The short answer
Adding AI features to software you already run typically costs $10,000 to $60,000, delivered in three to ten weeks. The price depends far more on your existing codebase and data than on the AI itself.
A single feature in a modern app, such as a “summarize this ticket” button or data extraction from uploaded documents, sits near the low end. AI that reads across customer records, respects user permissions and takes actions inside the product sits near the high end.
Common AI features and what drives their cost
| Feature | Typical effort | What makes it harder |
|---|---|---|
| Summaries and drafts | Low | Long inputs, strict tone or format rules |
| Classification and tagging | Low | Many categories, ambiguous edge cases |
| Document extraction | Low to medium | Scanned files, many layouts, validation rules |
| Search and Q&A over your content | Medium | Permissions per user, messy or duplicated sources |
| Natural-language reporting | Medium | Complex schemas, need for exact numbers |
| AI that takes actions | High | Confirmations, audit trails, rollback, abuse prevention |
Natural-language reporting is a good example of a feature that demos well in a day and takes weeks to make reliable, because wrong numbers erode trust faster than no feature at all.
Where the money goes
| Phase | Share of effort | What happens |
|---|---|---|
| Codebase and data review | 10–15% | Read the code, check the deploy process, find where the feature fits |
| Data access and permissions | 20–30% | Connect to the right tables or documents and enforce who sees what |
| Feature logic and prompts | 15–25% | Prompts, model calls, retrieval and structured outputs |
| User interface | 15–20% | Streaming output, loading states, edit and undo, feedback buttons |
| Evaluation and guardrails | 15–20% | Test sets, automated scoring, fallbacks for failures and slow responses |
| Release and handoff | 5–10% | Feature flags, monitoring, cost tracking and documentation |
Teams often expect prompts to be the main job. In practice, permissions and evaluation take most of the time on any feature that touches real customer data. Our guide on connecting an LLM to internal data safely explains why.
When not to build it
- Your vendor already ships it. Many CRMs, help desks and document tools now include AI features. If one covers your need, turn it on and save the project budget.
- The data is not ready. If the information the feature needs is scattered or out of date, fix that first. AI on bad data produces confident wrong answers.
- Nobody will own it. AI features need someone to review feedback and adjust prompts over time. Without an owner, quality drifts.
How to keep the cost down
- Ship one feature behind a flag. Release to internal users first, measure, then expand.
- Start read-only. Suggestions that a person accepts are cheaper and safer than automatic actions.
- Pick the smallest model that passes. Route simple tasks to a cheaper model tier and keep the expensive one for hard steps. See what it costs to run an AI agent for how model choice drives the monthly bill.
- Give us a working dev environment. A repo that builds and deploys cleanly on day one saves a week.
How Lumeor prices AI integration
We start with a short review of your codebase and data, then quote the feature work as a fixed-price project with a written scope: the features, the data they touch, the evaluation plan and the timeline. Examples of AI built into products include our Shopify AI copilot. If your app is not built yet, an MVP web app cost estimate is the better starting point.