What actually decides the cost of an AI app

"AI app" covers two very different builds. One is a normal web app with one AI feature that does one job. The other is a product built around agents, retrieval, or answer-quality checks. The demo can look the same. They are not the same amount of work.

Here is how I break it down when I scope one.

The two versions

1. One AI feature. A web app with accounts and one focused AI feature: drafting, classification, search, or a bounded assistant.

2. Agents or retrieval. An app where agents, retrieval over your data, or checks on answer quality are the core of the product.

In my scope estimator, the first one lands at $6,500 to $9,000 over about 4 to 5 weeks. The second gets no number at all until a written scope splits it into sprints, each with its own fixed price. Those are estimates from fixed rules, not quotes. They assume one user type with sign-in, one or two outside services such as a model API, standard customer data with nothing regulated, a public launch, and hosting and model usage on your own accounts.

Why one AI feature moves the price

In the estimator, any AI feature moves the work from the $3,500 to $5,500 tier to the $6,500 to $9,000 tier on its own.

The reason is that model output varies. A form or a list view does the same thing every time. An AI feature needs work a standard screen does not:

• prompt work
• testing against real inputs, not just a sample
• limits on cost and on failure cases

Agents and retrieval go further. They depend on your data, your quality bar, and how much a wrong answer costs you. That is why they get scoped before they get priced.

What pushes the cost up

• Agents, retrieval, or answer-quality checks
• Several outside systems or a custom integration
• Health, financial, or other regulated data
• Web plus a mobile app or an admin console
• Many users at once, with queues and retries
• Starting inside an existing codebase

What you can cut for a first version

• One AI feature that does one job
• One user type and one or two core workflows
• A small private beta before a public launch
• A Figma file or clear written spec up front
• The second AI feature left for the next sprint
• Starting fresh, not in an old codebase

Model usage is a separate bill

The app runs on your own hosting and model accounts, so the providers bill you for usage directly. A build price covers building and launching the app, not the tokens it uses afterwards. Ask about this whoever you hire.

Check this before paying anyone

For AI work, check who owns the prompts, the evaluation set, and the model accounts, whichever way you build. If those sit in someone else's account, you do not fully own the product.

Cheaper ways to start

If the full build is more than you want to spend up front, a paid scope is the smaller first step: a one-page scope, a fixed price, and a first prototype of the core flow. With me that is $500, credited to the build if you go ahead within 30 days. If you already have an AI-built app that broke, a diagnosis costs less than a rebuild.

The full breakdown for AI apps, with FAQs, is here: harjotrana.com/cost/ai-app. If you want a number for your own idea, the free scope estimator uses the same rules.

Happy to answer questions in the comments.