How much of this system is the agent, and how much is everything around it?
The demonstration was quick, and that is the fact worth examining. Somebody built something impressive in what sounded like a fortnight, which means either the work is genuinely small or you were shown the small part of it. Your last pilot already answers this. It worked, everyone was pleased, it never reached anybody outside the project, and nobody could name the obstacle.
A loop, and the four fifths around it
The reasoning part of an agent is a loop. It reads what it is given, decides whether to use a tool, uses it, reads the result, and goes round until it can answer. Around that loop sit the parts making it usable by a company rather than by its builders.
In ArkOne’s reference design for an executive agent, the Cabinet, those parts have names: the guards on what may leave and what happens when they fail, the ledger where every action and its reason are written and never edited, the roster of people it may address, the step where an action waits for a person, and the setting deciding whether a specialist may act alone. Each is a value on the Register, and each is code somebody writes, tests and operates.
Held against a working system, the machinery of reasoning is about a fifth of it. The rest is the guards, the record, the connections into systems the vendor did not build, and the screens your staff see any of it through. That proportion is the argument the anatomy primer makes part by part.
What proportion sounds like in a meeting
Suppose the question goes out in writing after the demonstrations, and three proposals attempt the division.
| What a good answer sounds like | What you will be offered instead | What that hides |
|---|---|---|
| A split you can read: the reasoning here, the guards, ledger and connections there, with days against each | The agent framework handles most of it | A framework supplies the loop and the tool calling. Ask which other parts it supplies, item by item, and watch the list shorten |
| The integration work is named per system, and those without a usable interface are flagged | Integration is straightforward with your existing systems | A category rather than a count. Ask how many systems, which a person operates by hand today, and who owns each connection |
| The approval screens and the ledger view are in scope, with an owner named | The pilot proves the concept; production is a scaling exercise | Scaling as a word for the larger share. What is scaled is a path built for one willing user, in one direction |
Exhibit 1. Illustrative. Three replies about proportion, and the work each of the weaker two leaves unpriced.
The third row explains more failed pilots than any technical fault. A pilot may skip the record, skip the approvals, connect to one system by hand and be run by whoever built it. Production is those omissions, each a piece of engineering rather than a matter of capacity, and calling the gap between them scaling is how a project loses a quarter.
The first row is a fair answer, narrowed. A framework genuinely supplies the loop, the parallel calls and the tool handling. What it does not supply is the part specific to your company, which is why that part never comes out of a box.
The follow-up that produces a number
Ask them to divide the build into two columns in front of you, the reasoning and everything around it, with an estimate on each line.
A vendor who has shipped this before writes the second column faster than the first, and it holds unglamorous entries: an identity setup, a permissions model, a ledger and a way to read it, a connection to a system whose interface is a screen. Ask which entries exist already and which are being built for you, because the second kind carries the risk.
Where the columns come out roughly even, ask what is missing rather than assuming efficiency. Where the reasoning column is larger, you have learned something: the guards and the record are thin, and the work that takes the time has not been done. At a pilot price that need not disqualify anyone. Whether an unrecoverable action can be undone is among the entries most often missing from that second column.
Then ask who operates that larger share once it is live: you, them, or nobody yet.
The same arithmetic decided at length, against building it yourself, is the comparison on an in-house build. A demonstration is honest about what it shows and silent about its own proportion. This question asks a vendor to say the silent part aloud, before the proposal is agreed rather than after a third change request explains it to you.
component: answer-card
Asked plainly
How much of an AI agent system is actually the model?
A small share of it. The reasoning loop and its specialists are the part a demonstration shows, and they can often be built in days. The guards on what may leave, the audit trail, the connections into existing systems, the approval steps and the screens people use are the larger share of the effort, and none of them appear in a demonstration.
Why do AI agent projects cost more than the demonstration suggests?
Because the demonstration shows the part that was quick to build. Everything that makes the same behaviour survivable in a real company sits outside it: identity, permissions, the record of what happened, the rules that stop a message, and the integration work into systems the vendor did not write. A quote that prices only the visible part is priced for the demonstration.
Why did our AI pilot work but never reach production?
Very often because the pilot was the smaller part of the system and the remaining work was never scoped. A pilot needs one working path; production needs failure handling, approvals, an auditable record and connections to systems that were built for people. Those are the parts that take the time, and they are the parts a pilot deliberately skips.