Studio services / No. 07
AI development
A good AI brief starts with a task you can grade. We turn examples from your workflow into a small evaluation set, then use it to decide whether a model, a search index or a simpler rule solves the problem.
Discuss your project ↗
01 / The brief
Who it’s for
Teams with a repeatable research, support or document task and someone who can judge the quality of the output.
If the answer must be an exact calculation, use deterministic code. A model is more useful for interpreting varied language or drafting a response that a person can check.
02 / Scope
Inside the work
Prototype against a baseline
Try a small slice of the workflow and compare answers with the current process. Track failure types alongside useful results.
Ground answers in permitted sources
Design retrieval around document ownership and access rights. Show the supporting passage when the user needs to verify a claim.
Set operating boundaries
Measure latency and cost per task. Define when to abstain, request approval or hand work back to a person.
03 / Discovery
Where we start
Bring representative examples, a few unacceptable answers, the sources a user may access and the action a reviewer should take when the system is unsure.
See how we work ↗04 / Field notes
Before you start
Do we need to train a model on our documents?
Often the first experiment is retrieval: find relevant passages at request time and use them as context. Test that approach before taking on training and model maintenance.
Can an AI assistant take actions in our software?
It can call a scoped integration. Start with read access, then add narrowly defined actions with approval, logs and a recovery path appropriate to their impact.