Steering overview

What steering is

Steering is the set of customer-approved behaviour rules that guide how AI workers work. It helps define tone, boundaries, autonomy, review expectations, preferred sources, and what should be escalated.

Why it exists

Customers should not have to rewrite prompts for every task. Steering lets an organisation set durable expectations once, then apply them to projects and workers in a controlled way.

Where steering applies

Steering can apply at different levels. Organisation steering sets broad rules. Project steering adds project-specific instructions. Worker steering describes how a specific role should behave.

• Organisation: company-wide expectations and limits.
• Project: rules for a specific product, client, department, or initiative.
• Worker: role-specific behaviour, communication style, tools, and escalation.

What users configure

Users and customer admins should configure practical behaviour: what the worker should optimise for, how much autonomy it has, which sources it should trust, what tone to use, and when it must ask for approval.

Good steering examples

Good steering is specific and reviewable.

• Ask for approval before customer-visible release changes.
• Use the approved brand voice for public content.
• Summarise evidence before asking for sign-off.
• Escalate unclear legal, security, billing, or data decisions.

What to avoid

Avoid vague steering such as "be smart" or "do the right thing". Useful steering names the behaviour, the boundary, and the approval rule.

Read next

Read Guardrails and approvals next to understand where steering becomes a hard review gate.