In consulting and change management work, automation is most valuable where the task is genuinely repeatable and the risk of error is bounded. Summarizing stakeholder interviews across a large sample. Generating first-draft communications from a template. Reformatting data for a report. The trap is treating complex, contextual work as if it were automation — asking the machine to make judgment calls you have not made yourself, then letting it run.

Field Notes — Running It Right

Inspect the first three outputs of any automated run before committing to the batch. Do not skip this.
Define failure clearly before you begin. If you cannot say what wrong looks like, you will not recognize it when it arrives.
Automation amplifies the quality of your setup. The time spent on a precise instruction pays at scale.
Keep a human in the loop for any output that reaches a client or makes a consequential decision.

Questions for the Trail

Would I check every output if I ran this a thousand times? If not, what is my sampling strategy?
Have I defined what failure looks like in this task before I start?
Is this task genuinely repeatable, or am I calling it automation to avoid thinking it through?

Augmentation.

“A good thinking partner asks the right questions and knows when to be quiet.”

The framework: Humans and AI collaborate as thinking partners. Neither directs; both contribute.

In the field: You with the context and judgment; the machine with the breadth and speed. The discipline in augmentation is keeping your own perspective while using the machine’s — not letting its framing become yours before you have thought it through yourself.

Chapter V — Mode B: Augmentation

Augmentation is where most productive knowledge work with AI lives. You and the machine working together — you with the context, the judgment, the domain expertise; the machine with the breadth, the speed, the inexhaustible patience for drafting and revising and thinking aloud. Neither of you is in charge. Both of you are useful. The question is whether you lead the conversation or let it lead you.

The risk in augmentation is subtle. It is easy to let the machine’s framing become your framing. It offers you three good options and you choose one. But the best option might have been something it did not generate, and you did not think to ask because you were reacting rather than leading. Good augmentation means you bring your thinking to the machine, not just your questions. Go in with a position. Use the machine to stress it, refine it, break it if it is breakable.

In the consulting room, augmentation looks like drafting alongside the machine, asking it to find the objection you missed, to argue the opposite position, to write the paragraph the skeptical board member would write in response to yours. It compresses the distance between an idea and a testable form of it. That kind of efficiency accumulates. Not because the machine did the thinking, but because you did not have to do the slow parts alone.

Field Notes — Working Alongside

Go in with a point of view. Augmentation sharpens existing thinking; it rarely generates thinking from scratch.
Ask the machine to disagree with you, find the flaw, or argue the opposite. The useful friction is in the pushback.
Rotate the perspective: ask what the skeptical CFO would say, what the resistant employee would feel, what your hardest client would challenge.
Do not accept the machine’s framing of the problem. Name the problem yourself first, then bring it in.

Questions for the Trail

Did I go into this session with a position, or did I let the machine form one for me?
What did the machine say that surprised me — and did I test it, or accept it?
If I removed the machine’s contribution from this work, what would be left? Is that enough to stand on?

Agency.

“A well-rigged boat sails true without you at the helm. But you built it, and you answer for where it goes.”

The framework: You configure AI to carry out future tasks on your behalf, without watching every step.

In the field: Agency is the mode that requires the most trust and earns it slowly. Every result the system produces in your name is yours. The four disciplines do not loosen here — they are load-bearing.

Chapter V — Mode C: Agency

Agency is the mode where the four disciplines stop being habits and become architecture. You set up the system — the instructions, the guardrails, the decision points, the outputs — and then you step back. The machine acts in your name while you are not watching. Every result that comes from it is yours. The people who receive those results did not build the system. They trust the person whose name is on the work.

This is where the disciplines carry the weight they were built for. Delegation: which decisions can the system make on its own, and which must surface to you? Description: are the instructions precise enough to work correctly in cases you did not anticipate when you wrote them? Discernment: have you reviewed enough output across enough time to know what the system gets wrong, how often, and in what direction? Diligence: do you have a clear understanding of what happens when it fails, and how you will know before the client does?

A well-designed agentic workflow can do in a night what a team could not do in a week. A poorly designed one runs quietly wrong for months before anyone notices. The productive gains in agency are real, and so are the failures. The difference between them is almost always in the design and the monitoring, not in the running. The boat sails itself. You built it. You are responsible for what it finds.

