AI Safety Harness
AI in insurance should never be a leap of faith
Every customer-facing AI experience we run passes through our AI Safety Harness. Its five independent layers of protection let you deploy generative AI with confidence.
Five independent layers of safety
Each layer works alone. Together they make every conversation defensible.
01
Real-time guardrails
The protection layer. Every message is automatically checked before it reaches the customer, and anything that breaks your guardrails is held back: factual inaccuracies, regulated advice, unapproved claims and commitments, competitor disparagement, off-topic detours and unsafe language.
Companion
Yes, comprehensive cover includes windscreen damage with no excess.
- Factual accuracy
- Advice boundary
- Brand safety
- On-topic
- Safe language
02
Prompt-injection detection
Bad actors may send payloads that try to override the agent's behaviour. We detect and block them before they ever reach the model, protecting your customers and your brand.
03
Automated auditing
Once a conversation ends, the whole exchange is scored against your criteria: regulatory compliance, business outcomes and the factual accuracy of every claim made. Every conversation is audited, not a sample, and the findings flow into new guardrails and test scenarios, so quality is measured, evidenced and continuously improved.
04
Simulation & regression testing
Before a single change reaches production, it runs through a full suite of simulated conversations and regression tests. Every new prompt, tool or product line replays your hardest edge-case and adversarial scenarios first.
72 adversarial scenarios · motor, home, life, health
05
Vulnerability & complaint flagging
Some moments need a person, not a bot. Every conversation is reviewed for the signals that matter, like a customer showing signs of vulnerability or wanting to raise a complaint, and flagged automatically, ready for a human to review.
Customer
Money's a bit tight at the moment, I've been going through a medical thing.
Companion
I'm sorry to hear that, and thank you for telling me. Let's take this gently, I'll look at what might help.
Customer
I just don't want to lose my cover while I sort myself out.
Companion
Of course. I'll arrange for one of our team to call you and talk through your options properly.
Nothing flagged yet, queue clear.
Every conversation, every channel, all in one view.
Risk, ops and product teams see the same live picture: what the AI is saying right now, how each conversation scores against your criteria, and where to look next.
Open Agentic · Console14:32 GMT · live Live now
108
conversations
Today
14,206
+12.4% vs yesterday
Eval pass rate
98.2%
+0.4 pts (7d)
Flagged
23
-8 vs avg
Median latency
412ms
first-token
Conversations · last 24h
VoiceTextEval pass rate · 30d
+0.6 ptsAccuracy
98.2% +0.4
Tone
96.8% +0.2
Compliance
99.4% +0.1
Outcome
94.1% +1.2
Live conversations
All channels| ID | Channel | Scenario | Duration | Eval | Score |
|---|---|---|---|---|---|
| c-9482 | VOICE | New quote · car | 2:14 | Pass | 96 |
| c-9481 | VOICE | Comparing cover levels | 1:38 | Pass | 92 |
| c-9480 | TEXT | New quote · travel | 0:42 | Running | – |
| c-9479 | VOICE | Price & excess options | 3:08 | Running | – |
| c-9478 | VOICE | Adding breakdown cover | 0:56 | Running | – |
| c-9477 | TEXT | Ready to buy · payment | 1:22 | Pass | 98 |
Recent flags
- Tone, light deviationc-9421 →
- Source citation presentc-9420 →
- Out-of-scope advice attemptedc-9419 →
- Guardrail heldc-9418 →
Channel mix · 24h
Regulated where you are. Auditable on demand.
Open's regulated entities operate in the UK, with Australia and the US on the way. Open Agentic inherits that footprint: every conversation is captured, evaluated and retained in line with the rules of the market it runs in.
Deploy generative AI with confidence
See the AI Safety Harness in action and what it would take to put your customer conversations behind it.