Adding AI to an app has two costs. The build cost depends on how much surrounding software is needed: data access, review screens, and safeguards. The running cost is billed by the provider on the volume of text processed, so it rises with the number of requests and the length of each one.
When we built Ask Tamou, a website chat for a law firm, it replaced a third-party chat widget and its monthly bill. That swap is a fair picture of how AI costs work. You trade a fixed subscription for two different things: a one-time build and a usage bill that follows how much the feature is used.
We will not put dollar figures on model usage here. Providers publish their own rates and change them, and any number we printed would go stale. What stays true is the shape of the cost and what pushes it up or down.
What are the two costs of adding AI to an app?
The two costs of adding AI to an app are the build cost, paid once to design and develop the feature, and the running cost, paid continuously to the model provider based on usage. A feature can be cheap to build and costly to run, or the reverse.
They are driven by different things, which is why they need to be estimated separately.
Build cost is developer time. It covers connecting your data to the model, writing and testing the instructions, building the screen where output appears, and adding the review and safety steps.
Running cost is what the provider charges. With an API model you pay for the amount of text sent in and the amount sent back, usually measured in units called tokens. Larger, more capable models cost more per unit than smaller ones.
How is AI usage billed?
AI usage through an API is billed by volume of text: the text you send to the model, called input, and the text it returns, called output. Each request is metered. More requests, longer documents, and longer replies all increase the bill, and more capable models charge more per unit.
A few consequences follow from that.
- Long inputs add up. Sending a full file with every request costs more than sending the relevant page.
- Conversations grow. In a chat, earlier messages are usually sent again with each new one so the model has context, so long conversations cost more per reply than short ones.
- Agents multiply calls. A feature that takes many steps makes many requests for one task. We explain why in what are AI agents.
- Idle features cost nothing. Unlike a subscription, there is usually no usage charge when nobody uses it.
What drives the cost of an AI feature?
The cost of an AI feature is driven mainly by how much work surrounds the model: getting the right data to it, checking what comes back, and presenting it for review. The model call itself is a few lines of code. The safeguards and the interface are most of the build.
| Driver | Affects | Why it matters |
|---|---|---|
| Number of distinct tasks | Build | Each task needs its own instructions and testing |
| Access to your data | Build | Reading from an existing system can be simple or a project by itself |
| Review and approval screens | Build | Someone has to see and correct the output |
| Accuracy requirements | Build | Higher stakes mean more grounding, checks, and testing |
| Integrations with other systems | Build | Each connection is designed, built, and maintained |
| Requests per day | Running | Each one is metered |
| Length of input and output | Running | Billing follows volume of text |
| Model size | Running | More capable models cost more per unit |
| Steps per task | Running | Multi-step features make several calls |
| Monitoring and upkeep | Both | Instructions and checks need adjusting as real use reveals gaps |
AI is often one part of a larger product. Regulars, our loyalty product for shops, lists a model API in its stack alongside a kiosk, an owner dashboard, and a campaign scheduler. Most of what was built there is ordinary software, and that is typical.
What kind of AI project fits each budget?
Our contact form uses four broad budget bands, and AI work can fall in any of them depending on scope. A single feature added to existing software sits at the low end. A new application built around AI, with accounts, data, and integrations, sits at the high end.
| Budget band | What typically fits |
|---|---|
| Under $5,000 | One narrow AI feature added to software that already exists, with data that is easy to reach |
| $5,000 to $15,000 | A self-contained tool, such as a website chat with human handoff, or a drafting feature with a review screen |
| $15,000 to $50,000 | Several AI features inside a custom application, with integrations and stored records |
| Over $50,000 | A full product where AI is one part of a larger system with multiple user roles, payments, or compliance needs |
These are broad bands, and every project is quoted on its own. They describe the build only. Running costs are separate and go to the model provider. For how the bands apply to software in general, see how much a custom web app costs.
How do you keep AI running costs down?
You keep AI running costs down by sending less text, calling the model less often, and using the smallest model that does the job well. Most overspending comes from sending whole documents when a section would do, or from calling a model for work plain code could handle.
- Send only what the task needs. Trim the input to the relevant records or passages.
- Store results. If a document was summarized once, save the summary instead of generating it again.
- Match the model to the task. Simple sorting does not need the most capable model.
- Limit output length. Ask for a short answer when a short answer is enough.
- Keep rule-based steps in code. See when to use AI and when plain code is better.
- Set a spending alert with the provider so a bug or abuse does not run unnoticed.
- Protect public features. A chat open to the internet needs rate limits.
Frequently asked questions
Is there a monthly fee for AI in my app?
With an API model there is typically no flat fee. You pay the provider for what is used, so the bill varies month to month with activity. Hosting for the app itself is a separate, ordinary cost.
Who pays the AI provider, me or the developer?
We recommend the business holds its own provider account and pays it directly. You see the usage, you control the keys, and the feature keeps working whoever maintains the code.
Is building custom cheaper than an AI subscription tool?
Not automatically. A subscription is the better deal when an existing tool already does what you need. Custom is worth it when the feature must work inside your own software or with your own data. We cover that in build vs buy.
Can I estimate running cost before building?
Roughly. Count expected requests per day, estimate the text in and out for each, and apply the provider's published rate. Testing with real examples gives a better figure than guessing.
Do AI costs go up as my business grows?
Usage cost follows usage, so more requests mean a larger bill. The build cost does not repeat. You can keep the running cost in check by sending the model only the text it needs and by using plain code for the steps that do not need judgment.
If you can describe the task and roughly how often it happens, send it to us and we will tell you which cost will dominate.