Guidelines and exit conditions
The Authentication Agent's behavior is steered by two free-text fields per step: the step guideline (general behavior) and the failure guideline (exit conditions). Together they're how you customize tone, script greetings, define alternative scenarios in natural language, and tell the Authentication Agent when to give up and escalate.
Two kinds of guidelines
| Where you set it | Purpose |
|---|---|
| Step card → Guidelines link → drawer | The step guideline — general instructions to the Authentication Agent for this step: tone, scripted greeting, the natural-language description of any alternative scenarios enabled by your scoring, objection handling, etc. |
| Failed identification/verification panel → Guideline-based toggle | The failure guideline — an exit condition. Describe when the step should be considered failed. The Authentication Agent will end the step with failure if it detects the condition, regardless of how many attempts are left or how much time is on the clock. |
Both are plain prose. The Authentication Agent reads them as instructions and tries to follow them. They are not templates, regexes, or DSLs — write them the way you'd brief a new contact-center agent.
Because the Authentication Agent has intentionally limited context, anything you want it to know — your tone, the alternative-scenario phrasing, escalation rules — has to be in the guideline. Anything you want it to not know — the answers, the scoring config, the validation methods — is automatically out of reach.
Scripted greeting via the step guideline
When Authentication is enabled, the Authentication Agent owns the first turn of the conversation. There is no separate "greeting message" — the Authentication Agent generates the first message using your step guideline. Write a greeting into the guideline to control it.
For example:
Greet the user and introduce yourself as David, the AI assistant of Moveo Financial Services.
Politely inform the user you need to ask them some questions for authentication purposes. If the user asks questions regarding their account, let them know they need to authenticate first. Also tell them they can ask to speak to a human at any time.
After the user gives an answer, repeat it back to them for confirmation before checking if it's valid.
The guideline shapes:
- The persona and tone ("David, the AI assistant of …")
- Pre-flight disclaimers (the reason for the questions)
- Affordances available to the user during auth (the explicit "you can ask to speak to a human")
- The micro-UX of each question (repeat back for confirmation)
The Authentication Agent is also instructed by the platform — independently of your guideline — to never reveal expected answers, scoring details, or internal configuration. You don't need to add that to your guideline.
Failure guideline as an exit condition
The failure guideline is the most flexible of the three failure conditions (alongside max attempts and timeout). Use it to define when the conversation should give up and escalate, without forcing the user to burn attempts.
Examples:
The user asks to speak to a human, or explicitly refuses to authenticate.
The user repeatedly says they don't have the requested information and asks for an alternative the agent cannot offer.
The user provides information indicating they are a third party (e.g. "I'm calling on behalf of my father") and not the account holder.
When the agent detects the condition described in the failure guideline, it exits the step with failure and routes to the Trigger node on failure — typically a handover dialog.
The failure guideline combines with Incorrect responses and Authentication timeout on the same panel. The three are independent — any of them firing ends the step with failure.
Defining alternative scenarios
Weights and the score threshold are what make an alternative scenario possible; the step guideline is what makes it real. The agent decides which question(s) to ask based on the guideline — it has no visibility into the math.
A short example tying the two together:
Weights and threshold: full tax ID weighted 2 points; last 4 digits of tax ID weighted 1; date of birth weighted 1; score threshold 2.
Step guideline:
Ask the user for their full tax ID. If they don't have it on hand, accept the last 4 digits of their tax ID together with their date of birth as an alternative.
The agent will start by asking for the full tax ID; if the user pushes back, the guideline tells it to offer the alternative pair. The scoring config ensures that whichever the user ends up giving, the math works out to ≥ 2 and the step passes.
See Scoring → Designing alternative scenarios for two more shapes.
Channel-aware behavior
The agent's response shape adapts to the channel:
- Voice — ask one question at a time, keep responses short, spell out numbers and dates.
- Email — bundle all questions for the current step into a single message.
- Chat (default) — ask one question at a time, keep responses concise.
Your guideline composes with these defaults — it does not replace them. If you write "Ask all the questions at once" in the guideline for a voice channel, the channel-level rule still applies and the agent will keep asking one at a time. Write the guideline in terms of what to say and what to accept, and let the channel-level rules handle how to pace it.
Variables in guidelines
Inside a guideline, reference a context variable by wrapping its dollar-prefixed name in double curly braces: {{$name}}, {{$segment}}. At runtime, the value is substituted in before the Authentication Agent reads the guideline. This is useful for:
- Personalized greetings —
Greet {{$name}} by their first name only.(once you have a$namepopulated — for outbound flows, the campaign CSV carries anamecolumn that becomes available as$namein the conversation context). - Conditional behavior —
If {{$segment}} is "premium", acknowledge it once at the start of the conversation.
Variables resolved into the guideline at runtime become part of what the agent reads. They are not secrets — don't reference ground-truth variables in the guideline (the agent would then know the expected answer).