"How much does an AI feature cost?" is the question we hear most often right now, and the honest answer is: it depends on far less glamorous things than the model. The model is usually the cheapest part. What drives cost is your data, your accuracy bar, and how much of the work happens after launch.
Here's how we break it down when we scope AI projects.
The anatomy of an AI feature budget
Every AI feature we've built has the same five cost centers, whether the client planned for them or not:
- Discovery and scoping — defining what "good enough" means, where the data lives, and what happens when the AI is wrong. Usually one to two weeks. Skipping this is the most expensive mistake on the list.
- Data preparation — cleaning documents, structuring knowledge bases, setting up pipelines. Often 30–50% of the total effort, and almost always underestimated.
- The build itself — prompts, retrieval, integration with your existing systems, the UI. This is the part everyone budgets for.
- Evaluation — a test suite that measures whether answers are actually correct, not just fluent. Without evals, you're shipping on vibes.
- Monitoring and iteration — usage tracking, quality drift, prompt updates as models change. This never reaches zero.
Realistic ranges by feature type
Every project differs, but based on what we've scoped and built, these ranges hold up in 2026:
A grounded support chatbot — answering questions from your docs with escalation to humans — typically lands in the $10k–$40k range for a solid v1, depending on how messy the source content is and how many systems it needs to touch.
A RAG assistant over internal knowledge — contracts, policies, technical docs, often multi-language — tends to run $25k–$80k. The spread is almost entirely explained by data quality and integration count, not the AI itself.
Custom analytics or document automation — extraction from invoices and forms, forecasting, anomaly detection wired into your operations — usually starts around $40k and climbs with the number of workflows and the accuracy requirements.
If a vendor quotes dramatically below these ranges, ask what's excluded. Usually it's data prep, evals, or both — which means you'll pay for them later, under pressure.
The ongoing costs nobody quotes
This is where AI budgets differ most from traditional software. After launch, expect:
- Token and API costs — often modest ($50–$500/month for typical SME usage), but they scale with adoption, so success has a price tag. Model prices have fallen steadily, which helps.
- Eval runs and regression testing — every prompt tweak and model upgrade needs re-testing against your quality benchmarks.
- Drift — your documents change, your products change, and models get deprecated. Budget maintenance the way you would for any custom system — we suggest the same 15–20% annually we recommend for custom software generally.
Scope a pilot, not a moonshot
The biggest cost lever isn't negotiating rates — it's scoping. We steer clients toward a 4–6 week pilot: one workflow, one measurable outcome, real users. A pilot answers the questions that matter (does it work on your data? do people use it?) for a fraction of a full build, and it produces the evidence you need before committing to more.
The failure pattern we see is the opposite: a six-month "AI transformation" that tries to solve five problems at once, stalls on data access in month two, and ships nothing. Small and measurable beats ambitious and vague, every time.
Build, buy, or API?
Not every AI need justifies custom development. Sometimes an off-the-shelf tool covers 80% of the value for a subscription fee, and sometimes a thin app over an LLM API is the sweet spot between cost and control. We've written a fuller framework in our build vs buy guide for AI — read that before committing budget in either direction.
Our short version: most SMEs should start with API-based custom features. Fine-tuning and self-hosting are rarely worth it until you've proven the workflow and hit a real limitation.
How to keep an AI budget honest
Three questions to ask any vendor — including us:
- What does the eval suite look like? If there's no plan for measuring accuracy, there's no plan.
- What are the monthly running costs at 10x usage? You want that answer before launch, not after.
- What's the smallest version that proves value? A good partner will try to shrink your first phase, not inflate it.
AI features are more affordable in 2026 than they've ever been — but the gap between a demo and a dependable production feature is still where most of the money goes. Budget for that gap and you'll be ahead of most buyers.
If you're trying to put a number on a specific AI feature, get in touch. We'll scope it honestly — including the parts of the budget other estimates leave out — and tell you if a pilot makes more sense than a full build.