We are often asked why Saolix, a company building AI into its products, refuses to let AI make the final call. The short answer: a decision in a regulated business has to meet three tests, and today's models pass none of them on their own.

Three tests a decision must pass

It must be explainable. When a bank blocks a payment or rejects a customer, someone will ask why. The answer has to point at something concrete: this threshold, this list entry, this rule. "The model scored it 0.87" is a number, not a reason.

It must be repeatable. Run the same case again with the same inputs and the same rules, and you must get the same answer. Language models are sampled. Even at their most constrained settings, their outputs can shift when the model, the prompt or the surrounding text changes slightly.

It must be owned. A named person or a versioned rule, approved by a named person, has to stand behind the outcome. "The AI decided" is not an answer any regulator, auditor or customer will accept.

Models will keep getting better. But accountability does not come with capability: however good a model becomes, it cannot answer for a decision the way a person or an approved rule can.

A more capable model is not a more accountable one.

What AI is genuinely good at

None of this makes AI less useful. It changes where it goes. Models are very good at the work that surrounds a decision: reading messy evidence quickly, noticing that four events in four different systems belong to one story, and explaining a match in language a busy analyst can act on. That is where analysts lose most of their time, and where the right help pays off fastest.

So every Saolix product is designed with AI above a line we call the determinism boundary. Above it, models suggest, rank, label and explain. Below it, versioned rules and named people decide, and every decision is recorded.

The same line in four products

  • Eagle (sanctions screening): a triage copilot explains each alert and an optional AI agent can label it, while configured thresholds produce the verdict. A code guard forces any real hit back to human review, whatever the model says.
  • K2M (identity verification): automated checks score the document and the liveness challenge, but minimums written into the code mark failed checks, reviewers decide, and decisions on high-risk sessions need a second reviewer. K2M is in pre-release.
  • F2X (fraud, in development): the copilot links related activity and explains the pattern, and governed rules decide whether to hold, block or allow.
  • Shield (security, in development): the copilot explains the likely attack path, and a response runs only under approved policy, with approvers where the policy requires them.

Three design rules we hold to

1. The AI can always be switched off

If turning the model off changes a decision, the model was deciding. In Eagle, AI assistance has its own switch, and switching it off leaves every verdict unchanged. The analyst loses help, not correctness.

2. When unsure, escalate to a person

If the model is unavailable, or its answer conflicts with the rules, the safe default is human review. Our copilots are designed to fail towards a person looking, never towards a case quietly closing.

3. Keep what the AI said apart from what was decided

AI output is stored as an advisory note, never as part of the decision record. You can always tell what the AI suggested and what was actually decided.

Better AI, not less AI

This design lets us improve the models freely. A better model makes analysts faster; it never silently changes outcomes. A customer can adopt the copilot gradually, measure it against their own analysts, and switch it off entirely without re-validating a single rule.

In our view that is what responsible AI looks like in regulated work. Not a promise about the model, but a structure in which the model's mistakes cannot become decisions. You can see the structure for yourself in the four product simulators.