Everything runs on the largest model
The fix: routing
Picking the right model for the kind of request, escalating only when the answer falls short. Decided by measurement, not by feel.
Assessment
Two starting points, one outcome: you know where you stand on AI today, what it costs and what to do next. Both assessments come at a fixed price, with no obligation.
Which assessment fits you?
You are just starting out
Readiness Check
€4,900
Fixed price, net. A workshop with one pilot team, on your site. A standalone engagement, separate from the price ladder below.
For companies that want to adopt AI seriously and need to know where to start. We work with a real team on a real process, instead of laying a strategy across every department that nobody reads afterwards.
- Inventory: what is already running, officially and unofficially
- EU AI Act risk classification, per use case
- Potential map across the departments
- A roadmap with a binding order: what first, what later, what never
You already use AI
Cost picture
€3,000
Fixed price, net. A clear read on the token consumption of your AI systems.
For companies whose AI bill has been running for a while and who cannot break it down. It runs on your providers' billing data and three or four conversations.
- The size and trend of your AI spend
- The model mix, and where the choice does not match the task
- A reasoned estimate per workflow
- The accounts and services nobody in the building knew about
What happens in the Readiness Check
- 1
Workshop with the pilot team
A day or two with the people who actually run the process. Not with the people who would describe it.
- 2
Inventory
What AI is already in the building, who uses it and for what. The shadow IT turns up almost every time.
- 3
Classification and gaps
EU AI Act risk class per use case, plus the gap against deployer obligations: oversight, logging, informing staff, AI literacy under Art. 4.
- 4
Roadmap and plan
An order you can hold yourself to: which process first, what it needs, what it returns, and what would tell us it is not worth doing.
- 1
Workshop with the pilot team
A day or two with the people who actually run the process. Not with the people who would describe it.
- 2
Inventory
What AI is already in the building, who uses it and for what. The shadow IT turns up almost every time.
- 3
Classification and gaps
EU AI Act risk class per use case, plus the gap against deployer obligations: oversight, logging, informing staff, AI literacy under Art. 4.
- 4
Roadmap and plan
An order you can hold yourself to: which process first, what it needs, what it returns, and what would tell us it is not worth doing.
What the cost picture says
- 1
Size and trend
What you spend on AI, how it has moved over the months and where it is heading.
- 2
The model mix
Which model carries how much, and where the choice obviously does not match the task.
- 3
An estimate per workflow
Reasoned, from billing data and conversations. Not a measurement, and we do not call it one.
- 4
Accounts and services
The ones nobody in the building knew about. They turn up almost every time.
- 1
Size and trend
What you spend on AI, how it has moved over the months and where it is heading.
- 2
The model mix
Which model carries how much, and where the choice obviously does not match the task.
- 3
An estimate per workflow
Reasoned, from billing data and conversations. Not a measurement, and we do not call it one.
- 4
Accounts and services
The ones nobody in the building knew about. They turn up almost every time.
What we typically find
A workflow running on the largest model when the smallest returns the same answer.
Repeated calls with identical input, because nothing is cached.
One team causing 60 % of the bill, without anyone having been able to name it.
Every finding comes with a fix.
The same call, paid for again every time
The fix: caching
A layer in front of the model that recognises what has already been asked and reuses the answer.
Five calls where one would do
The fix: rebuild the interaction
Often the model isn't too expensive, the dialogue is badly built. That is design work, not infrastructure.
Nobody knows whether the output is any good
The fix: an evaluation harness
Cost per result means nothing while nobody measures the quality of the result. We set up the harness that checks it.
Step 1 · Cost picture
€3,000
Fixed price, net. It runs on your providers' billing data and three or four conversations.
Cost picture, one-off€3,000
Savings found, per year€9,000
If the cost picture finds less than that, you get it refunded.
Applies from €10,000 of monthly AI spend
Step 2 · Measurement
10 % of your annual AI spend, €24,000 minimum
Inside your environment it is measured rather than estimated. The cost picture is credited in full if you commission within eight weeks.
What that means for you
| Annual spend | Measurement |
|---|---|
| €180,000 | €24,000 |
| €500,000 | €50,000 |
| €1,000,000 | €100,000 |
€
- Cost picture, one-off
- €3,000
- Measurement
- €102,000
- Savings we guarantee to find
- €9,000
The cost picture and the guarantee do not scale with usage. The guarantee applies from €10,000 of monthly AI spend.
Fixed prices instead of timesheets. We earn when you win.
You know before every step what it costs. Analysis, prototype, implementation. All at a fixed price. New scope means a new mini-analysis and a new fixed price, not a quiet addendum. And because we can measure the benefit, we offer a success component if you want one: after twelve months, measured against your own baseline, with Tokenometrics as the shared data basis. The metric, the method and the attribution are agreed up front, not negotiated afterwards.
- Fixed price
- Kill criteria
- Success component
And after that?
Tokenometrics is the platform you keep measuring with once we are gone.
Handover means handover: operations, documentation, runbooks and training sit with your team afterwards.
Where do we start with you?
Tell us where you stand. Whether the bill has been running for a while or you are still before the first step. I'll get back to you within 24 hours.
