段取り · A Dandori field guide

What "agentic AI" actually means.

Agent vs. assistant, in plain terms: why a prepared agent does the work while a chatbot just talks about it — and why the difference is preparation, not the model.

The confusion

The most-used, least-understood word in AI.

Everyone says "agentic." Almost no one can tell you what makes something an agent.

Ask ten vendors what makes their AI agentic and you'll get ten answers, most of which come down to the same claim: it's a more capable model. That answer is wrong — and the mistake is expensive, because it sends you shopping for the wrong thing. You go looking for a better brain when what you're missing is a prepared job.

The difference between an assistant and an agent has almost nothing to do with the model. It has almost everything to do with preparation.

What this guide does

Explains the distinction in plain terms, shows it with a single concrete example, and gives you the one question that separates a real agent from a chatbot in a costume.

Agent vs. assistant, in one breath

One talks about the work. The other does it.

Assistant · talks about the work

You bring it a request; it produces something — a draft, an answer, a summary — and you take that output and carry the work the rest of the way yourself.

The loop closes on you.
Agent · does the work

You give it a job and the authority to do it; it carries the work to a finished state inside your systems — checking, applying rules, acting, escalating — and tells you what it did.

The loop closes on the work.

That is the whole distinction. An assistant hands you a draft; an agent files it. An assistant tells you what to do; an agent does it and reports back. The gap is not intelligence. It is preparation — and you can watch it happen with a single request.

Same request, two outcomes

Watch one request meet the same model twice.

One request, same model both times: "A customer wants a refund on order #4471. Can you handle it?"

Assistant · talks about the work

Writes a clear, polite reply suggesting the next steps. Still left to you:

  • Open order #4471 and read its details
  • Check it against the refund policy
  • Decide whether it qualifies
  • Issue the refund, log the action, close the ticket
The loop closes on you.
Agent · does the work

Carries the refund workflow to a finished state:

  • Pulls order #4471 from the system of record
  • Checks it against the written refund policy
  • Confirms it's in the return window and under the auto-approve threshold
  • Issues the refund, logs every action to the audit trail, closes the ticket
  • Escalates to a person only if it falls outside its authority
The loop closes on the work.

Same model, both times. The difference is the six preparations wrapped around the agent: the job, the systems access, the rules, the authority, the audit trail, and the escalation path. None of them is the model. All of them are preparation.

Why it's preparation, not the model

Swap in a smarter model and the assistant writes a nicer reply. It still doesn't do the work.

Because the work was never a writing problem. Reading the order, applying the policy, issuing the refund, logging the action, closing the ticket — none of that is language. It's a prepared job: systems the agent can reach, rules it can follow, authority it can exercise, and a trail it leaves behind. A better model raises the ceiling on how well each step is done. It does not, on its own, create the steps.

This is why "we have the best model" is an answer to the wrong question. Capability is the ceiling; preparation is the floor you were actually missing. Give a modest model a fully prepared job and it will do the work. Give the finest model on the market a chat box and no preparation, and it will do what chat boxes do — talk about it, beautifully.

The model is the same blade in both hands. What makes one an agent is the prepared job around it.

段取り八分、仕事二分

Preparation is eight-tenths of the work.

The four shifts

Four preparations turn a chatbot into an agent.

Each is a shift the model doesn't make on its own — someone prepares it. They happen to share a first letter: Assignment · Access · Authority · Accountability.

一 · From prompt to assignment

A job, not a question

An agent is given work to complete, with a clear start and finish — not a message to answer. "Handle the refund" is an assignment; "write me a refund reply" is a prompt. The assignment is what lets the agent know when it is done.

The preparation gives it a job.
二 · From conversation to access

Inside the systems, not just the chat

An agent operates where the work actually lives — reading records and writing results in your systems of record — not producing text in a window you then copy from. Without access, even a perfect answer is just a suggestion.

The preparation gives it reach.
三 · From suggestion to authority

Permission to act, with boundaries

An agent has bounded permission, defined by rules: what it may do alone, what it may only recommend, what it must escalate, and what it must never do. Authority without boundaries is dangerous; a suggestion with no authority is just an assistant.

The preparation grants — and bounds — the power to act.
四 · From output to accountability

A measured result and a trail

An agent works to an agreed measure, leaves an audit trail, and hands off to a person at defined control points. Its output can be checked against a standard, and its actions traced after the fact. That is what makes it something an operation can trust, not just admire.

The preparation makes it answerable.
Why "agentic" demos mislead

A demo makes an assistant look like an agent by hiding the preparation: the systems are mocked, the rules are hardcoded for the happy path, and the authority is unbounded because nothing real is at stake. Strip the preparation away and any capable model can look agentic for five minutes. The trouble comes later, when the same "agent" meets a real operation with real systems, real exceptions, and real consequences — and turns back into what it always was: an assistant that talks about the work.

The one question to ask

"Who prepares the assignment, the access, the authority, and the accountability — and does it run on my systems, or only in your demo?"

The answer tells you whether you're being sold an agent, or a chatbot in a costume.

What good looks like

A prepared agent is the model plus the four.

We don't build assistants. We build agents — which means we do the four preparations as the actual work: a bounded assignment, access to your real systems, governed authority, and measured accountability. The model is the blade we start with. The preparation is the trade.

01Assess. Survey the workflow and its systems. Find the real edges.
02Prepare. The assignment, access, authority, and accountability.
03Build. One governed agent, prepared for your work.
04Prove. Run it on real work; measure against the line you set.
05Handoff. You own the agent, the rules, and the operating logic.
保証 · Pay after proof

If the proof does not meet the written success criteria, you owe no build fee. You keep the prepared workflow documentation, rules, success criteria, and integration design. Production operation begins only if you choose to continue.

The 30-Day Dandori Proof

One workflow. One governed agent. A measured result. Yours to keep.

An assistant talks about the work. An agent does it. The difference isn't the model — it's the preparation. Dandori builds prepared AI agents for real workflows, governed by your rules, integrated with the systems you already run, proven on your own work in thirty days, and owned by you.

Bring one workflow to a workshop