Services / 04
Put AI to work
Someone on your team spends their day answering the same questions, reading the same documents, and typing the same things into the same systems.
What you can answer afterwards
Outcomes, not features.
- Answers from your own documents
- An assistant that reads your policies, catalogue or records and answers customers or staff with a citation to the source.
- Manual steps done for you
- The reading, sorting, drafting and re-typing that fills a working day, handled before anyone arrives.
- AI inside your product
- A feature your customers use, built into the website or app you already have.
- Proof that it works
- Every answer checked against a test set before launch, and measured after it, so you know the error rate rather than hoping.
What we actually set up
For the person who wants to check the work.
Everything above is in your language. This is the same work in ours, so you can hold it against anyone else’s quote.
- Assistants built on the Claude API, or another model where the task calls for it
- Retrieval over your own documents, so answers come with a source rather than from memory
- Integrations with the systems you already run: email, help desk, CRM, spreadsheets, your own database
- Automation of back-office steps, with a person approving anything that leaves the building
- An evaluation set written with you, and a score before and after every change
- Usage and cost measured per task, so you know what each answer costs
- Data kept in your accounts, under your terms, with nothing used to train anyone else's model
What lands with you. A sample, so the shape is honest before the figures are yours.
We did this for a client.
An Australian taxi company: More calls, at less than half the cost, by telling Google which callers paid.
Questions people ask
The ones that actually arrive.
Will it make things up?
Left alone, any model can. That is why every answer comes with its source, why anything customer-facing is checked against a test set before launch, and why a person approves anything that goes out under your name until the error rate is one you have seen and accepted.
Where does my data go?
Into accounts you own, under terms that do not allow it to be used for training. We will show you the terms before anything is connected, and we do not connect anything you have not seen.
Do I need a lot of data?
No. Most useful work is an assistant reading a few hundred pages you already have, or automating a step someone does forty times a day. Large datasets are the exception, not the requirement.
What does it cost to run?
Usually pennies per task, and we measure it. You will know the cost of every answer before you decide whether to keep it.
What do you need from my team?
The documents or systems the work touches, someone who knows how the job is done today, and a few hours to build the test set with us. The test set is the most valuable thing in the project, and it has to come from you.
Next / 01
Know your customerEvery step a person takes on your website and in your app, recorded and named — so who buys, where they come from and where they give up stops being an opinion.