Cost control
You finally know what each use of AI costs you. We combine several models and technologies so that each task runs on whatever does it at the best price, with a meter and caps: the bill no longer surprises you.
What public statistics, research and industry studies measure. Each figure links to its source, and when the author sells consulting or AI, we say so.
54%
of AI requests are sent to a more powerful and more expensive model than the task needs, and 90% of more than 9,000 tasks studied do not need a frontier model.
Source: Accenture, “The CIO’s guide to AI tokenomics”, 10 September 2026
750 executives at companies with over $1 billion in revenue in 17 countries and 15 interviews, June to July 2026. Accenture sells consulting, integration and managed AI services.
26%
of the large US companies surveyed see in real time what running AI at scale costs them, and 36% directly control their token consumption.
Source: KPMG, AI Quarterly Pulse, Q2 2026, 24 June 2026
204 executives at companies with $1 billion or more in revenue, surveyed from April to May 2026. KPMG sells audit, consulting and AI services.
51%
of the large companies surveyed say they received higher generative AI bills than expected, and 57% of those still in the pilot phase.
Source: Capgemini Research Institute, “Harnessing the value of AI”, 10 September 2025
1,020 respondents among 1,100 executives at companies with over $1 billion in revenue, in 15 countries including France, May to June 2025. Capgemini sells consulting, integration and outsourced AI services.
Sorting an email does not call for the most expensive model, sometimes not for AI at all. Each task goes to what is enough (a rule, a small model, the most powerful one last), work already done is served again from the cache, and provider discounts, from 50 to 90%, apply to the rest.
Each use has its budget, its caps and its alerts, agents included: a runaway loop or spending that drifts shows up immediately, not at the end of the month.
Models are compared on your own content, cost measured and quality judged: you know why each one was chosen.
You do not depend on a single provider: the routing table says which model for which task, and changing it is decided on measurements. The single entry point that will make this change without redevelopment is in preparation.
A written grid says what may go to whom: public data at the best price, sensitive data on a model hosted on your side.
If it does not find at least its own amount in recoverable annual savings, we refund the difference.
In one week, you see your real costs and the first obvious savings.
You receive a costed verdict per task, on your own content.
Each system switches to the model chosen for each of its tasks, with its key and its caps.
A report every month, and the table reviewed every quarter.
Below it, the audit alone does not pay for itself. We tell you so and fold it into the wider Audit Sprint: we do not sell a useless audit.
What costs money is the wrong model in the wrong place, work done twice and the absence of a cap. The pack fixes all three.
No. A rule, a script or a search in your data is often enough, and consumes no tokens. The pack first identifies what genuinely needs a model, and entrusts only that to the model that is just sufficient.
An agent plans, acts, checks its result and starts again until it reaches the goal: each round is a billed call. The pack sets it a cap per task, hands repetitive checks to a small model and keeps the most powerful one for the difficult steps. A human review remains in place before anything goes outside.
A written grid classifies it by sensitivity: public data at the best price, internal data with providers that retain nothing, personal and regulated data hosted on your side.
Rarely just one service. Most of our clients combine several: build your cart and ask for a bundled quote, a set of services gets a better price.
Build a bundled quoteGo further on this topic
An AI agent does not do everything in one go: it runs in a loop, gathers context, acts, checks and starts again. Understanding that loop means knowing where to put the controls, the limits and the human review.
Adopted by the French Senate on 18 February 2026, the bill naming the AI Act enforcement authorities is still waiting for the National Assembly. Who would supervise what under the Senate text, what already applies without it, and what a business can prepare in the meantime.
In mid-October 2026, ChatGPT's Chrome extension learned to read open tabs and act in them, and Apple announced it would tighten disk access on Macs because of AI agents. What a browser assistant really sees, and the simple rule that says, tab by tab, where it belongs.
20-minute diagnostic
We look at your situation, identify what can be automated or improved, and tell you plainly whether we can help.