Model matched to the task
A cheap fast model for classification, a strong one for judgement. Defaulting everything to the most expensive option is the commonest way AI cost surprises people.
AI
The model is the easy part. The limits are not.
An agent with real access to real data needs scoped credentials, spending caps, an audit trail and a person to escalate to. That machinery is the integration.
Connecting to a model takes an afternoon. What takes judgement is everything around it: which model is right for a task, what the agent is allowed to touch, what happens when it is not confident, how much it may spend before someone is told, and how you explain afterwards why it did what it did. We treat model providers as an integration category like any other — swappable, metered, and governed.
Off until you need it. Turning it on is a toggle — not a project, not a rebuild. See how capability arrives as you grow.
Providers
What we connect to
Available means built and running in production today. On request means we will build it for your application — it is not pre-built, and we will not imply otherwise.
Available
- Anthropic Claude
- OpenAI
- pgvectorRetrieval over your own documents
On request
- Azure OpenAIWhere the customer requires it inside their Azure tenancy
- Amazon Bedrock
- Google Vertex AI
- PineconeWhere retrieval scale justifies a dedicated store
What you get
The part that is not just an API key
Answers grounded in your own documents
Retrieval over your policies, contracts and manuals, with the source shown so a person can check the answer rather than trust it.
Credentials scoped to the job
The agent gets access to exactly what it needs and nothing else — the same least-privilege rule applied to software that acts on its own.
Spending limits with alerts
A loop cannot become an invoice. Thresholds are set, and crossing one raises a notification rather than a month-end surprise.
A person in the loop where it counts
Confidence thresholds and explicit escalation. Autonomy is granted in proportion to the cost of being wrong.
Every action logged
Input, reasoning and outcome recorded, so when someone asks why the agent did that, there is an answer.
For example
What this looks like in practice
Document intake
Suppliers send documents in every format ever invented. Fields are extracted, matched, and only mismatches reach a person.
Screening at volume
Applications or leads read against the real criteria, ranked with reasons. The agent does the reading; a human decides.
Internal first line
Questions answered from your own documentation with citations, handing over the moment it is out of its depth.