My opinion, subject to change, on how AI is being priced today.

Value-based pricing involves setting prices primarily based on the perceived or estimated value of a product or service to the customer, rather than solely on the cost of production or historical prices. This is what we should expect, especially as we see that inference costs continue to plummet as shown in Mary Meeker’s Trends – Artificial Intelligence report from May 2025.

The cost of applying/using these models – known as inference – is falling quickly. Hardware is improving – for example, NVIDIA’s 2024 Blackwell GPU consumes 105,000x less energy per token than its 2014 Kepler GPU predecessor. Couple that with breakthroughs in models’ algorithmic efficiency, and the cost of inference is plummeting.

When pricing an AI agent based on the customer perceived value, three areas come to mind:

• automation in the form of time & cost savings
• client experience, seen as an increase in revenue or decrease in customer churn, and
• optimization, such as increasing throughput in a process or manufacturing line

Some observations about how Microsoft has priced things so far, generally AI features in premium SKUs like Dynamics 365 have been included at no additional cost to subscribers, whereas stand-alone AI products, such as Copilot for Finance are available at an additional charge.

Initial Impressions of Dynamics 365 Agent Pricing

Transform work with autonomous agents across your business processes – Microsoft Dynamics 365 Blog

With the introduction of agents, which automate tasks and optimize processes, Microsoft sees them more as a human intern, following a script, performing tasks in enterprise software, like a normal user would. So it does not surprise me that the licensing model is following how you’d normally license a human. Similar to how you’d assign a human a license for M365 or D365, Microsoft expects the agent to have the proper licensing. You pay your intern a wage, and you pay your agent in Copilot Studio Messages.

• Intern = Agent
• Salary = Copilot Studio messages
• Call Script or Instructions = AI Prompt / Instructions
• Both have an Entra identity and Microsoft licenses

Challenges posed so far is to understand how many Copilot Studio messages (CSM) would be used by a particular task. CSM in particular don’t represent just the messages a chatbot would send you, it represents the compute effort the agent took; for that reason it can be confusing. Current wisdom is to deploy or test the agent in a limited fashion to create a benchmark and to also use the estimator. Additionally there should be some governance in place to monitor the usage of the agents and the feedback being provided, to enhance quality.

Copilot Studio agent usage estimator (preview)

Challenges for Agent Pricing in General

There are things that true agent offerings and marketplaces need to work on:

• AI use cases can be hard to justify, and that means the perceived value of an agent is hard to pin down.
• If you are a software vendor, an MCP Server would enable any LLM to access your system… those need to be locked down, scaled according to usage, and therefore might need to be monetized for SaaS applications. I believe these features should be part of the cost, in other words, if an intern can access a web site, an LLM agent should be able to access an MCP Server (API).
• Agent as a Service (AaaS) sounds great: connect your data, connect your system, we automate the process, however it brings up questions of security and data replication that need to be addressed. Then do you price based on users, storage, or per transaction that was automated.

Feedback

What challenges and questions do you have around AI agents and how they are priced?

Original Post https://calafell.me/reflections-on-agents-and-pricing/