An Agent Is Not an App

A proposal reaches your desk for an AI agent. It has a price, a start date, and a per-seat line that looks like every other piece of software you have ever approved. Somebody has done the arithmetic against the hours it saves, and the number works.

Approve it on that document and you have made a category error, because the thing you just bought doesn’t behave like the things that document was built for.

Here is the rule underneath everything that follows. Software waits to be asked. An agent acts without being asked. Everything expensive about agents descends from that one difference, and none of it appears on the invoice.

What the buying document assumes

A license is a permission. You pay, it activates, people use it or they don’t, and the work around it stays exactly as it was. The failure mode of buying software badly is waste: a tool nobody opens, renewed out of inattention. Annoying, bounded, visible at renewal.

An agent inverts that. It doesn’t sit waiting for someone to open it. It takes an action inside a process, and that process was built on the assumption that a person would be standing there when the action happened. Remove the person without changing the process and you haven’t automated the step. You have removed the checkpoint the step relied on.

So the useful question at purchase isn’t what the agent costs or what it saves. It is what it now decides on its own, and what used to happen at that moment.

The evidence says the constraint is not the model

Deloitte surveyed 501 US senior managers and executives between April and June 2026, all at organizations already piloting agents, with the follow-up interviews done at large organizations. Among that committed group, only 15% had scaled what the study calls orchestrated, cross-functional multi-agent adoption. Only 5% said their business processes were highly prepared for agents at all.

The interesting part is what those leaders named as the obstacle. 72% said they lack unified, accessible data. 70% said they don’t feel they can trust and govern agents. 67% said integration is too costly and complex.

Notice what is absent from that list. Nobody says the models aren’t good enough. What they name instead is data, oversight and integration.

Part of that trust figure surely does belong to the agents themselves, which are genuinely uneven. The rest belongs to everything around them, and that part is the half a company controls.

There is a fair objection to leaning on a consultancy’s survey here, and it is better stated than buried: Deloitte sells the transformation work its own findings recommend, and the sample was drawn from organizations already committed enough to be running pilots. The proportions describe the committed end of the market, not the market.

Which is why the second source matters more than the first.

The US Census Bureau’s Business Trends and Outlook Survey has no remedy to sell. Its May 2026 release puts overall AI use among US businesses between 17% and 20%, a figure held down by the very small firms that dominate the count: fewer than 20% of businesses with four or fewer employees use AI at all. Read by size, the picture changes. 32% of firms with 100 to 249 employees said they used AI, against 37% of firms with at least 250. A Census working paper published in April 2026, by Bonney, Breaux, Dinlersoz, Haltiwanger and co-authors, found that among firms using AI at all, 57% have it in three or fewer business functions.

Two unrelated methods, one shape. A consultancy asking committed pilots about agents, and a statistical agency asking the whole economy about AI, both find that adoption does not spread across a company by itself. It stops at the edge of whatever function absorbed it first.

Those two studies measure different objects, and it is worth being precise. Census counts AI of any kind. Deloitte counts agents specifically. They agree on the pattern, not on a number, and anyone merging them into a single statistic is inventing one.

The four things that never appear on the invoice

If an agent behaves less like a license and more like a hire, the useful move is to budget and govern it the way you already know how to budget and govern a hire. You have done this hundreds of times, and the apparatus already exists inside your company.

What follows is our framework rather than a documented industry practice. It’s offered as a way to make the decision legible, not as a standard anyone else is following.

One: onboarding, which for an agent means context. A new hire spends weeks learning what your company knows: which customers are difficult, which numbers are trusted, where the real approval sits regardless of the org chart. An agent gets none of that from a model. It gets it from the data and instructions you put around it, which is exactly what 72% of Deloitte’s respondents said they don’t have in usable form. Budget the context work as the first cost, not as a surprise in month three.

