Your AI bill climbs month after month, and nobody can say who consumes what, or why. Managing AI token costs in the enterprise is not about restricting usage: it is about installing four governance levers — visibility, guardrails, optimisation, value-versus-cost arbitration — that turn the AI budget into something you steer rather than absorb. Across hundreds of users, integrations and agents, token consumption grows far faster than anyone anticipates; without steering, it becomes the budget blind spot of AI. This article explains why the bill spirals, and how to take it back — including with measurable technical levers such as prompt caching (a cache read is billed at 10% of the input price) and batch processing (an automatic −50% on the Anthropic API).
One point of framing: METASENSE is a Creative Tech agency based in Vélizy-Villacoublay. We design and operate AI systems every day — development, agents, automation, asset generation — so we steer our own consumption. What we describe here is a method you can transpose to your organisation, not the detail of our internal kitchen: the levers are the same, the steering stays yours.
Worth remembering — AI costs are not controlled by telling people to use less, but by knowing where every euro goes and what it returns. The four levers — see who consumes what, set rules and quotas, run the right model for the right task, weigh value against cost — turn an opaque bill into a budget line you can defend.
Frame the governance of your AI costs →
What is a token, and why does it decide your AI bill?
A token is the unit of text — a fragment of a word — that AI models read and produce, and that vendors bill: every request consumes tokens on the input side (your request plus its context) and on the output side (the answer). In English, a token averages a little under one word. Everything you send to the model and everything it returns is converted into tokens, then billed at a price per million tokens — with different rates for input and output.
Three mechanics explain why the token is the right level of reading for a CFO:
- Context is paid for every single time. If every request re-injects a long system prompt, reference documents or a conversation history, you re-bill that same content on each call. This is the first source of invisible drift.
- Input and output are not priced alike. Output tokens usually cost several times more than input tokens. A verbose answer therefore costs structurally more than a tightly framed one.
- The model you pick changes everything. A top-tier reasoning model costs a multiple of a lightweight model for the same task. Using the former for trivial work is paying Formula 1 rates to fetch the bread.
A useful definition — The token cost is the amount billed for the volume of tokens consumed (input plus output) by your AI usage. At organisation scale, that amount is not driven by unit price but by volume — so by usage, by model choice, and by how context is managed.
Why does the token bill spiral at scale?
The bill spirals because token consumption grows non-linearly: the more users, integrations and agents you add, the more the volume explodes — while visibility stays at zero. AI cost drift is almost never the result of a unit price that is too high: it is an accumulation of unsteered usage.
Four factors compound at enterprise scale:
- The sheer weight of users. A handful of enthusiastic pilots weigh nothing; hundreds of employees adopting AI do. Moving from proof of concept to rollout multiplies volume by a factor nobody budgeted for.
- Integrations and agents run without a human. A workflow plugged into your tools through MCP, or an orchestrated multi-step sequence (Cowork), consumes tokens automatically, sometimes in a loop, with nobody watching the meter. A poorly framed agent can burn tokens overnight.
- Context swells. As usage grows more sophisticated, longer documents, heavier histories and richer instructions get injected — each of them re-billed on every call.
- Nobody owns the budget line. AI is paid for sometimes by IT, sometimes by the business, sometimes on subscription, sometimes on consumption. With no owner and no reporting, the drift is discovered on the invoice, not before.
Market context sharpens the stakes. According to Gartner, worldwide generative AI spending was set to reach $644 billion in 2025, up 76.4% year on year — and the same firm notes that boards, disappointed by proofs of concept with high failure rates, are tightening budget scrutiny (see Sources). In other words: we spend more and more, and tolerate vagueness less and less. That is precisely the territory of AI FinOps.
The real problem — It is not that AI is expensive, it is that nobody knows what it costs or what it returns. A bill that doubles can be excellent news (value tripled) or pure leakage. Without governance, you cannot tell the difference — and that is the budget risk.
What is AI FinOps?
AI FinOps is the discipline of making AI costs visible, attributable and arbitrable, so that every euro spent on tokens is tied to the value it produces. It applies cloud FinOps principles to generative AI: the goal is not to spend less on principle, but to spend correctly — cut waste, invest where the leverage is real, and give leadership a clear reading of the cost-to-value ratio.
