A machine learning agency in Toronto built a fraud detection model for a fintech startup. The engagement was priced at $280,000 — a solid project. The client loved the results. Then the client got acquired by a larger financial services company that wanted to deploy the same model across 14 subsidiaries. The agency had structured the original deal as a work-for-hire arrangement where the client owned the model outright. The acquiring company deployed the fraud detection model across its entire portfolio without paying the agency another dollar. The agency estimated the missed revenue at over $2 million.
Compare that with a competing agency in London that delivered a similar fraud detection model to a European bank. They structured the deal with a base license fee for a single deployment, per-subsidiary expansion fees, and a retraining license for model updates. When the bank expanded the model to three additional business units, the agency earned $540,000 in additional licensing revenue — on top of the original engagement fee.
Same type of work. Same level of technical effort. Radically different business outcomes. The difference was not the quality of the AI model. It was the licensing structure.
AI model licensing is one of the most consequential business decisions an agency makes, and most agencies get it wrong. They either give away too much (work-for-hire everything) or lock down too aggressively (scaring clients away with complex licensing terms). The sweet spot is a licensing framework that protects your IP, creates recurring revenue opportunities, and gives clients the flexibility they need.
Why AI Model Licensing Is Different from Software Licensing
Traditional software licensing is well-understood. You license the code, the client runs it, and the terms define how many users or instances are covered. AI model licensing is fundamentally different for several reasons.
Models are not static artifacts. A software license covers a defined product. An AI model is a living system that changes when retrained, fine-tuned, or updated. Your licensing needs to address the model's lifecycle, not just a point-in-time delivery.
The value is in the training, not the code. The code that defines an AI model architecture might be a few hundred lines. The value is in the training process — the data curation, feature engineering, hyperparameter optimization, and iterative refinement that produces a high-performing model. Your licensing should reflect where the value actually sits.
Models depend on infrastructure and context. An AI model does not exist in isolation. It requires inference infrastructure, data pipelines, monitoring systems, and operational support. Licensing the model without addressing its operational context creates gaps.
Derivative works are common. Clients often want to fine-tune, retrain, or adapt models for new use cases. Your licensing needs to define whether and how derivative works are permitted.
Foundation model dependencies complicate ownership. If your model is built on top of a third-party foundation model, your licensing is constrained by the upstream license terms. You cannot grant rights you do not have.
Licensing Models for AI Agencies
Model 1: Perpetual License with Support
The client pays a one-time fee for a perpetual license to use the AI model, plus optional ongoing support and maintenance fees.
How it works:
- Client pays a project fee that includes model development and a perpetual usage license
- The agency retains ownership of the model IP
- The client receives a license to use the model for defined purposes within defined scope
- Support, retraining, and updates are sold as separate ongoing services
Advantages:
- Clear value proposition for clients — they pay once and own the right to use the model indefinitely
- Creates ongoing revenue through support and retraining services
- Agency retains IP and can license similar models to non-competing clients
- Aligns with how many enterprises budget for technology (capital expenditure for the license, operational expenditure for support)
Disadvantages:
- Lower upfront revenue compared to subscription models if the client does not purchase support
- Risk of clients running models without maintenance, leading to degraded performance and reputational damage
- Requires clear definitions of "use" to prevent scope creep
Best for: Agencies serving enterprise clients who prefer to own perpetual rights, particularly in regulated industries where vendor dependency is a concern.
Model 2: Subscription License
The client pays ongoing fees for the right to use the AI model, with the license terminating when payments stop.
How it works:
- Client pays a recurring fee (monthly or annual) for model access
- The fee includes model usage, standard updates, and defined support
- The agency retains full ownership and control of the model
- License terms renew automatically unless cancelled
Advantages:
- Predictable recurring revenue for the agency
- Natural alignment with ongoing model maintenance and improvement
- Lower initial cost for clients, reducing sales friction
- Agency maintains control of the model and can push updates
Disadvantages:
- Some enterprise clients resist subscription models for core business systems
- Client lock-in concerns can slow deal closure
- Revenue recognition is spread over time, affecting cash flow
- Client data and model portability questions arise at termination
Best for: Agencies offering AI-as-a-service, managed AI solutions, or products where continuous model improvement is a core value proposition.