Field Notes — Building for the Unattended Run

Document decision boundaries explicitly before you build. What can the system decide alone? What must it escalate?
Build in review checkpoints. Agency does not mean unmonitored — it means you review the right things at the right intervals.
Start with a small scope and earn trust from the outputs before expanding autonomy. Trust is demonstrated, not assumed.
Have a clear answer to: what happens when this fails, and how will I know within a day rather than a month?

Questions for the Trail

Have I documented exactly what this system can and cannot decide on my behalf?
Have I reviewed enough of its output to know its failure modes — not just its successes?
If this system runs for thirty days without my intervention, will the results reflect my judgment?

Effective.

“The work gets done well, or it does not count.”

The standard: Outputs are accurate, relevant, and useful to the person who needs them.

In the field: Effectiveness is the baseline. An AI-assisted workflow that produces impressive-looking outputs no one can use has failed at the first standard, regardless of how elegantly it was designed.

Chapter VI — Effective

Effectiveness is the baseline. Without it, none of the rest matters. An AI-assisted workflow that produces impressive-looking outputs no one can use has failed at the first standard, regardless of how elegantly it was designed or how quickly it ran. Effective means the work answered the actual question, served the actual reader, and could be acted upon without a follow-up conversation to explain what it meant.

The four disciplines drive effectiveness in sequence. Delegation ensures you are working on the right problem to begin with. Description ensures the machine has what it needs to produce the right output. Discernment ensures you catch the gap between what you asked for and what you got. Diligence ensures a final review by someone who cares whether it works. Remove any one of those links and the chain fails quietly, often long before anyone names it as failure.

In consulting, effectiveness has an additional test: the client does something with it. Not just nods and files it away. Does something. A strategy that sits in a drawer is not effective, regardless of its quality. An AI-assisted deliverable that gets acted upon is. Hold the work to the full standard, not just the surface one. The measure is what changes downstream, not what looked good in the room.

Field Notes — What Effectiveness Requires

Before you start: can you state clearly what success looks like for this output? If not, the problem is not yet defined.
After you finish: could the recipient act on this without asking you a clarifying question?
Effectiveness is not thoroughness. A five-page answer to a one-paragraph question is not more effective — it is less.
Test the output on a proxy reader before it goes out. Someone who represents the actual audience, without your context.

Questions for the Trail

Does this output answer the question that was actually asked, or the question I assumed was being asked?
Could the person who receives this use it without a follow-up conversation?
Is it thorough because thoroughness serves the goal, or because it was easier to include everything?

Efficient.

“More done — not just faster.”

The standard: AI amplifies human capacity without creating more waste than it saves.

In the field: Efficiency is not the same as speed. A workflow that produces a rough draft in two minutes and then requires forty-five minutes of repair is not efficient. It is redistributed labor.

Chapter VI — Efficient

Efficiency is not the same as speed. Speed is how fast the output arrives. Efficiency is the ratio of value produced to effort expended — and that denominator includes the time spent prompting, checking, correcting, and second-guessing. A workflow that produces a rough draft in two minutes and then requires forty-five minutes of repair is not efficient. It is redistributed labor with a confident face on it.

The test of efficient AI use is honest accounting. Where is the time actually going? If the machine handles the drafting but the review and revision cycle costs more than the drafting would have, the model is not saving time. It is moving it. Efficient workflows are ones where the machine’s contribution meaningfully exceeds the cost of the oversight it requires. That ratio is worth tracking, especially early in adopting a new workflow.

The deeper efficiency is cognitive, not temporal. A well-run augmentation session leaves you thinking more clearly about the problem, not less. It brings structure to vague thinking, surfaces objections you had not named, and compresses the distance between an idea and a testable form of it. That kind of efficiency accumulates in ways that pure speed does not. Hours saved on formatting and drafting are good. Thinking time freed up for harder problems is better.

Field Notes — Honest Accounting

Track where AI genuinely reduces your load versus where it just shifts the work to a different part of the process.
Optimize for output quality per hour of total effort, not for prompt speed alone.
If you are spending more time correcting than you would have spent doing, recalibrate the approach before the next task.
The most efficient use of AI is often the most boring one: repetitive, well-defined tasks where the machine adds volume and you add oversight.

