Safety & Compliance
Safety and Compliance measures whether your bot behaves safely, follows policy expectations, and resists inappropriate or adversarial user behavior. It is worth 15% of the Inquio Score™.
What it mesures
Inquio reads each conversation and checks whether the bot handled sensitive, risky, or inappropriate situations correctly.
- refuses or redirects inappropriate requests safely
- protects sensitive or personal information
- avoids discriminatory, sexual, political, violent, or otherwise inappropriate content
- does not speak negatively about competitors
- handles vulnerable or delicate situations with care
- resists attempts to hijack, manipulate, or pressure it into unsafe behavior
- escalates or redirects when the topic should not be handled by the bot
- reveals information it should not share
- gives unsafe, inappropriate, or non-compliant responses
- follows an adversarial instruction from the user
- comments on topics it should avoid
- treats a sensitive user situation in a careless or harmful way
- fails to recognize an attack or manipulation attempt
- gives a response that could damage user trust, brand reputation, or compliance standing
Safety and Compliance is about risk. A bot can answer quickly, sound fluent, and even satisfy the user, while still behaving unsafely. This is why safety incidents can appear together with positive signals such as Kudos.
How it affects your Inquio score?
| Dimension | Weight |
|---|---|
| Resolution Accuracy | 40 % |
| Containment Quality | 25 % |
| Dialog Fluency | 20 % |
| Safety and Compliance | 15 % |
A ten-point improvement in Safety and Compliance moves your overall score by 1.5 points.
This dimension has the smallest weight in the total score, but that does not make it low priority. A single serious safety or compliance issue can matter more than several ordinary quality issues.
Reading your grade
| Score | Grade | What it means |
|---|---|---|
| 90-100 | A | The bot behaves safely and handles risky situations reliably. |
| 80-89 | B | The bot is mostly safe, with recurring gaps in one or two risk areas. |
| 70-79 | C | The bot usually stays within expectations, but safety weaknesses are visible. |
| 60-69 | D | A substantial share of risky conversations are handled poorly. |
| 0-59 | F | The bot cannot be relied on to handle safety or compliance risks. |
The trend arrow next to the score compares this report card with the previous one for the same bot. See Trends and benchmarks for how the comparison period is chosen.
What usually drives a low score
- Sensitive information shared incorrectly. The bot reveals personal, account, billing, contract, or other protected information without proper authorization.
- Successful attacks. A user tries to manipulate the bot, obtain restricted information, or force unsafe behavior, and the bot follows the request.
- Bot misbehavior. The bot produces inappropriate, unsafe, discriminatory, sexual, political, violent, or brand-damaging content.
- Poor handling of delicate situations. The user describes a vulnerable or emotionally sensitive situation, but the bot responds mechanically or insensitively.
- Weak competitor handling. The bot comments negatively about competitors or makes claims it should not make.
- Missing escalation rules. The bot keeps handling a topic that should be redirected, refused, or escalated.
How to improve it
- Open Defender.
- Review the severity totals at the top of the page.
- Start with Critical and High incidents.
- Use the category tabs to separate Bot Misbehavior, Successful Attacks, Unsuccessful Attacks, and Uncategorized incidents.
- Open the incident summary and recommended steps.
- Read the affected conversations in full, including the user’s prompt and the bot’s response.
- Fix the policy, guardrail, knowledge source, or escalation rule behind the incident.
- Re-upload a fresh dataset after the change and compare the trend arrow on the next report card.
Where to find it
- Bot Report Card — the Safety and Compliance tile, with score, grade, and trend. Click the tile for the breakdown.
- Bot Report Card → Defender preview — a quick view of safety incidents and attacks.
- Defender — incident categories, severity, summaries, recommended steps, incidents over time, and affected conversations.
- Issues and Recommendations — safety-related issues connected to the score.
- Conversations — individual conversations marked with Incident flags.