In practice, AI FinOps answers three questions finance always asks, and that enterprise AI rarely answers:
| The board's question | What AI FinOps delivers |
|---|---|
| Who consumes, and how much? | Cost attribution by team, by use case, by workflow |
| Is it drifting? | Thresholds, alerts and quotas that warn before the invoice |
| Is it worth it? | Value-versus-cost arbitration use case by use case, to cut or invest |
AI FinOps is not a single tool: it is a loop — measure, frame, optimise, arbitrate, repeat. The four levers below are its pillars.
The four levers for managing AI token costs
Managing token costs rests on four levers, in this order: (1) visibility — knowing who consumes what; (2) guardrails — setting quotas and rules per department; (3) optimisation — the right model for the right task, standardised prompts, batching; (4) value-versus-cost arbitration — tying each use case to what it returns. You cannot optimise what you do not measure, nor arbitrate what you have not framed: the order matters.
| # | Lever | The question it settles | Effect on the bill |
|---|---|---|---|
| 1 | Visibility | Who consumes what, where does the money go? | Drift is visible before the invoice, by team and by use case |
| 2 | Guardrails | How do you stop one use case running away with the budget? | Experimentation stays bounded; no silent leakage |
| 3 | Optimisation | How do you pay less for the same result? | Direct reduction of the unit cost of every call |
| 4 | Arbitration | Is this use case worth what it costs? | Budget goes where it genuinely creates value |
Lever 1 — Visibility: who consumes what?
The first lever makes consumption visible: knowing, by team and by use case, how many tokens are consumed, by whom, and with what trend. You cannot steer a bill you cannot see. It is the prerequisite for everything else.
In practice, two sources feed that visibility. On the subscription side, the admin console of enterprise plans (Claude Enterprise, for example) exposes usage analytics — who uses AI, how often, for what. On the API and integration side, every call returns its token consumption, which you aggregate into a dashboard a non-technical owner can read (volumes, cost per team or per workflow, trend). The objective: make drift visible — a workflow consuming ten times the norm after an update, for instance.
Lever 2 — Guardrails: quotas and rules per department
The second lever sets safety rails: consumption quotas and usage rules per department, per application and per use case, so experimentation never turns into budget leakage. Visibility tells you what is happening; guardrails stop the worst from happening.
This takes the form of caps (a token budget per team or per project), alert thresholds (you warn at 70% of budget, not at 110%), and simple rules (which tasks default to a lightweight model, which documents are never sent in full). Enterprise plans allow spend limits per group; on the API side, budgets and alerts can be configured precisely. The point is not to police usage, but to make experimentation safe: a team can test without risking the quarter's budget line.
Lever 3 — Optimisation: the right model, the right prompts, batching
The third lever cuts unit cost at equal quality, through three concrete and measurable mechanics: routing each task to the right model, standardising prompts and caching repeated context, and grouping high-volume work (batching). This is where the measured gains are clearest.
- The right model for the right task (model routing). A lightweight, economical model for simple, high-volume work (classifying, extracting, summarising routine material), a reasoning model reserved for complex work. On the Anthropic price list, the gap between a lightweight model and a top-tier one is measured in multiples: routing correctly means you stop paying premium rates for trivial tasks (see Sources).
- Standardised prompts plus prompt caching. Framed prompt templates avoid verbose requests and back-and-forth. Above all, prompt caching stores stable, repeated context (system instructions, reference documents) instead of re-billing it at full price: on the Anthropic API, a cached token read is billed at 10% of the standard input price — up to 90% saved on the cached portion (see Sources).
- Batching for high-volume work. For non-urgent tasks processed in bulk (analysing thousands of documents, enriching a database), Anthropic's Batch API applies an automatic −50% on input and output tokens, with no loss of quality. Combined with prompt caching, the cost of a request can fall to a fraction of the standard rate (see Sources).
An honest caveat — These technical levers are powerful, but they do not undo poor framing. Halving the cost of a use case that serves no purpose is wasting half as much, not spending better. Optimisation comes after visibility and guardrails — never instead of them.
