The failure mode that matters is not an AI system that breaks. It is one that works beautifully and is wrong. A language model asked a question it has no information about will produce a fluent, confident and entirely invented answer, and nothing about the output indicates that anything went wrong. Somebody then repeats it to a customer. Every system we build is designed against that specific failure.
In practice this means retrieval that is grounded in your material and citations on every claim, so any answer can be traced back to the document it came from. It means the system stating plainly when it has nothing relevant, rather than filling the gap. And it means that anything with real consequences — a legal commitment, a payment, a change to access — requires a human decision rather than being executed autonomously.
Data residency is a design decision rather than a footnote. Models are deployed in Microsoft Azure's Australia East region with regional processing, so material stays onshore and no cross-border disclosure under Australian Privacy Principle 8 arises. Where an organisation has an existing obligation about where its data may be processed, we can show how the system satisfies it.
What is included
Use case assessment
An honest evaluation of whether AI is the right tool for the problem, including the cases where conventional automation is cheaper and more reliable.
Knowledge base construction
Your documents ingested, structured and indexed for retrieval, with the quality of the source material assessed rather than assumed.
Grounded assistant
A question-answering system that answers only from your material, cites its sources, and says so when it has nothing relevant.
Evaluation
A test set of real questions with known answers, so accuracy is measured before deployment and re-measured as the system changes.
Guardrails and approval
Explicit limits on what the system may do without a person, with anything consequential routed to a human for approval.
Cost and usage reporting
Token consumption and cost recorded per interaction, so running cost is visible rather than discovered on an invoice.
How it is delivered
- We start with a narrow, well-defined use case and prove it works before widening the scope. A broad AI programme that has not demonstrated value on one task tends not to demonstrate it on ten.
- Accuracy is measured against a written evaluation set rather than judged impressionistically. Answers that look right and answers that are right are different things, and only measurement distinguishes them.
- Tenant and document access controls are enforced in the data layer, not by instructing the model. An instruction is a request, and retrieved content can itself contain instructions.
What you receive
- Working system deployed to your own Azure tenancy
- Evaluation set with measured accuracy results
- Documentation of what the system may and may not do without human approval
- Source code and infrastructure configuration
- Running cost model based on measured usage
- Training for the people who will operate it
What is not included
Stated plainly, because unstated exclusions are where disputes begin.
- Model provider consumption charges, which are billed to your own account so you can see and control them
- Cleaning or rewriting source documents, which can be scoped separately where the material is not in usable condition
- Guarantees of accuracy — we measure and report accuracy; no honest provider guarantees it
- Autonomous execution of legal, financial or access-control actions, which we will not build
Common questions
Will our data be used to train a model?
No. Systems are built on enterprise model deployments in your own Azure tenancy, where inputs are not used for training. This is a contractual position from Microsoft, not an assurance from us.
Where is the data processed?
In Azure's Australia East region, using a regional deployment rather than a global one. The distinction matters: global deployments route processing to whichever data centre worldwide has capacity, and only a regional deployment keeps processing onshore.
How do we know the answers are correct?
Every answer cites the passage it came from, so it can be checked in seconds. Beyond that, we build an evaluation set of real questions with known answers and measure accuracy against it before the system is deployed.