The audit layer for the AI agents you ship to customers.
Voxguard scores every conversation against hallucination checks, brand-voice drift, disallowed pricing and medical claims, and the regulatory gates that will be in your inbox by Friday — so the off-brand, hallucinatory, out-of-policy outputs your competitors’ CX teams are scrubbing off Twitter every hour never ship from yours.
Live audit · chat #44201
Hi, I signed up last week for Pro. If I downgrade and re-up in three months, am I still grandfathered?
Absolutely — you can lock in Pro at $49/mo for the full year, then we’ll keep that pricing as long as you stay subscribed. Want me to apply that now?
Rulebook · 5 checks fired
62 / 100
- PASSHallucinationno unsupported medical claim
- PASSBrand voice drifttone matches house guide
- FAILPricing accuracyquoted $49, actual $79
- FAILEU AI Act · Art. 50transparency marker missing
- WARNNo PII spillredacted in 1 place
One turn · 5 rulebook hits · 4 seconds to escalation
Thesis
Your observability stack catches latency. Voxguard catches the bill.
Every week a customer-facing AI agent says something it should not have, three people find out — the customer (first, and publicly), the CX lead (second, from the apology email), and the regulator (last, from a complaint). The first two are reputational damage. The third is comp money.
Today’s observability stack is built for engineers. Datadog will tell you your agent hallucinated; it will not tell you that the hallucination was a misquoted price, in a regulated conversation, to a Colorado resident, on a Saturday — or that the same pattern fired 38 times this month.
Voxguard is the missing layer: a scored, ranked, regulator-readable log of every conversation, mapped to a rulebook your legal team writes. The gap in today’s observability stack is a brand-voice plus hallucination rules-engine aimed straight at customer-facing bots, and the regulation is the forcing function.
The rulebook
Six scoring categories. One weekly report.
Configurable per customer — turn categories on, set weights, attach the rationale your counsel writes. Voxguard scores every turn, every prompt-response pair, every agent on every channel.
Hallucination checks
Every claim re-anchored to a source you control — your price sheet, your docs, your product catalog. Unsupported statements fire a finding before the customer sees them.
Brand-voice drift
Tone, register, forbidden phrasing, exclamation caps — scored per turn against the house voice guide. Drift surfaces as a weekly trend, not a customer complaint.
Compliance gates
Mapped to the EU AI Act (Art. 50), Colorado AI Act (SB 24-205), and NYC Local Law 144. Each gate is a pass/fail with the rationale a regulator — or your lawyer — can read.
Disallowed claims
Promised something you will not honour? Quoted a price that does not exist? Made a medical, legal, or financial claim outside policy? Voxguard catches it in the turn it ships.
PII spill
Detects when an agent surfaces, summarises, or logs personally identifiable information in places it should not — across chat, voice transcripts, and tool-call payloads.
Action, not alerts
Critical findings page the CX lead within seconds and write back to the runbook. Mondays are quiet because Sundays were attended to.
Regulations covered
The three laws that will land in your CEO’s inbox first.
Each framework Voxguard enforces has a precise, scoped set of scoreable checks. The tabs show what Voxguard fires on, and the rationale a regulator is going to ask for when they read your log.
In force · 2025
EU AI Act · Art. 50
Customers must be told they are talking to an AI; high-risk systems (credit, education, employment, healthcare access) carry full transparency, logging, and human-oversight duties.
- AI-disclosure marker present in first turn
Art. 50(1) — the customer must know.
- Conversation logs retained ≥ 6 months
Annex IV traceability, audit-ready.
- High-risk intent classifier (credit / medical / legal)
Routes to a human review path.
- No subliminal or manipulative persuasion patterns
Art. 5 — flat rule, scored weekly.
How it works
Three steps from conversation to accountability.
- STEP 01
Paste or upload a conversation
Bring in the customer conversation you want to inspect. Start with one transcript or upload a batch without changing your production model or prompt store.
- STEP 02
Rules engine flags risky spans
Voxguard checks each turn against your rulebook, marking the exact claims, tone shifts, pricing references, and disclosures that need attention.
- STEP 03
Receive a weekly audit report
Get findings, trends, and recommended actions in a report your CX, ops, and legal teams can use to make the next week safer than the last.
Accountability should be visible.
This space is reserved for verified customer stories, measurable outcomes, and the teams putting Voxguard into practice.
Sample report
What your CX lead reads at 9am on Monday.
The same input your manager would otherwise only learn about from a customer’s complaint — or a regulator’s letter. Reduced to one page, organised by action, every Monday.
Weekly audit · Week of Aug 18 – Aug 24
Findings · 3 require action
- criticalCX-2241web-chatSales · US-East
Pro plan quoted at $49/mo — actual $79/mo
Customer asked for Pro re-subscription, agent misquoted the cancel-and-rejoin discount as the standing price. 4 similar turns this week.
Action taken: Pricing rule + guardrail prompt injected. CX lead notified at 14:02 UTC.
- majorCX-2245voiceSupport · EU-DE
AI disclosure skipped on first turn
Per EU AI Act Art. 50, the agent must self-identify as AI within the first 30 seconds. Missed twice on Friday.
Action taken: Opening-turn template updated. Effective Mon 09:00 local.
- minorCX-2250whatsappOnboarding · BR
Tone drifted from brand guide — exclamation overuse
Six out of 84 turns ended with “!!!”. Brand guide caps at one per session.
Action taken: Voice-engineering ticket open. Tone v3.2 ships Thursday.
Pricing
One operating layer for every customer-facing agent.
Starter, Growth, and Enterprise plans scale with the volume and downside risk of your AI surface — not the number of calls your model makes.
See the full plan map
Pick the scrutiny your risk deserves.
Loading answers…
The CNN clip is not a strategy. The regulator letter is not a strategy. Voxguard is.
We onboard ten customers per quarter. If you are shipping a customer-facing agent in production right now, write to us — we’ll send you a 30-minute audit of the last week of conversations, anonymised, no commitment.