What business model for AI in software?
Arthur Mensch, founder of Mistral, told the French Senate: at Mistral, AI consumption already represents 10% of payroll. Extrapolated to all of Europe, 10% of European payroll represents roughly €1 trillion per year. That's the business AI market within 3-4 years. For software publishers, the question is no longer "should we integrate AI?" but "how do we capture this spend?". Here's how to structure the monetization.
Jean de Bodinat
CEO & Founder, Rakam AI — March 25, 2026
Conviction
The virtuous cycle of AI monetization
The more useful AI is, the more indispensable the software becomes, the more the vendor can monetize. This is not theory — it is the mechanism driving the best SaaS vendors on the market.
Integrated AI
The vendor deploys AI features in its product
User productivity
Users save time and do more with less
Software value
The product becomes indispensable, churn drops, retention rises
Revenue captured
New AI revenue reinvested to improve the product
"AI is not about POCs, it is about profitability. The question is not 'should we integrate AI?' but 'how do we make it profitable as fast as possible?'"
Monetization
The 3 AI monetization models
Each model has its strengths and limitations. The right choice depends on your market, your customers' maturity and your cost structure. To understand the different technical integration tiers, see our article on the 3 levels of AI integration in a SaaS.
Usage-based revenue
Billing per token, per request or per completed task. The price is indexed to the value delivered. The customer pays when they use it — perfect alignment between revenue and costs.
Principle
The token is indexed to the margin created for your customers. You price on value, not on cost.
Advantage
Scalable, transparent, incentive-aligned. The more the customer uses, the more they pay — and the more value they get.
Risk
Unpredictable revenue if adoption is slow. Hard to forecast ARR over 12 months.
Example: Salesforce just launched Flex Credits for Agentforce — $0.10 per action (1 action = 20 credits), sold in packs of 100,000 credits at $500 (i.e. 5,000 actions). No longer priced per conversation, but per real action in the CRM. A major pivot in agent monetization.
Premium subscription
The AI module is included in a higher tier. The vendor secures monetization via a stable, predictable ARR. It is the classic CAPEX model applied to AI.
Principle
Fixed monthly or annual license. AI is an upgrade argument, not a separate product.
Advantage
Recurring, simple, predictable. The CFO loves it: ARR is smooth and forecastable.
Risk
The customer does not always perceive the value. If usage is low, they challenge the premium.
Example: ClickUp Brain at $7/user/month. AI is integrated into the premium plan, not as an option.
Hybrid model
A subscription base that secures ARR, combined with a usage layer that captures incremental value. This is the model we recommend at Rakam: the best of both worlds.
Subscription base
Guarantees predictability. The customer pays a fixed monthly fee to access basic AI features.
Usage layer
Captures the upside. Beyond a threshold, each additional request is billed on usage.
Example: Claude Code offers standard and premium licenses with different usage tiers — a model that's easier to sell to enterprise accounts who want predictability in their AI budget. This is the model we recommend for business agents.
"Combine usage-based revenue and standard subscriptions to boost your EBITDA."
Costs & Optimization
The cost structure of AI
Monetizing AI without controlling its costs is building on sand. Here are the cost items to anticipate and the levers to optimize them.
Costs to anticipate
Inference
~€0.01 to €0.10 per request depending on the model. This cost drops 40 to 60% per year, but will never be zero. Model it from the start.
Cloud infrastructure
VectorDB, streaming, monitoring, advanced cloud architecture. AI demands reliability and scalability that traditional SaaS does not require.
Specialized engineering
ML engineers, data engineers, evaluation and quality. Talent is scarce and expensive — this is often the most underestimated line item.
Optimization levers
Smaller models
Use Small Language Models for simple tasks (classification, extraction). Reserve LLMs for complex reasoning.
Intelligent caching
Frequent responses are cached. Up to 80% savings on repetitive requests with no loss of quality.
Continuous evaluation
Measure quality and cost per request in real time. Avoid cost overruns by routing requests to the optimal model.
Off-the-shelf agents
Favor pre-built agents before developing custom ones. Cheaper, faster, already evaluated.
Framework
The AI-EBITDA equation
AI-EBITDA = (AI revenue generated) − (Inference + infra + engineering costs)
When this number is positive, AI is no longer a cost center but a margin lever. Vendors who understand this build a lasting competitive advantage: every euro invested in AI generates more than a euro of revenue.
+
Positive AI-EBITDA
AI is a product. You sell it, it contributes to the margin.
=
Neutral AI-EBITDA
AI strengthens retention without generating direct margin. Acceptable short-term.
−
Negative AI-EBITDA
AI is a cost center. The model or costs need to be revised.
Pricing
How to set your AI pricing
AI pricing is not a theoretical exercise. It is an iterative 4-step process, guided by data.
Measure the productivity gain
How much time does AI save the user? How much money? Measure the before/after delta on a specific workflow. Without this data, you cannot price on value — you are just guessing.
Capture 10 to 30% of the value created
Rule of thumb: your price should represent 10 to 30% of the customer's gain. If your AI saves a customer €10,000/month, you can charge between €1,000 and €3,000/month.
Customer gain
€10,000
Your price (10-30%)
€1,000 — €3,000
Test usage vs. subscription
Run an A/B test on a customer segment. Offer both models (pure usage vs. premium subscription) and measure adoption and willingness-to-pay. Data will settle the debate better than any strategy meeting.
Iterate with data
AI pricing is not fixed. Adjust quarterly based on actual usage, customer satisfaction and the evolution of inference costs (which are dropping fast). The right price today will not be the right price in 6 months.
Concrete cases
How leaders monetize AI
Real examples, not projections. Here is how vendors are already monetizing their AI features.
Qonto
Task-based AI agents for 600,000 businesses. Usage-based model: each AI-processed operation is billed. The customer pays proportionally to what they consume.
Pennylane
ComptAssistant is included in the premium subscription. AI is an upgrade argument, not a separate product. It strengthens the value proposition of the higher tier.
Our Rakam clients
Support agent at €500/month (fixed subscription) + BI agent on a setup fee + subscription. The hybrid model in action: predictability for the customer, upside for the vendor.
To understand how these vendors positioned their overall AI strategy, see our article on how to position your SaaS in the AI era.
Conclusion
AI is a product, not an R&D project
Vendors that treat AI as a research project will never monetize it. Those that treat it as a product — with pricing, a cost structure and a margin target — build a lasting competitive advantage.
The virtuous cycle is simple: AI makes your users more productive, your software becomes indispensable, and you capture new revenue. The hybrid model (subscription + usage) is the best starting point for most vendors.
This is exactly what we do at Rakam with our services and our products. We help software vendors go from an AI cost center to a positive AI-EBITDA. In 3 months, not 3 years.
Build your AI business model
30 minutes between executives. Together we analyze your cost structure, monetization levers and AI-EBITDA potential.
Book a meetingJean de Bodinat
CEO & Founder of Rakam AI. Helps software vendors turn AI into revenue.
Go from buzz to real impact
Leave us your details, we will get back to you within 24 hours.