Adding a single AI feature built on a hosted model typically costs $10,000 to $40,000 (€8,500 to €35,000, £7,500 to £30,000) to build and a usage-based running cost that starts small and grows with users. Features needing custom data pipelines, fine-tuning or a review workflow cost $40,000 to $150,000 (€35,000 to €130,000, £30,000 to £110,000). Training your own model from scratch is rarely justified for a startup and costs far more. The hidden costs are collecting and cleaning data, building the evaluation that tells you whether the feature is good enough, and the human review step.
Prices are in US dollars. Euro and pound figures in parentheses are approximate, rounded conversions. Quotes are issued in the currency of your contract.
The cost of adding AI to a product has fallen sharply and become harder to predict, because it is now split between a build cost and a running cost that depends on how much users use it. This guide covers both, for the three common approaches, and the costs founders discover after the quote.
The three approaches
| Approach | What it means | Typical build cost | Typical running cost | Timeline |
|---|---|---|---|---|
| Hosted model, simple integration | Call a model provider's service with a well-designed prompt and your data | $10,000 to $40,000 (€8,500 to €35,000, £7,500 to £30,000) | Per-use billing, $50 to $2,000 (€45 to €1,500, £35 to £1,500) a month at early volumes | 3 to 8 weeks |
| Hosted model plus your data | Retrieval from your documents, fine-tuning, workflows with review steps, evaluation suites | $40,000 to $150,000 (€35,000 to €130,000, £30,000 to £110,000) | $500 to $10,000 (€435 to €8,500, £370 to £7,500) a month, growing with usage and data | 2 to 5 months |
| Custom model | Training a model on your own data for a task general models cannot do | $150,000 to $500,000 (€130,000 to €435,000, £110,000 to £370,000) and up | Infrastructure and a team to maintain it | 6 months and up |
Almost every startup feature belongs in the first row, and a good share of those in the second. The third is for companies with unique data and a task that is the entire business.
What drives the build cost?
- The feature itself. Classifying a support ticket is a week. A conversational assistant that answers from your documentation with sources is a couple of months.
- Data preparation. Collecting examples, cleaning them, structuring documents for retrieval. Often the largest and least visible item.
- Evaluation. A set of real examples with known good answers, and the tooling to measure the feature against them every time something changes. Without it you cannot tell whether the feature works or whether a change made it worse.
- The review workflow. The screens and logic that let a person check and correct outputs where being wrong matters. Frequently more work than the AI call.
- Integration. Connecting the feature to your product's data and flows, the same as any integration.
- Safety and privacy. Filtering what is sent to the model, handling personal data correctly, preventing misuse.
What drives the running cost?
Hosted models are billed by usage, typically by the amount of text processed. Running cost therefore depends on how many users use the feature, how often, how much text each use involves, and which model tier is chosen. A feature that summarises one short message per user per day costs little. One that processes long documents for every user on every visit can become the largest line in the hosting bill. Design decisions such as caching, choosing a smaller model for simple steps, and limiting input size change the running cost by an order of magnitude, so they belong in the design, not in the optimisation phase. Check current provider pricing at the time you design, because it changes frequently and mostly downward.
The hidden costs
- Data you do not have. If the feature needs examples of your specific task and none exist, someone has to create them. Budget for it.
- The accuracy gap. A prototype at 85 percent accuracy is impressive. A product needs to know what happens to the other 15 percent, and that is the review workflow.
- Evaluation maintenance. Models change, providers update, prompts drift. The evaluation suite has to be run and updated. This is ongoing.
- Provider dependence. Switching model providers is usually feasible but not free. Design the integration so that the model is a component, not the foundation.
- Support load. AI features generate a different kind of support question. Plan for it.
Keeping it in proportion
- One feature first. The one that passes the five questions in should your app use AI.
- Prototype in a week with a hosted model and fifty real examples before committing to a build.
- Measure before building. If the prototype is not accurate enough on real examples, the build will not fix that.
- Design for cost. Cache, limit input, use the smallest model that works for each step.
- Budget the review workflow. It is part of the feature.
For context, the AI feature is usually a fraction of the product around it. Our guide to what an app costs covers the rest.
How 7L scopes AI features
Artificial intelligence is one of our technical services and AI Business Empowerment is how we approach it: the process and the decision first, then the technology. We prototype with real examples before quoting a build, we include the evaluation suite and the review workflow in the scope, and we design for running cost from the start. If you have an AI feature in mind, describe the task and the volume and we will give you a build and running estimate.
Frequently asked questions
Can I add AI to my app for under $10,000?
A simple feature using a hosted model with a good prompt and light integration can come in around that figure, especially as an addition to an existing product with a team that knows it. Anything needing your own data, review workflows or evaluation costs more.
Why is the running cost hard to predict?
Because it depends on how users behave. Usage-based billing means the cost follows the number of uses and the size of each. Estimate from expected usage, set alerts, and design to limit input size from the start.
Should I train my own model?
Almost certainly not at the start. Hosted models handle most tasks well, and retrieval from your data or light fine-tuning covers most of the rest. Custom training is for unique data and a task that is the core of the business, at a scale that justifies the cost.
What is an evaluation suite and why does it cost money?
A set of real inputs with known good outputs, and tooling to measure the feature against them. It is how you know the feature works and stays working when models or prompts change. It is part of the build and part of the maintenance.
Is it cheaper to use a no-code AI tool?
For internal use and prototypes, often yes. For a feature inside your product, no-code AI tools tend to limit control over accuracy, privacy and cost, which matter more as users grow.