Usage standards
AI tools on a programme either produce business results or they quietly become another line item. The difference is usually not the tool. It is whether anyone agreed what good use looks like.
The failure this prevents
Twelve consultants each have a licence. Each works in their own chat window. The same client question gets asked and answered eleven times, slightly differently. None of the output is written down anywhere the next person will find it. At the end of the quarter someone asks what the tooling delivered, and the honest answer is that nobody can tell — there is a bill, and there is a general feeling that things went faster.
That is not a tooling problem. It is the absence of a standard.
Three spend classes
Almost every argument about AI cost dissolves once the team agrees which actions are which. The asymmetry is deliberate: nobody should be asking permission to breathe, and everybody should be asking before money moves.
| Class | Examples | Rule |
|---|---|---|
| Free | Recalling something from the session, reading a file in the repository, one lookup | Just do it. Never ask. |
| Cheap | One web search, one page, one short document, one commit | Just do it. Never ask. |
| Expensive | Deep research, large batches, anything run across many entities, anything in a loop | Confirm first, with a reason. |
The test before anything expensive
Say in one sentence what it will do, why the cheaper approaches could not answer, and what the cheap version would have produced. If that sentence cannot be written honestly, the action is not justified.
This applies to the person as much as to the tool. It takes ten seconds and it is the only cost control on this page that survives contact with a deadline.
What to measure
Not tokens. Nobody on a steering committee has ever been moved by a token count.
Measure cost per artefact that survived — how much was spent, divided by the number of decisions, analyses and deliverables that ended up in the repository and were still being referenced a month later. It is a crude number and it is the right one, because it prices the thing the programme actually wanted.
A team where that number is high is usually not overspending. It is producing output that evaporates.
What things actually cost
Worth knowing, because most people's intuition is badly wrong in both directions. Prices below were taken from the vendors' own pricing pages on 18 September 2026 and will move.
A worked case: regenerating this entire website — five pages, every guide, around 11,300 words plus the markup — done iteratively over roughly fifteen exchanges, which is what building it honestly looks like.
| Tool | Cost |
|---|---|
| Claude Haiku 4.5, with caching | $0.59 |
| Claude Sonnet 5, with caching | $1.19 |
| Perplexity Sonar Pro | $2.87 |
| Claude Opus 5, with caching | $2.96 |
| Claude Opus 5, no caching | $5.95 |
Three dollars. An entire site, top to bottom, on the most capable model available.
Two things that follow from that
Producing the document is not where the money goes. If a programme's AI bill is large, it is not because people are writing too much. It is because sessions start cold and re-derive what the team already knew — the same analysis, commissioned three times by three people who could not find the first one. That is a filing problem wearing a cost problem's clothes, and the repository fixes it.
Caching roughly halves a real workload, and not every tool offers it. When most of what you send is unchanged from the last message, the unchanged part can be charged at a tenth of the normal rate. In the table above that is the whole difference between $2.96 and $5.95 on the same run. It is worth knowing which of your tools does this before standardising on one.
A standard you can adopt as-is
Six lines. Put them in the programme's ways-of-working pack, or in canon/CANON.md so the tools follow them too.
Repository first
Every workstream has one. Output lands there or it did not happen.
Read before searching
The repository is checked before the internet. Most waste is here.
Files, not transcripts
Ask for a file. A chat window is not a deliverable.
- 4 · Confirm before expensive. One sentence naming what it will do and why the cheap path could not.
- 5 · Decisions get recorded. What was chosen, the alternatives, what it rests on, who agreed. Eight lines.
- 6 · Review monthly. Spend against artefacts that survived. Ten minutes, not a workstream.
What not to do
- Do not set a token budget per person. It punishes the people doing the most thinking and it measures the wrong thing.
- Do not standardise on one tool because of the headline rate. Caching, context limits and how well each one reads a repository matter more than the per-million price.
- Do not make the standard long. A standard nobody reads is worse than none, because everyone assumes it is being followed.
- Do not put client data in a prompt without checking what your firm and the client have agreed. This is the one on the list that can end an engagement.
If you adopt this and it works — or it does not — say so. Numbers from a real programme are worth more than this page.