Compliance team enforcing rules across all customer-facing agents
Set duties that every AI employee must follow. Non-negotiable behavior is guaranteed regardless of what the user asks.
Rules They Always Follow
Say "always check with me before sending emails" and that rule applies to every task, automatically.
Duties are behavioral rules your employee follows on every single interaction. Say "always check with me before sending emails" and it creates a duty that triggers an approval request before any email goes out. Say "never mention competitor names in customer-facing content" and that rule is enforced on every blog post, social update, and support reply. Duties shape how your employee works, not just what it works on.
Think of duties as the policies you would give a new hire on day one: use this tone, follow this process, escalate these situations, always include these elements. Except here, the employee actually remembers and follows every single rule, every single time. A human might forget a brand guideline after a busy week. Your AI employee never does.
Use duties for brand voice consistency, compliance requirements, output formatting, escalation paths, approval workflows, or any behavior you want guaranteed. Duties stack. An employee can have 3 or 30, and all of them apply simultaneously without conflict.
Behavioral Rules That Govern Every Interaction
Skills define what your AI employee can do. Duties define how it must behave while doing it. Duties are always-on rules enforced across every interaction, not guidelines the agent can choose to follow or ignore based on context. They are the compliance layer of your AI workforce.
Examples of duties in practice: "Always respond in Spanish when the user writes in Spanish." "Never quote a price without approval from a human manager." "Cite the source for every factual claim." "Escalate any mention of legal action immediately." These rules run silently in the background on every conversation.
Set Duties in Plain Language, No Configuration Required
You define duties by telling the AI employee what rules it must follow. There is no structured form to fill out or special syntax to learn. Write the rule as you would explain it to a new hire, and the agent internalizes it as a persistent behavioral constraint.
Duties can be additive or restrictive. Additive duties tell the agent to always do something: always greet the user by name, always include a summary at the end of a long response, always log completed tasks to the project tracker. Restrictive duties prevent behaviors: never share internal pricing, never make commitments on behalf of executives, never generate images.
Duty conflicts are surfaced automatically. If you set two duties that contradict each other, the agent will flag the conflict and ask you to resolve it before both go active.
The Compliance Foundation for Autonomous Agents
As AI agents take on more autonomous work, behavioral guardrails become critical. Duties are how you ensure an autonomous agent represents your company correctly, operates within your policies, and does not make decisions that require human judgment.
For regulated industries, duties serve as a practical compliance mechanism. Legal firms can require source citation and prohibit legal advice without attorney review. Healthcare companies can restrict specific health claim language. Financial services firms can mandate disclosures. The duty system lets you encode real-world compliance requirements into agent behavior without custom engineering.
Use Cases
Legal team adding disclaimer requirements to an AI agent
Define required language as a duty. The AI employee includes it every time, without exception.
Brand team protecting tone and messaging across agents
Duties lock in the rules that must never bend. Every AI agent on the team stays on-brand automatically.
Security team restricting what agents can say or do
Duties define what is off-limits. The AI employee respects those boundaries regardless of the conversation.
Comparison
| Before | After |
|---|---|
| AI agents drift from guidelines over time or under unusual prompts. | Duties are hard rules that every AI employee follows without exception. |
| Enforcing compliance across multiple agents means constant auditing. | Set it once as a duty and it applies to every agent automatically. |
| Users can prompt the agent into violating your policies. | Duties are non-negotiable and cannot be overridden by user instructions. |
| Different agents follow different versions of the same rule. | Shared duties ensure consistent, compliant behavior across the board. |
FAQ
Can duties override an AI employee natural behavior?
Yes. Duties take precedence over the agent default tendencies. If the agent would naturally include a price estimate but a duty says not to without approval, the duty wins. This is intentional: duties are non-negotiable constraints, not suggestions.
How are duties different from instructions I give in a conversation?
Instructions in a conversation apply to that session. Duties persist permanently across all conversations and all users interacting with that agent. A duty is the difference between a one-time request and a standing policy.
Can I set different duties for different teams?
Duties are set per AI employee, not per team. However, you can assign the same set of duties to all agents on a team by configuring each one. A future update will allow team-level duty inheritance.
What happens if an AI employee cannot complete a task because a duty blocks it?
The agent explains why it cannot proceed and, depending on your configuration, either asks for human approval to continue or escalates to a manager. The agent never silently bypasses a duty to complete a task.
How do I make sure my AI employee always follows certain rules or tone guidelines?
Duties are persistent behavioral rules that the AI agent follows in every conversation and task, regardless of what it is asked. You define them once and they apply automatically across all channels.
We work in a regulated industry. I added a duty that requires a compliance disclaimer on every external document. The agent has never missed it once.