GOVERNANCE · BUILT INTO AXIOMAI
AI chatbot governance, built in.
Enterprise guardrails for customer-facing AI. Every conversation passes through a policy layer you control: prompt-injection guards, reversible masking of personal data (PII), compliance-framework rules, cost budgets — and a hash-chained audit trail that proves what happened, decision by decision.
Or run the interactive simulator belowVisitor asks
What the cloud model sees
WHY IT MATTERS
The three questions every security review asks.
Put an AI on your website and your buyer's security team will ask the same three things. Governance is how you answer them with evidence instead of assurances.
01
What reaches the model?
Visitors paste card numbers, IDs, and medical details into chat boxes — they don't read your data policy first. Governance classifies and masks sensitive values before anything leaves your server for a cloud model.
02
What can the bot leak?
Your indexed documents contain more than you think. The egress guard screens every answer for secrets and personal data on the way out — and never-emit rules keep values like card numbers out of replies entirely.
03
Who can prove it?
When the audit comes, "trust us" is not an answer. Every decision is recorded in a hash-chained trail, stamped with the exact policy version that made it. Tamper with a record and the chain breaks visibly.
THE ROLLOUT LADDER
Turn it on without breaking anything.
Governance that flips from off to blocking in one step gets turned off again by Friday. AxiomAI rolls out in three deliberate steps — watch the audit trail between each one.
STEP 1
Monitor
record only
Nothing changes for visitors. Every request is classified and every would-be decision is logged, so you see exactly what the policy would do before it does anything.
STEP 2
Monitor + masking
PII masked for real
Masking goes live — sensitive values stop reaching cloud models. Everything else still shadows. Answer quality holds, because masking is reversible by design.
STEP 3
Enforce
blocks, masks, and scrubs
The full policy is live: injection attempts blocked, hard-block frameworks never egress, never-emit values scrubbed from every answer. The trail keeps recording.
Prefer the conservative default? Enable fail-closed and a guard error blocks the request rather than waving it through unchecked — the right posture for regulated intake.
THE GUARDS
Five guards. Both directions.
Input and output are guarded independently — what visitors type is screened before it reaches your AI agent, and what your agent says is screened before it reaches visitors.
INPUT
Prompt-injection guard
Screens what visitors type before it reaches your bot. Prompt injection — hidden instructions in a visitor message that trick the bot — is the #1 risk on the OWASP Top 10 for LLM applications (the standard list of AI security risks) — and the industry already has its cautionary screenshots of bots talked into promising $1 cars. Smuggled instructions are caught at the door, not after the model has obeyed them.
OUTPUT
Secret / PII egress guard
Screens every answer on the way out. Keys, credentials, and personal data that surface from your indexed documents are caught before a visitor ever sees them.
PII + DATA CLASS
Classification
Every request is classified against your data-protection frameworks — what kind of sensitive content it carries, and which policy list that puts it on.
TUNABLE
NER person-name detection
Named-entity recognition (NER) — AI that finds names in ordinary text — catches person names that pattern rules miss, with a confidence threshold you control — turn it up for regulated intake, down for low-risk FAQ bots.
THE KEY MOVE
Reversible masking on the cloud lane
Detected values are swapped for typed tokens like [PAN_REF_1] before the request leaves your server, and rehydrated in the reply. The cloud model never sees the real value. The visitor never sees the token.
Why reversible masking matters: most "PII redaction" destroys the value, so the model can't answer the question. Typed tokens keep the structure — the model helps with the double-charge without ever holding the card number.
TRY IT
Run a message through the policy layer.
Pick what a visitor types and which mode you're in, then watch every decision land — including the one your auditor will ask about. Scripted demo; the real pipeline ships in every portal.
1 · Pick what the visitor types
2 · Pick the governance mode
Masking is live — sensitive values never reach the cloud model. Everything else still shadows.
-
Prompt-injection guard · input
waiting -
Classification · frameworks
waiting -
What the cloud model sees
waiting -
Secret / PII egress guard · output
waiting -
What the visitor sees
waiting
Every run appends a record carrying a hash of the one before it — edit history and the chain breaks. This demo is scripted; the same pipeline runs for real in every AxiomAI portal, including the trial.
DATA PROTECTION
Three lists decide what leaves, what never does, and what always escalates.
Detected content is classified into compliance frameworks. Each framework goes on one of three policy lists — and the lists, not vibes, decide how every request is handled.
Maskable
mask and send
Values are replaced with typed reference tokens before leaving for the cloud model, and rehydrated in the reply on your server. Useful data stays useful; sensitive values stay home.
Hard-block
never egress
Content in these frameworks never leaves your server for a cloud model — masked or not. The request is handled without external egress, full stop.
Always-on
tenant escalation
Frameworks added to every request's classification regardless of what detection finds — the floor your tenant (workspace) policy never drops below.