Model 3: Usage-Based Licensing
The client pays based on actual usage of the AI model — per prediction, per API call, per document processed, or per active user.
How it works:
- Client pays a base platform fee plus variable costs based on usage metrics
- Usage is metered and billed at defined intervals
- Pricing tiers may offer volume discounts
- The agency operates the model and manages infrastructure
Advantages:
- Directly aligns agency revenue with client value
- Low entry barrier for clients — they pay as they grow
- Natural scaling — as the client's usage grows, agency revenue grows
- Easy to justify ROI because cost correlates with utilization
Disadvantages:
- Revenue volatility if client usage fluctuates
- Requires robust metering and billing infrastructure
- Clients may optimize usage to reduce costs, potentially underusing the model
- Complex pricing negotiations for high-volume clients
Best for: Agencies delivering AI products with measurable, quantifiable outputs — document processing, content generation, classification services, prediction APIs.
Model 4: Tiered License
The client selects a licensing tier based on deployment scope, usage volume, or feature access.
How it works:
- Define clear tiers (Starter, Professional, Enterprise) with specific inclusions
- Each tier has defined limits on users, deployments, API calls, or features
- Clients upgrade tiers as their needs grow
- Custom enterprise tiers are available for large deployments
Advantages:
- Simple pricing that clients can understand and budget for
- Natural upsell path as clients grow
- Predictable revenue within each tier
- Reduces negotiation complexity
Disadvantages:
- Clients may feel constrained by tier boundaries
- Pricing cliffs between tiers can frustrate clients near tier boundaries
- Requires careful tier design to avoid leaving money on the table or pricing out smaller clients
Best for: Agencies with standardized AI products that serve a range of client sizes.
Model 5: Revenue Share or Outcome-Based Licensing
The agency's compensation is tied to the business outcomes the AI model produces.
How it works:
- Agency receives a percentage of revenue generated or costs saved by the AI model
- Base fees may apply to cover development costs
- Outcome metrics and measurement methodology are defined in the agreement
- Revenue share continues for the duration of model usage
Advantages:
- Directly aligns agency incentives with client outcomes
- Can produce significantly higher total revenue than flat-fee models
- Demonstrates agency confidence in their model's performance
- Reduces client risk — they pay when the model delivers value
Disadvantages:
- Revenue depends on client's ability to measure and report outcomes
- Client may underreport outcomes to reduce payments
- Difficult to attribute outcomes solely to the AI model versus other factors
- Revenue is unpredictable and delayed
Best for: High-confidence engagements where the AI model's impact is directly measurable — revenue optimization, cost reduction, lead scoring, conversion optimization.
Key License Terms to Get Right
Scope of Use
Define exactly what the license allows. Ambiguity in scope of use is the most common source of licensing disputes.
Specify:
- Authorized users — Who within the client organization can use the model
- Authorized purposes — What business purposes the model can be used for
- Authorized deployments — How many instances or environments the model can be deployed in
- Geographic scope — Where the model can be deployed and used
- Subsidiary and affiliate rights — Whether the license extends to the client's subsidiaries or affiliates
- Duration — How long the license lasts
Modification and Derivative Works
Clients often want to modify, fine-tune, or build upon your model. Your license needs to address this explicitly.
Options:
- No modifications — Client uses the model as-is. Any modifications require agency involvement
- Limited modifications — Client can adjust configuration parameters, prompts, and thresholds but cannot modify model weights
- Full modification rights — Client can retrain, fine-tune, and modify the model. Agency retains ownership of the original model, client owns their modifications
- Derivative work licensing — Client can create derivative models, subject to defined terms and potential additional fees
Exclusivity
Should the client have exclusive rights to the model, or can the agency license similar models to other clients?