Lever 4 — Value-versus-cost arbitration: cut or invest
The fourth lever is a leadership decision: tying each AI use case to the value it produces, in order to cut what costs without returning and reinforce where the leverage is real. This is what turns cost governance into a strategic steering tool rather than a simple spending hunt.
Arbitration depends on the three previous levers: without visibility and attribution, you cannot tie a cost to a value. Once the data exists, leadership can decide with full knowledge — a legal workflow that costs a lot but saves two days per case deserves the investment; a gimmick that consumes with no measurable effect gets cut. The right habit: review this portfolio of use cases regularly, the way you review a budget.
The board argument — Cost control is not a brake on AI, it is what makes AI sustainable and defensible at budget time. AI spending you can attribute and justify survives financial arbitration; an opaque invoice gets cut at the first tightening. Cost governance is often what unlocks — and secures — the decision on the finance side.
How much can you realistically save?
The orders of magnitude are documented and significant, but they apply to the optimisable share of the bill, not to its total: prompt caching cuts up to 90% of the cost of cached context, batching applies −50% on bulk processing, and good model routing avoids paying premium rates for trivial tasks. These gains compound: Anthropic states that a cached request processed in batch can come down to roughly 5% of the cost of a standard, uncached request (see Sources).
Two honest caveats, essential for a finance decision-maker:
- These percentages apply to part of your usage, not to the whole invoice. Batching only applies to non-urgent work; caching only to genuinely repeated context. Real savings depend on your usage mix — which is exactly what an audit quantifies.
- The biggest reserve is often not technical. Cutting three worthless use cases and routing models correctly frequently weighs more than any fine-grained optimisation. Technique optimises; governance decides.
Worth remembering — We do not promise a savings percentage on your invoice without having seen your usage: that would be dishonest. What the method guarantees is a readable, arbitrable invoice — and, on that basis, documented savings levers (caching, batching, routing) applied where they are worth it.
Why bring in support on AI cost governance?
Because the difficulty is not knowing the levers — they are public — but instrumenting them in your context: wiring visibility into your subscriptions and integrations, setting quotas that do not smother usage, routing intelligently, and holding value-versus-cost arbitration over time. That is a matter of method and tooling, not a recipe you apply once.
Our standing is concrete: we design and operate AI systems every day — development, agents, automation, asset generation. So we steer our own consumption, and we know the difference between optimisation that holds in production and a trick that breaks quality. This is the Creative Tech double role applied to cost: advisory (framing, attributing, arbitrating) and technical expertise (genuinely understanding where the tokens go and how to save them without degrading the service).
And we have tested it in real skills-transfer work, with public and private organisations:
- Grand Angoulême — agentic AI talk and workshop. A format designed to help non-specialist participants understand and frame agentic AI use cases — exactly the logic of "we bring the method, you keep control". (Public-sector organisation, citable.)
- Bertrandt — HR innovation. An ideation workshop and the design of an AI CV-screening assistant: a targeted business use case, framed with the teams concerned rather than rolled out blind.
- Strate (design school) & Gocad Lab. Training and Design Thinking modules (Gocad Lab: 4 workshops, 24 employees) — our habit of transferring towards autonomy, until the teams steer on their own.
Honesty — Cost governance is built with your finance, IT and business teams: they are the ones who know your budget constraints and your usage. We bring the method, the tooling and the technical command; the steering stays yours. And the way we steer our own tokens stays internal: what we pass on is a repeatable method.
This cost governance sits within a broader effort to embed AI in your processes.
Where do you start to take back control?
Start by seeing before cutting: a picture of actual consumption, then guardrails, then optimisation, then arbitration. You do not optimise in the dark. A realistic path:
- Map current consumption — subscriptions and API together: who consumes, which use cases, what trend.
- Attribute the cost by team and by use case, so finance can read the invoice.
- Set the first guardrails — quotas and alerts on the heaviest use cases.
- Switch on the high-leverage optimisations — model routing by default, prompt caching on repeated context, batching on bulk processing.
- Put arbitration in place — a regular cost-versus-value review of use cases, with leadership.
- Measure and iterate — cost governance is a loop, not a one-off project.
Worth remembering — The classic trap is waiting for a nasty surprise on the invoice before reacting. Cost governance is wired in from rollout, not afterwards. And the first deliverable is not a saving: it is visibility — without it, every AI budget decision is a bet.
