Deterministic AI decisions.
The model talks, your rules decide.

Deterministic AI is an architecture where the consequential decisions in an AI system are made by explicit rules that return identical outputs for identical inputs, while the language model handles conversation, extraction, and phrasing. An LLM samples from a probability distribution, so even at temperature zero its behavior can shift across model versions and prompt changes. A deterministic decision layer like GoRules removes that variance from the part that matters: eligibility, limits, pricing, routing, and escalation are computed by versioned decision tables, and the model only communicates the result.

Most LLM guardrails inspect input or output: they block unsafe content, redact PII, or constrain the model to a topic. They filter what the model already generated. GoRules moves the decision itself out of the model - the agent calls a decision table as a tool and receives a structured, versioned, explainable outcome. The two are complementary: keep content-level guardrails for safety and tone, and use a decision layer for AI agent guardrails around business policy, where you need the same answer every time and a reason you can show an auditor.

The model can always generate text that claims something, so the pattern is to make the tool result the only source of truth: the prompt instructs the agent to state outcomes only as returned by the decision tool, and your application acts on the structured result, not on the prose. Because every real decision is logged with its trace, a reply that diverges from the recorded outcome is detectable, and the decision of record is still correct - the engine decided, the model merely phrased it. Teams commonly template the consequential sentence directly from the structured output for exactly this reason.

In two places. At runtime, escalation is a rule outcome: rows route specific cases to human review, and anything that matches no rule escalates by default instead of letting the model improvise - the review queue receives the case with its full trace. At change time, the policy itself is governed: edits happen on branches, reviewers see diffs and simulation results, releases require approval, and every change records who approved it. Humans stay in the loop where they add judgment, not on every routine call.

Almost nothing on the scale an agent operates at. The ZEN Engine evaluates rules in microseconds to sub-millisecond, and a networked call to a GoRules agent or REST endpoint typically lands in single-digit milliseconds. An LLM spends hundreds of milliseconds to seconds generating tokens, so the decision call is invisible in the response time. If you want zero network hops, embed the engine in the agent runtime and evaluate in-process.

Yes. Run the full BRMS with Docker or Kubernetes inside your own network, deploy on AWS, Azure, or Google Cloud, or embed the MIT-licensed open-source ZEN Engine directly in your services, so prompts, conversations, and customer data never cross a vendor boundary. SSO and role-based access control govern who can see and change rules, and the audit trail covers both decisions and policy changes. GoRules Cloud is available when managed hosting is the better fit.