What an AI feature actually involves
The model call is the smallest part. A dependable AI feature needs the data pipeline that feeds it, an evaluation set to measure quality, guardrails, a review step where mistakes are costly, logging, limits and an interface your users understand. That surrounding engineering is where most of the hours go.
- Preparing inputs: uploads, parsing, cleaning and splitting documents
- Choosing a model and testing it against your own examples
- Prompt and output design, with structured results your code can use
- Guardrails for what the feature may answer, and fallbacks when it fails
- Cost caps and usage limits enforced in code
- Monitoring of quality and cost after launch
Document extraction in an existing app worked through
Your customers upload invoices or contracts, and your app should pull out the key fields, show them for review and export them.
- Base hours for an AI feature: 96
- AI document extraction 40 to 80, file uploads 12 to 24, one third-party integration 16 to 48, reports and export 16 to 40
- Subtotal: 180 to 288 hours
- Design, QA and project management, plus 45 percent: 261 to 418 hours
- At $89 per hour, rounded: $23,000 to $37,000
An assistant on your own data worked through
Your users ask questions in plain language and get answers grounded in your help center, documents or database, with links to the sources.
- Base hours for an AI feature: 96
- AI assistant on your data 48 to 96, search and filters 16 to 40, analytics events 8 to 16
- Subtotal: 168 to 248 hours
- Design, QA and project management, plus 45 percent: 244 to 360 hours
- At $89 per hour, rounded: $21,500 to $32,000
AI features in our model range from $12,500 for the base package to $58,000 for a large scope. Combining several AI features, or building a product around one, moves into the MVP range.
The running cost nobody budgets
Every request to a model provider costs money, priced by the amount of text or images processed. For a feature used a few hundred times a day, that is often modest. For a feature that processes long documents, images or runs in the background for every account, it can exceed the hosting bill many times over. Three habits keep it under control:
- Cap usage per account and per day in code, and show the limit to users
- Run AI only for paying accounts, not for every trial or test account
- Cache results, so the same document or question is not processed twice
We estimate running costs in Discovery and build the caps into the feature, because an AI bill that doubles overnight is a design flaw, not bad luck.
How to choose a model
There is no best model, only a best fit for your task, your data rules and your budget. Large commercial models are strong at reasoning and language. Smaller models are cheaper and faster for classification and extraction. Open models can run in your own infrastructure when data must not leave it. The only reliable way to choose is to test candidates on an evaluation set built from your real examples, and to measure accuracy and cost side by side.
Data, privacy and ownership
Before you send customer data to any model, check where it is processed, whether it may be used for training and how long it is kept. Choose provider settings that exclude training, keep logs free of secrets and give users a way to correct results. Our approach is described on our secure software development page.
Start small and measure
The best first AI feature is narrow, measurable and attached to a task people already do. Extraction, classification and search over your own content usually pay back fastest. Assistants that answer anything about everything are the hardest to make reliable. Build one feature, measure it on real usage for a month, then decide on the next.
Price your AI feature
Our AI development page explains what we build and how. To see your own numbers, choose AI feature in the estimate form, add the supporting features and get the hours, the range and the team in about a minute.