What Is Agentic AI — and How Should You Price It?

“Agentic AI” gets used for everything from a single chatbot to a full autonomous team, so the term is worth pinning down before you try to buy it. An agent, in the useful sense, is an AI system that doesn't just answer a question — it plans a small sequence of steps, calls tools or APIs, checks its own output, and produces a finished artifact: a report, a working script, a first draft of an integration. The difference from a chatbot is the same difference as between a search engine and an employee: one points you at information, the other hands you a result.

That distinction matters for pricing. A chatbot subscription is priced like software: a flat fee for access. Work delivered by an agent behaves more like labor: it varies by task, some tasks are trivial and some are substantial, and the buyer has no reliable way to tell which is which before the work starts. Most vendors solve this the way agencies always have — an hourly rate, billed after the fact, with an estimate that's really a guess.

Why hourly billing breaks down for AI work

Hourly billing assumes the buyer can't see inside the work, so they're trusting the seller's clock. That trust problem gets worse with AI, not better: a task that takes a model four seconds of inference time isn't worth per-minute billing, and a task that requires five iterations of tool calls and human review isn't worth a flat per-request fee either. You end up either overpaying for simple requests or underpricing complex ones — and you find out which one happened only after the invoice arrives.

Standard-hours: pricing the work, not the clock

A standard-hour is a normalized unit of AI-delivered work, not a measurement of wall-clock time. An estimator — in our case, an AI router that has seen the shape of similar requests — scores a task's actual complexity (integration work, data volume, ambiguity, review needed) and converts that into a number of standard-hours, independent of which model or how much compute it takes to deliver it. That number gets a price attached to it before any work starts, using a fixed rate per standard-hour from your plan.

• You see the price before you commit, not after.
• The price doesn't change if the vendor's underlying AI provider gets more expensive or slower.
• Two similar tasks cost similarly, even if one happened to need more model calls behind the scenes.

The other piece worth having is an alternative to hours altogether: paying a share of a measured effect (time saved, revenue recovered, cost avoided) instead of a rate. That only works if the first deliverable comes with a real metric attached — a baseline and a target, not a vague promise — which is why it should be tied to an actual working prototype, not a proposal.

If you're evaluating an AI vendor and the pricing conversation starts with “it depends,” ask them to show you the number before the work starts, in a unit you can compare across tasks. That's the whole idea behind standard-hours.