How to budget for AI, and whether costs climb once agents are embedded
Not a percentage of profit. Budget per job, in three lines – the licences, the build, and the owner's time. Costs do climb once agents are embedded, in two specific ways, and both can be capped before you start.

Why a percentage of profit is the wrong question
A percentage of profit tells you how much you can afford to lose. It doesn't tell you what anything costs or returns. AI isn't a category of spend like marketing. It's the cost of doing a particular job a different way, and the only budget that makes sense is the one for that job, set against the number you measured before you started. The strategy article is where that before number comes from. You wouldn't hire someone on a percentage of profit without knowing what they'd do, so don't budget for tools that way either.
The three lines
| Line | What drives it | Shape | How to cap it |
|---|---|---|---|
| Licences and usage | Seats for the assistants, and metered usage for the agents, which rises with the volume they process. | Small and recurring. Often the smallest line, and the one that gets the most debate. | An alert per agent at a monthly ceiling. A seat review each quarter. |
| The build | A builder's days: design sessions, the build, the test cases, the first month of fixes. | The big one-off. Days to weeks per tailored agent. | Fixed scope through the gate, and the five-case acceptance test agreed before building starts. |
| The owner's time | About an hour a week per live tool, more in the first month, for as long as it runs. | Recurring, invisible, and the one that gets missed. | The maintenance budget: a fixed number of live tools, so a new one means an old one retires. |
In the businesses we work with, the first year's cost is mostly people's time – the build and the owning – not licences. A budget that only counts the licences is off by a multiple.
The two ways costs climb once agents are embedded
Volume: an agent that reads every reply and updates the CRM costs more in a busy month than a quiet one, because it did more work. That's good news dressed up as bad – the cost is tied to activity. It only becomes a problem when nobody set a ceiling, so the first big month comes as a surprise. Set an alert per agent, at a level you'd be happy to pay if the work was real.
Change: once an agent is embedded, a change to the process is also a change to the agent. The quoting rules move, so the agent needs its rules moved, tested and republished. This is the maintenance band, and it's why every agent needs an owner and why the number of live tools has to be capped at what the owners can actually look after.
Neither one is sudden. Both are predictable, as long as the lines were on the budget to begin with.
How much a change like this needs
We won't put a figure on it. The right one depends on your builder's day rate, which jobs you pick and how many people are involved, and a number from someone else's business is the wrong number for yours. What we can give you is the shape. For a company of 20 to 200 people starting from nothing: the general assistant is cheap and quick, the first tailored agent costs mostly in a builder's days and an owner's hours over a quarter, and the cost of the change as a whole is overwhelmingly your own people's time – which is why protecting that time is a budget decision rather than a culture one.
Build the budget as a list of jobs, each with its three lines and its before number. Add them up. That's your number, and you can defend it in a way you never could a percentage.
The rules we'd set
- A usage alert on every agent, at a monthly ceiling the owner's agreed to.
- A quarterly review of every live tool: what it cost, across all three lines, against the number it moved.
- A kill rule: a tool that costs more than it moves for two quarters gets retired, and the owner's hour goes to the next job.
- No new live tool without an old one retiring, once the maintenance budget is full.