A year ago the cost of AI in a professional firm was a seat licence and a rounding error. McKinsey's 2026 State of AI survey reports that for roughly one in five organisations AI-related operating costs are now beginning to constrain use, even as most expect to increase investment. The reason is structural. A chatbot answers a question and stops. An agent runs a workflow, reads forty documents, calls six tools and reasons at every step. The same firm doing more useful work with AI is, by definition, spending more on it.
1. What changed in the bill
Three things. Agents consume tokens for as long as a task takes, so a diligence pass over a large data room costs what it costs. Reasoning-heavy models charge for thinking as well as for output. And usage moved from a few enthusiasts to the whole firm. None of these is a problem in itself. Together they produce a number that a finance director notices.
2. The four levers
- Routing. No single model is best for every task, and the most capable model is rarely the right one for classifying an invoice. Firms that route each job to the cheapest model that meets its quality bar, with automatic fallback when it does not, report inference costs a fraction of firms that default to the top model.
- Effort. Vendors now expose settings that trade depth of reasoning against cost. A first-pass triage does not need the same effort as the memo that goes to the investment committee.
- Caps. Hard spend limits per agent session, now available on the major platforms, turn a runaway task from a surprise invoice into a stopped job that asks for approval.
- Caching and memory. A firm's own context, its templates, precedents and how it does things, should be supplied once and reused, not re-read from scratch on every run.
The cheapest token is the one the workflow did not need. The second cheapest is the one a smaller model produced.
3. Pricing follows cost
Cost visibility changes how firms price. When the cost to serve a fixed-fee engagement includes a known inference line, the fee can be set with a margin rather than a hope. Several firms now track AI cost per matter alongside hours per matter, and a few have started showing clients both. That is a stronger position than absorbing an unknown cost into a rate card designed for a different delivery model.
4. What to ask a vendor
- Can we route across models and set fallback rules ourselves?
- Can we set effort and spend limits per workflow, not just per account?
- Can we see cost per run, per workflow and per client, exported to our own systems?
- What happens to cost when the model we use today is replaced next quarter?
A vendor with good answers has designed for a world in which AI is a line item. A vendor without them is designing for the world before.
Hive Newsroom follows what is changing in AI and professional services. Sources are linked in the text; figures are as published at the time of writing. Comments and corrections: press@get-hive.ai. Back to the .