Two: supervision, which means a defined line. Nobody hires a person and lets them act on anything at any value. That line exists implicitly for humans and has to be written down for agents, because an agent won’t hesitate where a person would. Where the line sits is a real decision with real trade-offs, and it deserves more room than this article gives it. We have written separately about where the override threshold belongs.

Three: a probation period, which means reversibility. A hire who isn’t working out gets managed or moved, and everyone accepts a stretch where the verdict is open. Agents rarely get this. They get deployed and then quietly become load-bearing, at which point pulling one back is a deprovisioning exercise rather than a decision. Set the review date at purchase, while removing it is still cheap. There is a failure mode on the other side of this, where probation gets reached for as a way to approve something risky rather than as a decision anyone intends to review, and it shows up clearly in what an agent’s recommendation is actually made of.

Four: an owner, which means a name. Not a team, not a function, not a committee. A person accountable for what the agent produces, the way a manager is accountable for the work of their reports. This is the one most companies skip, and it decides whether anything else on the list actually happens. The person who ends up doing this job is often already on your payroll and unrecognised.

Read the four together and the real cost becomes obvious. None of it is the license. All of it is time and attention from people who are already busy, which is precisely the cost a per-seat line makes invisible.

Why this takes years, and why that is not failure

In the Deloitte data, 74% of leaders expect nearly half their business processes to be redesigned or rebuilt around agents within four years, while fewer than a third expect a majority redesigned within two.

Four years, from a group already running pilots. Set against how fast an agent can be switched on, that is a striking gap: a thing you can deploy in a week, on a clock measured in years.

The gap isn’t a technology failing to arrive. It’s what absorption looks like when the thing being absorbed acts on its own. The redesign is the work. The agent is what makes the redesign necessary.

Which reframes what being behind means. A company running fifteen agents inside one function hasn’t travelled further than a company running two agents through a rebuilt process. It has moved faster along the axis that does not compound. The Census finding, that most AI users sit in three or fewer functions, is what that looks like at national scale.

The uncomfortable version of the same point: if your agent work has produced no argument about how the work should change, it probably hasn’t touched anything that matters yet.

The counter worth taking seriously

Two objections are worth taking seriously, and both say the same thing: the argument is too generous to the technology.

The argument runs roughly like this. When 70% of leaders say they cannot trust and govern agents, the plain reading is that agents aren’t yet trustworthy, and calling it an organizational readiness problem quietly relocates a product shortcoming onto the buyer. Reliability in multi-step autonomous systems has been the binding constraint through several cycles and hasn’t been convincingly solved. On that reading the four-year horizon is not an absorption clock at all. It’s what discovering a capability ceiling looks like from the inside.

There is a second edge, aimed at the hire analogy itself. A hire who delivered nothing for four years would be gone in month four. Recasting a stalled deployment as a maturing one hands executives a vocabulary for postponing the question of payback, and the discipline of the software frame is precisely its impatience: a license that produces nothing gets cancelled at renewal.

Both are fair, and the second one changes what to do rather than whether to do it. Treat the agent like a hire, including the part where a hire who isn’t working out gets managed out. A probation period only means something if it can end badly. Install the four items above without a date on which someone decides the agent is not earning its keep, and the analogy has been used for reassurance rather than for management, which is exactly the failure the objection predicts.

What to do with the next proposal

One question first, and the answer routes everything else: does this thing act without a human pressing go?

If no, it is software. Install it, use it, judge it at renewal.

If yes, then before the signature, put four things in writing. Which context it needs, and who is producing it. Where the line sits between what it decides and what a person signs. The date on which someone reviews whether it is earning its keep. And the name of the person accountable for its output, who should also be the one with authority to switch it off.

If those four can’t be answered, the agent is not ready to buy, and the honest reading is that nobody has yet examined the process it is meant to enter. Budget for the examination, because examining the process is what decides whether an AI budget returns anything at all.

Nobody calls a new hire a failure in week three. Agents are getting judged on a software clock, and on a hiring clock the same numbers read completely differently.