段取り · A Dandori field guide

Why most AI fails in complex operations.

The quiet failure nobody talks about: software that arrives and leaves the hard part — the preparation — to you. And what to look for so it doesn't happen to you.

The quiet failure

It doesn't fail with a bang. It fails with a shrug.

Most AI does not explode, crash the business, or reveal itself as useless. It simply arrives, impresses everyone in a demo, produces a few promising outputs — then quietly stalls when it meets the actual operation. Six months later the work runs exactly as it did before. No disaster. No error to point at. The pilot never quite became the practice.

When someone asks why, the answer is always the same short list of suspects: the model wasn't good enough, the data wasn't ready, the organization wasn't mature enough. The real reason almost never comes up — because the whole industry has trained us to talk about models instead of preparation.

The model is rarely the problem. It can summarize, draft, classify, and reason through examples. But complex operations don't run on "something that looks right." They run on the preparation — and that is the part the software leaves to you.

What "capable" still leaves undone

A capable model still has to be told what work to do, which system holds the truth, which rules matter, which exceptions must stop it, who approves the edge cases, what counts as correct, what must be logged, what must never happen on its own, and what success looks like in numbers.

Until someone supplies all of that, the tool is not an agent. It is a burden — and that someone is usually you.

The gap · 段取り八分

Most AI is sold as the small part of the work.

段取り (dandori) is the preparation a craftsman does before the work begins — laying out the tools, measuring the stock, sequencing every cut so that when the blade moves, it moves once and moves right. There is an old proverb, and it is the most honest thing anyone has said about AI in operations.

What shipsthe capability · the blade
vs
What complex work needsthe preparation · left to you

段取り八分、仕事二分 — preparation is eight-tenths of the work; the work itself, only two. Most AI is sold as that other two-tenths: the model, the platform, the capability. The eight-tenths arrives as your homework — unmarked, unmentioned, and due immediately. The gap between what arrives and what the work needs is exactly the size of the preparation nobody did.

Why complexity punishes it

Simple tasks forgive loose AI. Complex operations do not.

A summary can be slightly imperfect and still useful; a first draft can be rewritten; a search result can just point someone in the right direction. Complex work has no such slack.

Pricing

A quote must use the right pricing logic — not a plausible-looking number.

Support

A support reply must reflect the customer's status and your actual policy.

Finance

A finance workflow must know when not to touch a dispute.

Scheduling

A scheduling workflow must respect real capacity, not just open calendar slots.

Procurement

A procurement workflow must know which rules apply to which vendor.

Operations

An operations workflow must understand dependencies before it suggests a change.

In each of these, the work is not just the visible task — it is the surrounding system of rules. That is why generic AI works in a demo and stalls in practice. The demo shows the model's capability; the operation reveals the missing preparation.

The five preparation gaps

Nearly every failed effort has one of these five.

"Preparation" sounds soft until you name what's missing. Five gaps account for most of it — each one quietly handed to the buyer.

一 · The workflow was never bounded

AI gets pointed at a direction, not a job — "improve customer service," "help finance." Those can't be traced from trigger to done state. A bounded workflow has a clear beginning and end. Unbounded, the agent wanders and nothing can be measured.

The tell: language that sounds strategic but names no trigger and no finish.

二 · The rules were never written down

Many operations run on rules people know but never documented: who gets a discount, which customer gets escalated, which field wins when two systems disagree. If those live only in heads, the agent guesses or constantly asks for help.

Without written rules, autonomy becomes hope.

三 · Systems connected, meaning not

Vendors talk about integrations as if connection equals understanding. An API can pull data; it doesn't know which data matters, which system wins when records conflict, which status is stale. The agent must understand how systems relate, not merely reach them.

Connection is plumbing. Preparation is meaning.

四 · No one owned "correct"

Every proof reaches a moment where two people disagree about whether an output is right. A proof can't survive without one accountable person who can settle what "correct" means. Without that owner, success criteria drift and the agent gets tuned toward whoever spoke last. One owner — not a committee.

五 · The outcome was never measurable

Projects begin with a feeling — "this should save time." Feelings aren't proof. A proof needs a line drawn in numbers before the work starts: time saved per week, turnaround time, error rate, escalation volume, days to payment.