Questions for the Trail

Is AI actually saving me time on this task, or just making a different part of it harder?
Could I achieve the same result with less prompting and checking if I invested more in the initial description?
Where is my real bottleneck — the generation or the review?

Ethical.

“Do it right by the people on the other end.”

The standard: AI collaboration respects people, data, privacy, and truth.

In the field: Ethics in AI work is not abstract. It lives in the specifics: whose data is in the prompt, who will be affected by the output, whether the work is honest about what it is and how it was made.

Chapter VI — Ethical

Ethics in AI work is not abstract. It lives in the specifics: whose data is in the prompt, who will be affected by the output, whether the work is honest about what it is and how it was made. The people on the other end of your AI-assisted deliverable — the employees whose roles the strategy touches, the clients whose trust you are working within — are real. The discipline of diligence exists precisely to keep them present in your thinking before the output goes out.

In change management and consulting, ethical AI use has particular weight. You are often working with sensitive information about organizations in transition: workforce data, performance concerns, strategic plans that are not yet public. The question is not only whether the machine can handle this information, but whether it should, and what your obligations are to the people the data represents. Those obligations do not change because the tool changed.

Ethical use also means being honest about what the machine did. Not as a confession, but as a professional standard. When AI substantially shaped a deliverable, you know what that means for how you stand behind it. Your client deserves to know what your quality process was. Honesty here is not vulnerability. It is the foundation of a durable working relationship, and it is the only kind worth building.

Field Notes — The People in the Work

Know whose data is in your prompt before you paste it. Sensitive client data requires explicit handling decisions.
Be honest with clients and colleagues about AI’s role. Have a clear, confident answer about what your review process was.
Consider the people downstream — not just the client, but the people the client’s decisions will affect.
Do not use AI to create a false impression: attributing machine-generated analysis to independent expert judgment you did not apply.

Questions for the Trail

Is there anyone who would be harmed if this output were wrong — and do they know AI was part of how it was made?
Am I being honest with my client about how this was produced?
Are there people whose interests are represented in this work who have not been considered?

Safe.

“Know where the ice ends and the open water begins.”

The standard: AI interactions do not create undue risk — to clients, organizations, or the people affected by the work.

In the field: Safety is not the absence of AI. It is the presence of limits you have thought through. A professional who uses AI carelessly is not more capable — they are more exposed.

Chapter VI — Safe

Safety is not the absence of AI. It is the presence of limits you have thought through. A professional who uses AI carelessly is not more capable — they are more exposed. A professional who uses it with clear boundaries and honest self-assessment is more capable than they would have been without it. The difference is not the tool. It is the discipline applied to the tool’s limits.

Safe AI use in professional work means knowing the edges. Knowing where the model’s reliable knowledge ends. Knowing which tasks are too consequential for the model’s error rate. Knowing when a human expert needs to verify something the machine touched. Knowing when to stop. The model will always give you an answer. The discipline is in recognizing the cases where an answer from the model is not sufficient evidence for the decision at hand.

Risk in AI-assisted work is often invisible until it is not. The output looks correct. The reasoning sounds sound. The client is satisfied. And then, later, the assumption that was wrong surfaces in a consequential place. Safe practice means building in checkpoints before that moment arrives — not because you expect failure, but because the professional who is never surprised by their own errors is the one worth trusting. That trust is built case by case, and it is lost fast.

Field Notes — Operating Within the Limits

Know your organization’s and your client’s AI policies before you use the tool. Not after.
Identify the highest-risk outputs in any engagement and apply your heaviest human review there, not everywhere equally.
When the stakes are high and the model’s confidence is also high, that is exactly the moment to slow down.
Never let AI make the final call on a decision that significantly affects people. The recommendation is yours to make, or to withhold.

Questions for the Trail

If this output turns out to be wrong, what is the worst plausible consequence — and have I reviewed it at that level of care?
Do I know the limits of the model’s reliable knowledge in this specific domain?
Have I checked the things that would hurt most if they were wrong — not just the things that are easiest to check?

The Creed

Fluency is not knowing what the machine can do. It is knowing what to ask of it, how to ask, when to doubt it, and what to put your name on. The rest is just typing.

— Filed from Mount Pleasant, South Carolina —

The Four
Disciplines

One‑Eighth Above the Water

The Disciplines

The Three Modes

The Four Standards