Responsible AI commitment
Powerful AI, safe, traceable and under human control.
AI should help people, automate tasks and improve processes, without removing human responsibility or compromising rights, data or trust.
We support providers outside the EU, and on Enterprise, at the client's request, we limit your workspace to providers that process its data in the European Economic Area.
Code of conduct
Clear rules.
What we demand of our AI, our team and any use of the platform.
- Legality and ethicsLawful, proportional use, aligned with people's rights.
- Data protectionMinimal data, clear purpose, security and GDPR-compliant processing.
- ConfidentialityNo unauthorized tools for customer information, credentials or business secrets.
- People and employmentAI is never used to fire, sanction, discriminate or critically evaluate people automatically.
- TransparencyNo output is presented as infallible and sensitive uses are reviewed first.
- Unacceptable usesManipulating, impersonating identities, generating fraud or accessing systems without authorization.
How we enforce it
Real controls, inside the product.
Not a policy in a drawer. Every commitment maps to a concrete mechanism in the platform.
Minimum permissions
Each agent accesses only the tools, data and actions its task needs.
✓ Tools assigned per agent
Roles and workspaces
Users, teams and connections are configured with differentiated permissions.
✓ Per-workspace permissions
Approved agents
Recurring tasks run through configured agents and controlled flows, not improvised instructions.
✓ Every scheduled task starts from an agent
Auditable record
Reasonable evidence of instructions, relevant actions and reviews is preserved.
✓ A trace for every run and action
Risk-based validation
Sensitive actions require human review, specific rules or preventive blocking.
✓ Per-tool approval · human handoff
Detection and escalation
The platform can warn, block or escalate instructions that violate this commitment.
✓ Runtime blocks and alerts
Transparency
You know when an AI is acting.
How the platform tells people an AI is acting and how it marks what it generates, under Article 50 of the EU AI Act.
An AI notice per channel
An admin sets on each channel the notice that the contact is talking to an AI. The platform checks the first message and, when the agent does not convey it, sends the notice first.
✓ Notice per channel · first message check
Content Credentials (C2PA)
Images, video and audio the platform generates carry a signed, timestamped C2PA manifest, the open standard for content provenance.
✓ C2PA signature with a key in AWS KMS
Marked in several layers
The file itself declares in its metadata that an AI generated it, and images, video and music from Google keep its invisible watermark.
✓ IPTC trainedAlgorithmicMedia · SynthID
On calls too
On a call, the agent opens with the channel's AI notice, and the platform checks the transcript afterwards to confirm it was said.
✓ Call opening · check after the call
Only what the AI made is marked
A file a contact or a teammate sent, or one the agent only downloaded, goes out as it arrived.
✓ Marking follows who made the content
No scoring of people
No agent in our catalog evaluates, scores or selects people for employment, education, credit or essential services.
✓ Catalog policy for high-risk uses
Agents in production
Tested first, watched after.
What keeps an agent within what you asked of it, from the first test to the daily run.
Rehearsals with no consequences
A rehearsal answers every outward action without carrying it out, so an agent is tested on scenarios without writing to anyone or changing data outside the platform.
✓ Rehearsal with scenarios
Rules of conduct
On Enterprise, every agent in your workspace receives its rules above its own instructions, with a history of who changed them. They are policy, not a switch: the platform's own protections always run.
✓ Code of conduct per workspace
Supervision you choose
A supervisor can watch each automation: it suggests fixes or applies them within its limits, and leaves the key decisions to you.
✓ Supervision: off, suggest or autonomous
Spend and loops in check
The platform learns each agent's usual spend and stops a run that spikes, and halts loops that repeat, fail or stop making progress.
✓ Spend guard · loop detector
Continuous evaluations
Evaluation cases score an agent's real runs and open a notice when one fails them.
✓ Sampled evaluation cases
Daily review
Failures, anomalies and thumbs down are reviewed every day, and a person decides on the improvements that come out of it.
✓ Daily review with proposals
Design and operation
It applies to us and to our products.
- Security by design
- Limited access, authorized connections, traceability and control of relevant actions.
- Customer control
- You define tools, permissions, users, agents and supervision levels.
- Useful automation
- Ordinary business tasks are automated when approved rules and agents exist.
- People in charge
- AI proposes, prepares or executes under rules. Responsibility never disappears.
- Risk governance
- We distinguish low-risk tasks from sensitive actions by legal, labor or reputational impact.
- Continuous improvement
- We review usage, incidents and configurations to reinforce security, quality and trust.
Commitment to clients and partners
We don't sell AI without control.
We offer configurable, traceable agents governed by rules. AI can automate tasks, but it cannot remove human oversight, security or respect for people's rights.
Powerful AI, but always safe, traceable and under human control.