Sovereignty lists pair with data-residency scoping for EU AI Act–era deployments — high-risk obligations have applied since August 2026, and "we'll get to governance later" stopped being a plan.
NEVER-EMIT RULES
Some values should never appear in an answer — card numbers are the canonical case. Never-emit types are scrubbed from every reply, and you can block outright when a never-emit value shows up that the visitor didn't supply. That's your leak detector, not just your filter.
// answer, before scrubbing
"…the card ending 4929 1156 8732 0041 [SCRUBBED: PAN] was charged twice…"
DEPLOYMENT MODELS
Run it where your law requires.
Policy lists control what egresses. Deployment models control where anything can egress to. When sovereignty or data-residency law is the requirement, we change the architecture — not just the configuration.
DEFAULT
Cloud SaaS.
Our managed infrastructure, fastest to live. Vendor LLMs run under zero-retention API terms, the full governance layer sits in front of every call, and EU or custom-region residency can be scoped in.
SOVEREIGN
On-premise, on your hardware.
The same portal and engine, deployed on infrastructure you own and operate — for government, banking, and regulated deployments where customer content may not leave your estate at all.
FULLY AIR-GAPPED AI
Open-source LLM stack.
Pair on-premise with open-source models running on your GPUs and there is no external model call to govern — the hard-block "never egress" guarantee becomes physically true, not just enforced.
On-premise and open-LLM deployments are Enterprise engagements, scoped and priced to the scale of the deployment — hardware footprint, tenants, and users. Talk to us about sovereignty requirements →
THE AUDIT TRAIL
Tamper-evident, or it didn't happen.
Every governance decision — pass, mask, block, scrub — is written to an audit trail where each record carries a hash of the record before it. Alter one historical entry and the chain breaks visibly. The policy version is stamped automatically on every save, so each decision names the exact policy that made it.
-
Scoped, not smeared
Policies apply at tenant default or per chatbot. A more specific policy replaces the whole less-specific one — fields never merge, so there is exactly one answer to "which policy decided this?"
-
Versioned on every save
The policy version stamps automatically. Your auditor sees not just what was decided, but which ruleset was live at that moment.
-
Independently verifiable
Export the trail and verify the hash chain yourself. Tamper-evident means your auditor doesn't have to take our word — or yours.
BUDGETS & BRAKES
Guardrails for your bill, too.
Runaway AI cost is a governance problem like any other. The same policy layer enforces rate limits and budgets — per visitor, per tenant, per month.
Rate limit
Requests per minute, keyed by visitor IP on the widget or by user in the portal. Abuse stops at the door.
Monthly token budget
A per-tenant ceiling on tokens — the units AI usage is measured in — for the calendar month. Zero means unlimited — you choose.
Monthly cost budget
The same ceiling, denominated in dollars — the number your finance team actually asked for.
Per-request token cap
Caps estimated outbound prompt tokens on any single request, so no one conversation eats the budget.
Question length flag
Soft-flags questions above a character limit before they become oversized prompts.
Degrade before you block
On a soft budget breach, governance can suggest a cheaper model instead of cutting visitors off — service degrades gracefully, not abruptly.
A NOTE ON WHAT THIS IS — AND ISN'T
A control layer, not a certificate.
Governance gives you enforcement and evidence: what was detected, what was masked or blocked, and a verifiable trail of every decision. It does not make your organization GDPR-compliant or HIPAA-certified by itself — no software does, and we'd rather tell you that here than let a badge imply it.
Our own security posture — sub-processors, retention, data-protection terms, and the certifications we don't yet hold — is documented in the same plain language.
Read our data practicesCOMMON QUESTIONS
What security reviews ask about governance.
What are AI guardrails for customer service?
Does putting a chatbot on our site drag us into PCI scope?
What does "governance" mean in AxiomAI?
How does PII masking work without breaking answers?
What happens if a guard itself fails?
Can I turn this on without breaking my live bot?
Does the compliance-framework tagging make us HIPAA or GDPR certified?
Is governance an enterprise add-on?
Can we deploy on-premise, with open-source models, for sovereignty requirements?
Can my auditor verify the trail, or do they have to trust you?
SEE IT IN YOUR OWN PORTAL
Governance is in the trial.
The Governance workspace ships in the same portal you get for 14 days, free. Start in Monitor, watch the trail, and show your security team real decisions instead of a datasheet.
THE SAME DISCIPLINE · SERVICEORBIT
AI you can audit, inside the service desk.
ServiceOrbit applies the same thinking to internal requests. Every field a service collects is classified, the rules decide what a model may see, your policy picks the model for each kind of data, and every AI call is logged with the model, the reason and the result.
- Each service is certified against your rules before it goes live
- Change the rules, and the certificate must be issued again
- Runs in your environment, so your data stays yours
Pricing: ServiceOrbit is scoped and quoted for each organization, usually starting with a pilot. It is not on the AxiomAI plans.
The working product on invented organizations. Every person and record shown is invented.