Governed AI systems, for work where being wrong costs something.
Retrieval, drafting, classification and agents — with the controls a regulated business needs designed in from the start: answers grounded in your own documents, rules enforced in code, a person approving what matters, and a record of every call.
What we do
- Retrieval over your content
- Your documents ingested from SharePoint, the web or your own systems, so answers come from your material rather than the model’s memory — and every citation is checked against the source before it is shown.
- Agents with defined limits
- Each agent has a stated role, defined inputs and outputs, the actions it may take and the conditions under which it must refuse. Predictable by design, not by hope.
- Decisions in code
- Eligibility, pricing authority, compliance checks and calculations live in deterministic rules. The model explains and drafts; the code decides.
- Human approval
- Draft, review, compliance, approved, published: a state machine that nothing skips, so no AI output reaches a customer without a person signing it off.
- Audit and validation
- Input and output validation, prompt-injection checks, and an audit record for every model call — so any output can be traced back to what produced it.
- Integration with your systems
- AI delivered as a service with a clean boundary to your business systems, passing only the data it needs and writing results back where they belong.
Three ways to start
AI projects most often fail by starting with the model instead of the decision. Each of these starts with the decision.
Where it usually hurts
If one of these sounds familiar, it is the kind of problem we are called in for.
- A pilot impressed in the demo, and nobody will sign off on putting it in front of customers.
- Staff are pasting company documents into public AI tools because no approved route exists.
- The model answers confidently and cites policies or rules that do not exist.
- Compliance asked how a decision was made, and nobody could say.
- A vendor’s demo ran on their data, and nobody knows how it behaves on yours.
Where this experience comes from
Our lead analyst's career history, described by sector — not an Elarion client list. More on the About page.
- Two AI pilots for a lumber manufacturer: an order-validation pipeline of three agents over around forty deterministic rules, and an order-intelligence advisor for sales representatives — each taken from requirements through agent design and build to a stakeholder demonstration.
- A six-agent tutoring system for an exam-preparation product, with deterministic learning controls beneath the model and rubric-based scoring against the official exam framework.
- Conversational automation for customer service at a global retailer, designing the flows with business experts on Microsoft’s conversational agent platform.
A governed AI platform we built and run.
KriftAI routes every model call through validation, a deterministic compliance gate and an audit log, and nothing it writes reaches a customer until a person approves it.
Questions buyers ask
- Which AI models do you use?
- Whichever suits the job. The model is chosen per use case on quality, cost and speed, and the system is built so it can be changed later. Nothing ties you to one provider.
- Will our data be used to train someone else’s model?
- We choose providers and terms under which your data is not used for training, keep personal data out of the AI layer where it is not needed, and document where every piece of data goes.
- Can the AI make the decision for us?
- For anything with consequences, we design it not to. The model retrieves, drafts, explains and flags; rules in code make the decision; a person approves what matters. That is what makes it defensible when someone asks how a decision was made.
- How do you stop it making things up?
- By grounding answers in your own documents, checking every citation against the source and withholding answers that fail the check, and keeping calculations and eligibility rules out of the model entirely.
- Where should we start?
- With one process where the cost of the status quo is clear. Usually that means a pilot on real data, demonstrated to the people who would have to approve it.
Further reading
- Using AI in a law firm or regulated practice without taking on the liabilityThe question is not whether AI is useful to a regulated practice.
- Privacy by design for custom software in CanadaPrivacy obligations are much cheaper to meet in the specification than in a finished system.
Related services
Putting AI in front of something that matters?
30 minutes, free. Bring the problem, or the change you already have in mind.