Non-exclusive (default): The agency retains the right to build and license similar models for other clients. This is standard and allows the agency to leverage its expertise across multiple engagements.
Exclusive within a vertical: The agency agrees not to license a directly competing model to competitors within the client's specific industry vertical. This commands a premium.
Fully exclusive: The agency grants exclusive rights to the model and agrees not to build similar models for any other client. This should command a significant premium — potentially 3x to 5x the non-exclusive price.
Audit Rights
For usage-based or outcome-based licensing, the agency needs the right to verify that the client is complying with license terms and accurately reporting usage or outcomes.
Include:
- Right to audit usage logs and deployment configurations
- Reasonable notice period for audits (typically 30 days)
- Client obligation to maintain accurate usage records
- Remedies for discovered underpayment or unauthorized usage
Termination and Post-Termination Rights
What happens to the model when the license ends? This is a critical question that many agencies fail to address.
Upon termination:
- Subscription licenses — Client must stop using the model and delete all copies
- Perpetual licenses — Client retains usage rights but loses access to updates and support
- Data obligations — Define what happens to client data used with the model
- Transition period — Provide a reasonable transition period for the client to migrate away from the model
- Survival clauses — Specify which terms survive termination (confidentiality, IP ownership, indemnification)
Pricing Your AI Model Licenses
Cost-Plus Pricing
Calculate your development costs and add a margin. This is the simplest approach but often undervalues your work.
Formula: Development cost + overhead + desired margin = license price
When to use: Early-stage agencies establishing pricing benchmarks, or for commoditized AI applications where market pricing is well-established.
Value-Based Pricing
Price based on the business value the model delivers to the client, not your cost to build it.
Approach: Estimate the annual business impact of the model (revenue generated, costs saved, risks mitigated). Price the license as a fraction of that value, typically 10-25% of first-year value.
When to use: Engagements where the business impact is quantifiable and significant. A fraud detection model that saves a client $5 million per year can justify a $500,000 to $1.25 million license.
Market-Based Pricing
Price based on what comparable AI solutions cost in the market.
Approach: Research pricing for comparable AI products and services. Position your pricing relative to the market based on your differentiation.
When to use: When there are clear market comparables for your AI solution.
Hybrid Pricing
Combine approaches for a pricing structure that balances cost recovery, value capture, and market positioning.
Example structure:
- Base development fee (cost-plus) — covers the agency's costs and a reasonable margin
- License fee (value-based) — reflects the business value of the model
- Usage fees (market-based) — competitive with market alternatives for variable costs
- Support and retraining fees (cost-plus) — covers ongoing operational costs
Protecting Your Licensing Revenue
Register your IP. While AI model weights may not be copyrightable in all jurisdictions, the training processes, architectures, and supporting code can be protected. Document your IP and consider trade secret protections.
Build technical enforcement. License terms are only as good as your ability to enforce them. Consider technical measures like license keys, usage metering, model encryption, and deployment verification.
Monitor for unauthorized use. Periodically verify that clients are using the model within the scope of their license. Usage-based licenses require robust metering infrastructure.
Include meaningful remedies. Your license should specify consequences for violations — additional fees for over-usage, termination rights for material breaches, and injunctive relief for IP theft.
Your Next Step
Review the last five AI models your agency delivered. For each one, identify the licensing structure used and estimate the total revenue earned. Then estimate what the revenue would have been under an alternative licensing model. If the gap between actual and potential revenue is significant, it is time to restructure your licensing approach.
Build a licensing menu with three to four standard licensing models. Define the terms, pricing methodology, and use cases for each. Train your sales team to present licensing options as a strategic choice rather than a negotiation battleground. The right licensing structure does not just protect your IP — it becomes a growth engine for your agency.
The Toronto agency gave away $2 million in potential licensing revenue with a single work-for-hire clause. The London agency built a recurring revenue stream from a licensing framework that took two days to design. The licensing structure you choose today determines the revenue you earn for years to come.