AI area
AI runs on an open budget.
We hand you the bill.
73% of companies overshoot their AI cost plan, and agentic projects overshoot it by 2.4x on average. Almost nobody knows the return, because nobody puts tokens in the denominator.
The problem in numbers
Nobody puts tokens in the denominator
Licences show up on the invoice, tokens don't. And the code AI produces carries a maintenance cost nobody is counting. The result is a budget you discover after the fact.
- 73%
- Companies overshooting their AI cost plan
- 2.4x
- Average overshoot on agentic projects
- $200-600
- Real cost per developer per month
- 2.5-3.5x
- Healthy ROI, but only if you count tokens
Public sources: FinOps Foundation State of FinOps 2026, GitClear study on 153 million lines, 2026 productivity benchmarks.
Diagnosis
2 weeks
Free
The AI bill
Real spend — licences, tokens, GPU — broken down by team and feature, share of AI-generated code and what it costs to maintain, top 5 wastes in €/year.
Pilot Build
2-3 weeks
AI unit economics
Every call tagged by team, feature and environment on a narrow scope: you go from «we spend a lot» to «this thing costs X».
Build
6-8 weeks
Token governance
Full attribution, model right-sizing, routing, caching, budgets and alerts, usage policy.
Stewardship
Retainer
AI watch
Continuous monitoring of cost, quality and risk. Monthly report, quarterly review.
Inside your perimeter
The AI passes compliance, or it doesn't ship
Regulated companies are not afraid of AI: they are afraid of where the data ends up and of who answers when it gets it wrong. Three precise fears, each met with a written guarantee and a deliverable with a name.
Local models
Open-weight models in your infrastructure or in your European cloud tenant. No prompt, document or line of code leaves the perimeter for an outside API. The model is picked per task, not by habit, and where a local model is enough that is the one we use.
Deliverable: deployment architecture and register of the models in use, with where the data sits for each.
GDPR and privacy by design
Impact assessment before the first prompt, minimisation of the data that enters the model, defined logs and retention, roles and legal bases written together with your DPO. Privacy is in the design, not in an annex signed afterwards.
Deliverable: DPIA and updated record of processing, ready for the DPO.
AI Act
Risk classification of every system, transparency duties, human oversight where the regulation asks for it, an audit trail of every decision. Most obligations apply from 2026: no longer a plan, a requirement.
Deliverable: register of AI systems and a compliance file for each.
What you get
From "we spend a lot" to "this thing costs X"
Every item is a named deliverable you can put in a contract and check on delivery.
AI spend inventory
Licences, tokens, GPU and infrastructure in a single statement, including the share that appears on no invoice.
Attribution by team and feature
Every call tagged by team, feature and environment. The argument ends and the numbers begin.
Cost per unit of value
Cost per resolved ticket, per active user, per delivered feature. Numbers a CFO reads without translation.
Model right-sizing and routing
The right model for each task, with caching and fallbacks. On targeted workloads, documented reductions of 40-80%.
Budgets, alerts and usage policy
Spend caps per person and per team, alerts before the overrun, and written rules on what can be used and how.
Generated-code quality
Separate measurement of AI-written code: how much survives, how much gets rewritten, what it costs to maintain.
Questions we get
What people ask before signing
- No. We work with what you have: Claude Code, Copilot, Cursor, direct API calls. We add measurement around them, we don't replace them.
Contact us
A quote, a particular request, or simply a coffee to meet us? Write to us and we will reply to a "nano-second".