Catch unsupported refund promises before they reach a customer.
A scoped policy-middleware pilot for customer-support AI. Compatibility, policy coverage, latency, and acceptance thresholds are measured before a production claim is made.
A fluent answer can still contradict the written policy.
Moffatt v. Air Canada is a useful warning that an organization can be held responsible for information delivered by its chatbot. The exact legal and regulatory effect depends on jurisdiction and should be reviewed by counsel.
Moffatt v. Air Canada
In 2024, a Canadian tribunal ruled Air Canada liable for a bereavement-fare policy its chatbot invented. The company argued the bot was a "separate legal entity." The tribunal disagreed. Your chatbot speaks for your company -- legally.
One answer, a million refunds
A single hallucinated promise, broadcast to every customer asking a similar question, becomes a contractual obligation you didn't agree to. CX AI hallucinations don't fail in isolation -- they fail at scale.
EU AI Act: high-risk
Regulatory obligations depend on the system's role, classification, and jurisdiction. A scoped build can produce technical logs and controls, but qualified counsel determines legal applicability.
Drop-in middleware. Every message checked before it ships.
The proposed guardrail sits between a compatible model endpoint and the customer. It evaluates a draft against configured deterministic rules and experimental scores, then follows the selected allow, block, rewrite, or escalation path. Pilot tests measure what it catches and misses.
Customer
Asks a question via chat, email, or voice
Your LLM
OpenAI, Anthropic, local, or any API
CX Guardrail
Policy check · refund ceiling · escalation · audit
Customer
Receives an approved response or a human handoff
Policy match
Every proposed response is scored against your actual written policies -- not a vibes-based guardrail prompt, but a manifold check against your versioned policy files.
Refund ceiling
Hard caps on any monetary commitment the bot can make autonomously. Anything above the threshold requires human approval -- non-negotiable.
Escalation trigger
Legal language, distressed sentiment, regulated domains, or ambiguous requests are flagged and handed to a human agent before the bot answers.
Audit log
The pilot records the input, draft, matched rule, action, and final output in a reviewable receipt. Cryptographic signing and retention requirements are selected in scope.
Three indicative operating bands.
These are design targets for scoping, not active self-serve subscriptions. Final throughput, integrations, support, and price follow the pilot evidence and written agreement.
For a small team evaluating whether a bounded policy layer fits its support workflow.
- Up to 10K messages per month
- Standard policy templates (refunds, returns, shipping, eligibility)
- Email support, 1 business-day SLA
- Monthly audit log export (CSV + JSON)
- One agreed LLM API compatibility target
- Latency measured and reported during pilot
For scaling CX teams who need custom policies and real-time oversight.
- Up to 100K messages per month
- Custom policy authoring (YAML or JSON, versioned in your repo)
- Slack integration for escalations
- Real-time dashboard with block/rewrite/escalate counters
- Weekly audit reports delivered to your compliance inbox
- Everything in Starter
For regulated industries, multi-brand deployments, and teams needing a contracted SLA.
- Unlimited message volume
- White-glove onboarding (policy migration, LLM integration, QA)
- Dedicated Slack channel with our engineering team
- Custom integrations (Zendesk, Salesforce, Intercom, bespoke)
- Negotiated SLA and response windows if the pilot advances
- Everything in Growth
Measure the behavior before promising the outcome.
The pilot tests whether the selected rules, thresholds, and escalation path reduce specific failure modes without creating unacceptable false positives.
- Measure how often configured checks intercept unsupported refund statements before delivery
- Produce technical evidence for review by your compliance and legal teams
- Cryptographic audit trail for every decision your bot makes
- Measure integration effort for the selected API and response format
- Policy-as-code you version in git alongside the rest of your infra
- Human-in-the-loop escalation paths that actually trigger when they should
- Real-time alerting when your bot tries to go off-policy
- Generate technical reporting for review by the organization's compliance team
Built on a governance framework we've been hardening for years.
Powered by SCBE-AETHERMOORE
The CX Refund Guardrail is a proposed application of the patent-pending 14-layer governance pipeline (USPTO #63/961,403). Geometry-based scores are experimental features to evaluate alongside deterministic policy checks; they are not a guarantee that every forbidden response is caught.
The core framework is open-source on GitHub. The Guardrail is what happens when you take that framework, pre-configure it for customer-support policy enforcement, and wrap it in an API.
The questions we get asked first.
Does this work with any LLM?
The design is model-agnostic, but every API and response format must pass a compatibility test during the pilot.
How fast is it?
Latency depends on the selected checks, deployment, and API. The pilot records a baseline and an observed latency distribution before any target is promised.
Can we bring our own policies?
Yes. Policies are YAML or JSON files, versioned in your repo alongside the rest of your infrastructure. Change control, code review, and rollback are git operations you already know. No policy is ever edited through a hosted dashboard without leaving an audit trail.
What happens on a policy violation?
Configurable per rule: block the message entirely, escalate to a human agent, or rewrite the response using a safe template. You choose the default, and you can override per policy. Every action is logged.
Is there a free trial?
There is no public self-serve trial. A bounded proof-of-concept can be scoped with explicit test data, cost, and acceptance criteria.