Service · Integration
AI integration and automation
Small, specific features that remove a manual step
Most useful AI work in an established business is narrow: a support inbox that drafts replies from your help pages, a contract folder you can search by meaning, an intake form whose attachments are read and filed, a weekly report assembled from three systems. We connect the model to the system where the work already happens, so nobody has to learn a new tool to use it.
Each feature ships with an evaluation set built from your real cases. It measures whether the AI gives the right answer on your material, and it is what we rerun when a provider releases a cheaper model.
Feature hours plus usage, each on its own line
The build hours for each AI feature come from the model's AI-native tables. These are the figures it uses for a feature added to an app, before any integration with your own systems:
| Feature | App | Server | Testing | Usage assumed per user |
|---|---|---|---|---|
| Chat assistant20 requests per user a month, 3,000 tokens in, 400 out | 16 | 44 | 12 | 20 requests per user a month, 3,000 tokens in, 400 out |
| Semantic search30 searches per user a month, 50 tokens each, embeddings only | 6 | 28 | 4 | 30 searches per user a month, 50 tokens each, embeddings only |
| Content generation10 drafts per user a month, 1,500 tokens in, 800 out | 10 | 22 | 6 | 10 drafts per user a month, 1,500 tokens in, 800 out |
Integration with your own systems is quoted on top, from the model's per-integration hours, adjusted after we have seen the systems. The monthly usage estimate uses the model's formula: users × requests per user × tokens per request × the provider's published price per million tokens. For a chat assistant used by 1,000 people, that is $10 a month on gpt-6-luna and $200 on Claude Sonnet 5.5, at the prices the model read on 2 October 2026.
Both appear on the quote with their assumptions, and usage is billed at cost. See how we price for the rest of the method.
Prompts, test sets and configuration, in your repository
In an AI feature, the prompts, the evaluation set and the model configuration are as much a part of the product as the code. We keep them as files in your repository, versioned alongside the code, from the first commit. If you change supplier, the next team starts from everything we wrote. The five questions page sets out the ownership terms in full.
Questions people ask us
How much does AI integration cost?
For a single feature added to an existing app, our model puts the AI-native build and testing at 38 to 72 hours depending on the feature, before integration with your own systems. The larger cost over a year is often the usage bill, which we estimate on the quote from your expected users and pass through at the provider's price.
Do you mark up AI or API usage?
No. Usage during a build appears at cost on its own invoice line with the provider's invoice behind it. For production we set the keys up in your own provider account, so the bill goes to you directly.
Which AI model do you use?
The one that passes your evaluation set at the lowest running cost. We build a small test set from your real cases first, run candidate models against it, and show you the accuracy and the monthly cost of each before choosing. The choice stays in configuration you own, so it can change as prices and models change.
Is our data used to train the AI?
We use provider plans and API settings under which your inputs are not used for training, and we tell you which ones before work starts. Where data must stay in the UK or the EU, that narrows the model choice, and the quote states the effect on cost.
Start a conversation
Send us a scope, or a link to your numbers in the model. We reply with questions or a quote.