AI deployment can involve model selection, integrations, monitoring, safety controls, data work, and organizational review. A configuration wizard can simplify part of that process, but it does not remove the need for engineering or governance where the use case requires them.
Where the Agent Creator is enabled, an authorized admin can configure an agent profile and policy. Completion time and required expertise depend on the role, connectors, data, and risk tier.
The 5-Step Wizard
Step 1 — Persona: Give your AI employee a name (Priya, Arjun, Maya), a designation (Senior AP Analyst - Mumbai), and assign it to a domain (Finance, HR, Marketing, Ops). This isn't cosmetic — the name and persona appear in audit logs, approval requests, and the agent fleet view.
Step 2 — Role: Choose an available agent type or define a new one. Specialization and routing filters are configuration inputs whose behavior must be tested before live use.
Step 3 — Instructions: Start from an available template or write custom instructions. Templates can include processing steps, escalation rules, safety constraints, and output schemas, but they are not production evidence on their own.
Step 4 — Behavior: Set the confidence floor (below which the agent escalates to a human), HITL conditions, LLM model selection, and retry policy.
Step 5 — Review & Evaluate: Review the configuration and, where verified non-writing evaluation is available, run representative tasks before considering promotion.
Multiple Agents, Same Role
The data model can represent multiple agents of the same type for regional or functional variations when that capability is enabled.
Illustrative example: Priya is scoped to domestic invoices, Arjun to imports, and Maya to a subsidiary. Their amounts, connectors, and routing behavior are hypothetical and require tenant-specific configuration and testing.
Prompt Templates: The Agent's Training Manual
A useful template structure includes identity and scope, processing sequence, escalation rules, prohibited behavior, and an output schema. Each template still requires use-case testing.
Templates use {{variable}} placeholders (e.g., {{org_name}}, {{hitl_threshold}}) so the same template works across different organizations and configurations. Built-in templates are read-only — clone them to customize.
From Shadow to Production
A governed release policy should start an unproven agent in a non-writing state, compare representative results with an approved baseline, and require explicit promotion evidence. No example accuracy or sample count guarantees readiness.
Pause and rollback are important control objectives. Verify their authorization, latency, scope, state handling, and audit evidence in the deployed environment.
Topics
Ready to try it?
Explore the public playground, or create an account to evaluate an agent with your own approved data and controls.