ServicesLLM integrations
Chat, search, summaries and writing help, built into the product you already have.
An LLM integration puts a large language model to work inside your product, on your data, for your users.
Anyone can call a model’s API. The work is everything around that call: which model answers which request, what context it may see, how the answer reaches the screen, what it costs, and what happens when it is wrong.
We build that whole path, and we make it visible, so you can see how the feature performs and change it without fear.
Features where people read, write or look things up.
None of these needs a chat window. Many useful features are a single button or a field that fills itself in.
A working feature in your product, and the means to run it.
The integration in your code base, its prompts kept as versioned files, and the interface parts your users see.
A test set built from real examples, and the checks that run on every change to a prompt or a model.
Cost limits, rate limits, model settings, and a switch that turns the feature off without a deploy.
Documentation, a walkthrough with your developers, and dashboards for quality, speed and cost.
A typewriter does what you tell it. A language model guesses what you meant.
That is its strength and its risk. So we treat every prompt like code: written down, versioned, reviewed and tested before it reaches your users.
Most products do not need the largest model for every request. We send each kind of request to the smallest model that does it well, and we cap what each feature and each user may spend.
Below is the kind of settings screen we hand over with an integration. Your team can change it without asking us.
| Request | Model | Max tokens | If it fails |
|---|---|---|---|
| Reply draft | Small | 400 | Medium model |
| Ticket summary | Small | 600 | Medium model |
| Policy question | Medium | 800 | Large model |
| Contract review | Large | 2,000 | Hand to a person |
| Limit | Set | Used | Status |
|---|---|---|---|
| Reply drafts | €300 | €212 | 71% |
| Summaries | €150 | €48 | 32% |
| Per user, per day | €0.50 | €0.50 | Reached by 2 users |
An example. The settings and the figures are made up for this page.
Language models fail in ways ordinary software does not. We plan for each of these from the first week.
The model states something false with confidence. We give it the facts it needs, ask it to cite them, and check answers against their sources.
Text in a document or a message tries to give the model new orders. We keep instructions apart from data and limit what any answer can set off.
The model sees, or repeats, something a user should not see. We pass only what that user may see, and remove personal data that is not needed.
A loop, a bot or a very long file sends the bill up overnight. We set limits per user, per feature and per day, with alerts before they are reached.
A provider updates a model and the answers change. We pin model versions where we can and run the tests before any switch.
An integration reaches its first users in a few short rounds. The order is the same every time. The length depends on your product and your data.
You see working software at the end of each round, with its test results.
One use, one group of usersagreed
Real examples collected50–100
First prompt and its testswritten
Built into your interfacebehind a switch
Two or three models comparedon your cases
Cost and rate limitsset
Safety and privacy reviewwith you
Dashboards and alertslive
Handover to your teama walkthrough
A live feature, its tests and its controlsin production
Which model will you use? The one that fits the request, the cost and your rules. We test two or three on your own examples before we choose, and we keep the choice easy to change, because better and cheaper models arrive every few months.
Will our data be used to train the model? Not with the business offers we use. The API terms of the major providers exclude training on your data by default, and we check the terms that apply to you and write them down before we build.
Does it work in languages other than English? Yes. Current models handle most European languages well. We test in every language your users write in, because quality is not the same in all of them.
What will it cost to run? It depends on how many requests you get and which model answers them. Before we build, we estimate the monthly cost from your real numbers. After launch, the limits keep it inside the budget you set.
Can we change the prompts ourselves? Yes. Prompts live as files in your repository, next to their tests. Change one, run the tests, and you see what the change did before your users do.
What if the provider goes down? The feature fails gently. We can route to a second provider, fall back to a simpler answer, or show a clear message, and we decide which with you up front.
Write a few lines about your product and what you want the feature to do. We will reply with questions, or with a time for a short call.
A few lines are enough. We reply within one working day.
Start a project