To go further on the AI pillar: building reusable Claude Skills and orchestrating AI agents without blowing the budget.
Let's frame the governance of your AI costs
Watching the invoice climb without being able to explain or defend it? METASENSE (Vélizy-Villacoublay) installs governance over your AI costs — visibility, guardrails, optimisation, arbitration — with your finance, IT and business teams, for an AI budget you finally steer. Controlling token costs often more than pays for the engagement: avoiding drift across the organisation weighs more than the cost of framing it.
Explore AI enablement & cost governance →
Frequently asked questions
What is a token, and why can its cost spiral?
A token is the unit of text — a fragment of a word — that AI reads and produces, and that vendors bill. Every request consumes tokens on input and output. At organisation scale (hundreds of users, integrations, agents), volume grows far faster than expected, especially without visibility. It is the budget blind spot of enterprise AI.
How do you manage token costs at enterprise scale?
Through four levers, in order: visibility (who consumes what, by team and use case, with alerts), guardrails (quotas and rules per department), optimisation (the right model for the right task, standardised prompts, prompt caching, batching) and value-versus-cost arbitration (tying each use case to what it returns). You cannot optimise what you do not measure.
What is AI FinOps?
AI FinOps is the discipline that makes AI costs visible, attributable and arbitrable, tying every euro of tokens to the value produced. It applies cloud FinOps to generative AI: not spending less on principle, but spending correctly — cutting waste, investing where the leverage is real.
How much can you save on your AI bill?
The documented levers are significant on the optimisable share: prompt caching cuts up to 90% of the cost of cached context (a cache read is billed at 10% of the input price at Anthropic), batching applies an automatic −50% on bulk processing, and good model routing avoids paying premium rates for trivial tasks. Real savings depend on your usage mix.
What is prompt caching, and how much does it save?
Prompt caching stores stable, repeated context (system instructions, reference documents) so it is not re-billed at full price on every call. On the Anthropic API, a cached token read costs 10% of the standard input price, meaning up to 90% saved on the cached portion. It is one of the most profitable levers where context repeats.
What is model routing, and why does it cut the bill?
Model routing means directing each task to the most suitable model: a lightweight, economical model for simple, high-volume work, a reasoning model reserved for complex work. Since the price gap between models is measured in multiples, routing correctly avoids paying premium rates for trivial tasks — often one of the largest savings reserves.
What is batch processing in AI?
Batching groups non-urgent requests to process them in bulk, asynchronously. On the Anthropic API, Batch mode automatically applies −50% on input and output tokens, with no loss of quality and processing within 24 hours. Ideal for mass document analysis or database enrichment, where instant responses are not required.
Why is our AI bill rising without anyone knowing why?
Because token consumption grows non-linearly with users, integrations and agents, while visibility stays at zero. Agents and automated workflows consume with no human at the meter, injected context swells, and nobody owns the budget line. The drift is discovered on the invoice, for lack of upstream reporting.
Does cost governance slow AI adoption?
No, quite the opposite. Without governance, leadership ends up cutting the AI budget at the first doubt, unable to defend it. A readable, attributed, arbitrated invoice makes the spending sustainable and defensible — it survives arbitration. Guardrails also make experimentation safe: a team can test without running away with the quarter's budget.
Do you need Claude Enterprise to steer token costs?
Not necessarily, but it helps. The Claude Enterprise admin console exposes usage analytics and spend limits per group, useful for visibility and guardrails. On the API side, consumption can be measured and capped just as precisely. What matters is not the plan but the method: aggregate the data, attribute the cost and arbitrate the value.
Sources
- Gartner — Forecasts Worldwide GenAI Spending to Reach $644 Billion in 2025 (+76.4%; tightening budget scrutiny).
- Anthropic — Prompt caching (API documentation) (cache read billed at 10% of the input price).
- Anthropic — Pricing (per-model rates, input/output, batch).
- Anthropic — Message Batches API (automatic −50% on tokens).
- Flowt — LLM FinOps : maîtriser la facture cloud de vos projets d'IA générative.
- Safebrain — Pilotage de la consommation de tokens IA : reprendre le contrôle.