Without measurement, "better" is only an opinion.

The warning signs before you buy

The cheapest time to avoid failure is before the software arrives.

The quiet failure hides in one question: who ends up owning the preparation? Five signs give it away.

01

The vendor shows features, not workflow preparation.

Ask — Who maps the process? Who writes the rules? Who defines exceptions? Who identifies the systems of record? Who tests the edge cases?

The tell — "Your team." Then you're not buying a finished solution — you're buying a tool and inheriting the hard part.

02

The vendor says "it learns" but can't explain what is captured.

Ask — What exactly improves? Where is it stored — model, prompt, or workflow logic? Can you inspect it? Can you take it with you? Could you run it on a different model later?

The tell — Vague learning language with no answer. If the vendor can't explain what compounds, you may be renting a black box.

03

The proof has no written pass/fail line.

Ask — What counts as success? What's the baseline and the target? Who decides whether the result is good enough? What happens if it misses?

The tell — "You'll see the value once it's running." A measure defined after the fact can never be failed.

04

The system requires your team to become the AI team.

Ask — Who does the integration? Who writes the instructions? Who maintains the rules? Who validates outputs? Who updates the workflow when the process changes?

The tell — The answers all point back at you. If software is sold but delivery is inherited, the risk has only moved.

05

The AI can act, but no one can show the control model.

Ask — What can it do alone? What can it only recommend? When does it ask for approval? What's blocked? What's logged? What can be reversed? Who sees the audit trail?

The tell — Autonomy discussed before control. When a vendor talks about what the agent can do before what it can't, slow down.

"It learns your process automatically."
"It works out of the box."
"We have the best model."

Each sells the two-tenths while quietly keeping the eight for you. In complex operations, the model is rarely the constraint. The preparation always is.

What good looks like

Good AI delivery starts before the build — with preparation.

A good delivery process doesn't ask the model to figure out the business. It gives the agent a prepared operating environment — the workflow map, the source systems, the business rules, the approval model, the exception paths, the success criteria, the audit requirements, the human control points — and only then builds.

01Assess. Survey the workflow and its systems. Find the real edges.
02Prepare. Write the rules, map the systems, agree the measure. The eight-tenths.
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.

Five stages, and the second is the one everyone else skips. The build is the easy fifth — Prepare is most of the value.

What to demand from an AI partner

A weak vendor says yes to every workflow. A strong partner helps you choose the right first one — and is honest when a workflow isn't ready.

Before you commit · 五

Five signs a workflow is ready — or isn't.

A quick read on any workflow before you buy AI or approve a pilot. If more than one of these is missing, the work isn't ready yet — which doesn't mean AI can't help, only that the preparation has to come first.

一 · Bounded

It has a clear start and a clear finish.

You can trace it from trigger to done state — not a direction like "improve service," but a job with edges.

二 · Written

The rules can be written down.

Who gets a discount, which case escalates, which field wins when systems disagree — captured, not living only in someone's head.

三 · Reachable

The systems have a way in — and the meaning is clear.

Not just an API that pulls data, but a clear picture of which record holds the truth when they conflict.

四 · Owned

One person can settle what "correct" means.

Not a committee. One accountable owner who can say: this is right, this is wrong, this must escalate.

五 · Measurable

"Done right" can be stated in numbers.

A line drawn before the work starts — turnaround time, error rate, days to payment. Without it, "better" is only an opinion.

Ready to test a specific workflow?

These five expand into a ten-question self-check — with a live scorer — in our companion guide.

Choose your first AI workflow

How we stand behind it · The 30-Day Dandori Proof

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

We prepare the workflow you chose the way a craftsman prepares a job: the rules written down, the systems mapped, the success criteria agreed in writing before a single piece of work is done. Then we build one governed agent, run it on your real work, and measure it against the line you set.

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 one lesson

Most AI doesn't fail because it lacks intelligence. It fails because the work was never prepared — the workflow too vague, the rules unwritten, the systems connected but not understood, the exceptions ownerless, the result unmeasured. That is avoidable: choose the right workflow, prepare it properly, and prove it on real work before the agent ever acts.

Bring one workflow to a workshop