AI solutions
AI agents
AI agents and AI features on GPT and Claude over your own data. Around the model sits an ordinary system: an API, a queue, a database, logs. First a prototype, then production.
- Reply to your brief
- within 24 hours
- First call
- 30 minutes, no commitment
- First working version
- in 5–6 weeks
Process
What happens, and when
The work is split into stages, and each one has a named result. You see it yourself rather than reading about it in a report.
Days 1–5
Limits and the measure
We decide which jobs the agent takes end to end and where it stops and calls a person. Two or three dozen real jobs with known correct outcomes are collected here too — from that point on they, not impressions, are the measure of quality.
You get — A test set and a defined hand-off
Weeks 2–3
Tools and permissions
We design the agent’s tools: what it may read, what it may write, which actions can be undone. Write access is granted separately from read access, and every action lands in a log together with what the agent saw before taking it.
You get — The agent’s tools, split access and a log
From week 4
Build and measure
We build the agent and run it against the test jobs, counting correct answers, wrong ones and refusals separately. A prompt or tool change is kept only if the set improves; what the agent does when someone else’s API goes quiet is tested on its own.
You get — A working agent scored on the test set
Before launch
A thin slice live
It goes onto live traffic in a thin slice: the agent prepares, a person confirms every irreversible action. In go a spend cap, a ceiling on steps per job, and a switch that stops writes to external systems without taking the rest down.
You get — Live traffic under confirmation, caps and a switch
After launch
Log and corrections
We read the log and the corrections people made to finished work; each one goes back into the test set. When the model version changes the whole set is run again — behaviour shifts without warning, and it is better not to hear about that from a customer.
You get — A growing test set re-run on model changes
What the work covers
Working out the approach
where an agent with tools is needed, where search across your documents is enough, and where a single model call does the job
Connecting your data
indexing documents and databases, search by meaning, answers limited to what was found
Tools for the agent
calling your APIs, writing to CRM and spreadsheets, sending mail and messages
Limits and safety
what the model does on its own, what a person confirms, what goes into the log
Quality checks
a set of test questions with expected answers, measured before and after prompt changes
Going to production
keys and rate limits, caching, token spend, behaviour when the model fails
What we need from you
- A description of the task with examples: what people ask, and what counts as a correct answer
- Access to the data the model works over: documents, a database, an export
- Who on your side confirms that an answer is correct
What people usually ask
Contact
Send a description of the task
A reply with the scope, the timeline and a budget estimate comes within 24 hours.
The first call is 30 minutes, with no commitment on your side.
01
You describe the task
Five questions in the form, or a plain email — whichever suits you.
02
We answer within a day
With the scope, the timeline and a budget estimate, based on what you told us.
03
We talk for 30 minutes
To clear up whatever is unclear. It commits you to nothing.