The GPT-5 demo last week seemed to be mostly about incremental development, as opposed to a step-change leap into the future. Sure, people poked fun at the bizarre charts and Bernoulli Effect Blooper in the demo, but analysts like Gartner gave it high marks for its new coding functionalities and substantially lower hallucination rates (though I’m still not sure I’d trust my health or my finances to a system with a roughly 10% error rate).
And of course, everyone is still trying to figure out how to use this stuff to build real tools that solve real problems and make real money. But a fact’s a fact: the AI race is on. Hundreds of new AI-enabled services are doing things that weren’t possible six months ago, and the companies behind them seem to be chasing one thing — outcome-based pricing.
Why? Because it’s one way to navigate the “Impossible Triangle” problem I discussed last week. But what exactly constitutes an “outcome”? Is it X? Is it Y? It’s an admittedly squishy word. Thanks once again to Michael Mansard of the Subscribed Institute, however, we can break it down simply. There are four common ways agentic AI companies are defining outcomes in order to determine their pricing:
1. Per Agent
Think of this like hiring an employee. The customer is paying for the agent to be available — whether it’s working 24/7 or just sitting idle. This pricing model works well when customer outcomes are diffuse and long-term. It’s predictable and easy to model, but it puts more risk on the customer. If the agent doesn’t deliver, they still get paid. This is used when the AI is replacing a broad, ongoing role — like an assistant or security engineer. For example, Nullify charges $800 per agent per year to fix security vulnerabilities. You’re licensing the AI, not paying per task.
2. Per Activity
This is the “metered” model — you pay when the AI does something: answers a question, writes some code, runs a process. Each action is small, but they add up to real productivity. This model makes sense when the work is granular and frequent. It’s good for customers who want to pay only when the AI is actually doing something. For example, Devin charges per “Agent Compute Unit” for its software engineer assistant services while Microsoft Copilot charges by “Security Compute Unit.” It’s all about resource use per action.
3. Per Output
Here, you pay for what the AI produces — a finished product, not the steps to get there. A document, a resolved ticket, a snapshot, a completed conversation. This makes sense when the customer cares about results, not activity, and when the output is tangible and measurable. They want to pay for deliverables, not time. For example, Replit charges $0.25 per “checkpoint” (a meaningful change in code), while Salesforce will charge you “$2 per customer service interaction.”
4. Per Outcome
Here, you only pay when the AI achieves a real business case result: a resolved issue, a cost saving, a new lead, a sale. In this case, the AI can be directly tied to strategic impact that can be clearly measured and agreed upon. You’re sharing in the upside with your customer. For example, Vantage charges 5% of actual savings on AWS storage costs delivered, or Zendesk charges only when a support ticket is fully resolved by its agent.
So, which model is best? You would be forgiven for jumping straight to number four, the “Holy Grail” of business case outcomes. But you’d be wrong. It actually represents less than 10% of the roughly 60 agentic AI services that Michael has studied. Why? Because defining (and tracking) this stuff materially is still very complicated — just look at the massive amounts of documentation that Zendesk needs just to define a “resolution.”
The truth is that there’s no perfect answer — just trade-offs. Some models are easier to explain. Some align better with customer value. Some give you more cost control. Some services use a combination of pricing methods. They all have their strengths and weaknesses. But they’re all being used to find the latest answer to the eternal question: What are your customers really paying for — and does that match how they experience value?
In order to bring the COMPASS Agentic AI Pricing Framework to life, next week we’re going to be looking at these four types of outcomes through the lens of attributability. It’s a process akin to deciding if your solution is more like an executive assistant who keeps things running smoothly behind the scenes (diffuse), a marketing campaign manager whose work contributes meaningfully to outcomes but relies on others to deliver results (medium), or a sales rep whose sourced and closed deals are directly tracked and credited (strong).
If you’ve ever struggled to prove your agentic AI service is worth its price, this will give you the blueprint. Stay tuned.
