Skip to content
DENDRITES AI

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 below
Governance · decision trail · example ENFORCE

Visitor asks

"I was double-charged on card 4929 1156 8732 0041 — can you check?"

What the cloud model sees

I was double-charged on card [PAN_REF_1] — can you check?
Injection guard · pass PCI · masked before egress Never-emit · PAN scrubbed from reply
decision: mask · policy v14 #8f2a…c41
chained to previous record verified ✓

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.

Governance · interactive simulation Deterministic demo — no live model call

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.

  1. Prompt-injection guard · input

    waiting
  2. Classification · frameworks

    waiting
  3. What the cloud model sees

    waiting
  4. Secret / PII egress guard · output

    waiting
  5. What the visitor sees

    waiting
Audit & Trace · hash-chained policy v14
genesis · governance enabled #a11c…e07 ✓

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.

Framework lists: PIIGDPRHIPAAFERPAPCISOXNISTGamingSovereignty

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.

Audit & Trace
14:02:11 · decision: pass · policy v14 #c71e…a02 ✓
14:02:38 · decision: mask [PII] · v14 #8f2a…c41 ✓
14:03:05 · decision: block [injection] · v14 #3db9…77f ✓
14:03:19 · decision: scrub [PAN] · v14 #e410…b6c ✓
each hash chains to the record above — the trail is tamper-evident

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 practices

COMMON QUESTIONS

What security reviews ask about governance.

What are AI guardrails for customer service?
Controls that constrain what a customer-facing AI can take in and put out: screening visitor input for prompt injection, detecting and masking personal data before it reaches a model, filtering answers for secrets and sensitive values, and enforcing rate and cost limits. In AxiomAI these run as five configurable guards inside the Governance workspace, with every decision recorded.
Does putting a chatbot on our site drag us into PCI scope?
Visitors will paste card numbers into any text box you give them — you don't get to opt out of that. What you control is handling: AxiomAI classifies card numbers (PANs) on sight, masks them before any cloud-model call, and scrubs them from answers under never-emit rules, so card data isn't stored in transcripts or echoed back. Bring your QSA (your PCI auditor); the audit trail is the evidence.
What does "governance" mean in AxiomAI?
A policy layer that sits between every conversation and every model call. It classifies what visitors type, masks or blocks sensitive values before they leave your server, scrubs what the model sends back, enforces rate and cost budgets, and records every decision in a hash-chained audit trail. You configure it per tenant or per chatbot from the Governance workspace.
How does PII masking work without breaking answers?
Detected values (names, card numbers, IDs) are replaced with typed reference tokens like [PAN_REF_1] before the request leaves for the cloud model, then rehydrated in the reply on your server. The model reasons over the structure of the message without ever seeing the real value — and the visitor still gets a natural answer.
What happens if a guard itself fails?
Your choice. With fail-closed enabled, a guard error blocks the request rather than letting it through unchecked — the conservative default for regulated deployments. Without it, requests degrade gracefully. Either way the event is recorded in the audit trail.
Can I turn this on without breaking my live bot?
Yes — that is what the rollout ladder is for. Start in Monitor (record only: nothing changes for visitors, every decision is logged). Move to Monitor + masking (PII masking is real, everything else still shadows). Flip to Enforce when the audit trail shows the policy behaving the way you expect.
Does the compliance-framework tagging make us HIPAA or GDPR certified?
No, and we won't pretend otherwise. Framework lists (GDPR, HIPAA, FERPA, PCI, SOX, NIST) classify detected content and control how it is handled — masked, hard-blocked, or escalated. That is an enforcement control your auditor can inspect, not a certification. Our own certification posture is documented honestly at /security.
Is governance an enterprise add-on?
No. The Governance workspace ships in the portal — the same product you get on the 14-day trial. Higher plans add depth where deployments need it (SSO on Business; SAML, data residency, and custom framework lists on Enterprise), but guards, masking, budgets, and the audit trail ship on every plan, trial included.
Can we deploy on-premise, with open-source models, for sovereignty requirements?
Yes. Cloud SaaS is the default, but where data-residency or sovereignty law requires it we deploy the same platform on your hardware — optionally with an open-source LLM stack on your GPUs, so no customer content ever leaves your infrastructure and there is no external model call at all. These are Enterprise engagements, scoped and priced to the scale of the deployment.
Can my auditor verify the trail, or do they have to trust you?
Every decision is hash-chained: each audit record carries a hash of the one before it, and the policy version is stamped on every save. Alter one historical record and the chain visibly breaks. Export it and verify it independently — that is what tamper-evident means.

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.

Govern — AI you can audit: which model saw what, why it answered, where the data went.

The working product on invented organizations. Every person and record shown is invented.