What an in-app AI agent really costs

A bias to declare: I sell an in-app agent, so I benefit if you buy. The build list below is everything I had to build for Verb, which is why I can write it down; price it yourself and decide.

Cost one: the model, which is the small part

Every request your users make becomes one or more model calls. The bill is input tokens times the input price plus output tokens times the output price, per call, times calls per request, times requests a month. An agent that uses tools usually makes several calls per request, because it reads the request, calls a tool, reads the result and sometimes calls another.

On current prices that lands at a fraction of a cent per request on a small model and a few cents on a large one. The LLM API cost calculator works it out for Claude, GPT and Gemini side by side from your own numbers. For most products it is the smaller line on the bill.

Cost two: everything after the first demo

The demo takes days: a model, a prompt and a tool that calls your API. What takes months is making it safe to put in front of every user. This is the list, each item something Verb needed before it could ship:

• Identity. Knowing who is asking from a token your server signs, never from the browser's word, and forgetting them when they sign out.
• Acting as the user. Calling your API with the signed-in person's own access instead of a master key, so the agent cannot exceed them.
• Confirmations. Showing the real values before any change and waiting for a yes.
• Honest results. Never reporting an action as done when it failed or the result is unknown.
• Abuse protection. Limits for visitors who are not signed in, a bot check, daily caps, defence against prompt injection.
• An audit trail. Who asked, what ran, with what, and what came back.
• A test environment that runs real actions against development data, never production.
• An off switch that takes it off every open page at once.
• The widget itself, working inside other people's CSS and Content Security Policies, on phones, without breaking their page.
• Keeping it working as your product, your API and the models change.

Build or buy

Build when the agent is your product: it is what you sell, it is your differentiator, and you want to own every line. The build list is then the work, not overhead. The CopilotKit alternatives post compares the toolkits for that path.

Buy when the agent is a feature that helps users get their work done in a product that is about something else. You get the list on day one and pay per use or per month. Why most in-app AI agents never ship is the longer argument.

The verdict

The model is cheap and getting cheaper; the engineering around it is what you are really pricing. With Verb there is no build bill: it starts with a free 30-day pilot, and the price after it is agreed from what the agent measurably does for your users. How to measure an AI agent pilot covers those numbers.