Frugal AI: controlling the AI bill in production

−74%annual CRM agent cost
−65%monthly reporting cost
−73%carbon footprint
−94%on a regulatory search

01 · SITUATION

The context 

3 AI agents in production: a CRM banker assistant (pre-meeting client summary), a KYC document analysis engine, a fund report generator (220 funds, 387 share classes, 5 languages, monthly). The cost had never been broken down, token consumption never quantified, carbon footprint invisible.

02 · METHOD

5-dimension frugal audit 

TCO Framework

Breakdown: build / run / decommission. OPEX represents 60-70% of total cost. That is where real gains happen.

Cost axis

'Right model for the right task': most calls don't need the most powerful model. Each use case is reassessed and steered towards the most frugal option that holds the required quality bar.

Token axis

Consumption is audited line by line: redundant queries, documents reprocessed in loops, chatty architectures. The cheapest tokens are the ones you never consume. The most underestimated lever.

Carbon axis

For the same service delivered, footprint varies widely depending on where and how workloads run. The audit identifies possible shifts (with no impact on quality or deadlines).

Cost governance

Making the gains stick: budget alerts, a cost, token and carbon dashboard per agent, validation rules. So the bill doesn't creep back up six months after the audit.

03 · RESULTS

Quantified gains 

−74% annual cost on the CRM assistant
−65% monthly cost on fund reporting
−73% carbon footprint on reporting
−90% tokens on 80-page document analysis
−94% cost on a regulatory search
Estimated figures, documented audit (AFNOR Spec 2314, OpenAI Pricing 2025)

Next case

CV agent: from manual to production

An AI bill to analyze?

Let's book 30 minutes.

Frugal AI: controlling the AI bill in production — Diane Maurin