The questions your security team will ask, answered.

Your data lives in your own cloud environment, hosted in the UK in AWS’s London region. Prompts, outputs and conversation history are stored and processed there. Nothing crosses a border, and nothing is pooled with anyone else’s data.

The detail of what the providers can and cannot see has its own section below.

No. Every model KAL runs, it runs inside managed model infrastructure that the providers never touch, so your prompts never reach Anthropic, Meta or any other provider’s systems. They build the models. They never see what you send to them.

That infrastructure stores nothing: prompts and responses are not retained, not logged and never used to train a model. They are processed to produce your answer, and then they are gone. The contracts commit to it in writing anyway, through a data processing addendum, so the promise does not rest on the architecture alone. The infrastructure carries its own SOC 2 and ISO 27001 certifications, and those are the cloud provider’s, not ours. Where we stand on our own is further down this page.

Yes. Where regulation requires it, KAL runs the model inside your own environment: your data never leaves it, and no provider is ever in the path. Everything runs inside a private network that never touches the public internet, and you can hold the only encryption keys, so not even Kallidin can read your data. We call it sovereign mode.

Yes. Every finding is checked by an independent AI agent before it goes into an answer. We measure that independent agent against a large body of cases where the right answer was already known, and give you a confidence score for each finding and for the answer as a whole.

We built it this way to avoid a trap: trusting one AI because a second AI approves. The independent agent’s judgement is benchmarked against known answers, so its reliability is a measured number, not a claim, and every answer carries an audit trail of what was checked and what was assumed, so you can follow the reasoning step by step. If your governance team asks for a chain of command for AI outputs, this is that chain, recorded at the level of individual decisions.

Two things, and they are the same idea underneath.

The independent agent that checks every finding is itself a model, so we measure it rather than trust it. It is benchmarked against a large body of cases where the right answer was already known, and that is where the confidence score on every finding comes from.

And the estates we test on are built with wrong answers planted in them: signals that look like textbook early warnings and are not. A system that only ever meets data rewarding the obvious answer tells you nothing about what it does when the obvious answer is wrong.

Access is built around how you already govern it. We don’t impose a fixed permissions model, we fit yours, so the roles, approvals and restrictions your organisation already runs carry over rather than being redesigned around a vendor’s assumptions.

As few people as possible, and never invisibly. Access is least privilege and logged: only the named engineers assigned to your engagement, granted for its duration and revoked when it completes. We use a short, disclosed list of subprocessors, our cloud host and the model providers, each bound by the same terms described above: nothing retained, nothing used for training. The full list, and our data processing terms, come with our security pack.

Yes. The rules, guardrails and permissions that govern what the agents may do are versioned and can be rolled back independently. Your data stays fixed: the agents work on it, they never alter it. And every action is logged, so there is a full record.

The agents are designed to improve over time, which is different from traditional software, so the controls are built for it. If something changes and you don’t like it, you have options: roll back the configuration, tighten the guardrails, or revert to a previous deployment.

Within 30 business days of the contract ending, we return all of your data in its original format, or an agreed export format, by secure transfer. Once you confirm the return, we delete it: from active systems within 30 days, and from backups within 30 days after that.

You receive written confirmation of deletion, signed by our Chief Product Officer. Where a legal or regulatory obligation requires longer retention, that obligation takes precedence, and we tell you exactly which data is retained and why.

Cyber Essentials certification is in progress. We are building our processes towards ISO 27001. We claim neither until it is actually held. Accreditation takes time for a new business, and it gets harder with every hire, which is exactly why the work is happening now. The model providers we work with carry their own certifications, and as we said above, those are theirs, not ours. If your procurement process needs the specifics on either, ask us and we will show you exactly where things stand. We will update this page as our status changes.