Procedures
A guide to writing Procedures
Overview
Procedures tell the AI how to answer the customer, not what internal actions to perform. Each procedure starts with a short trigger line:
When a guest asks … if … then …
Below are the main patterns you can use:
Pattern
Description
When to use
Single If‑Then (Binary Rule)
One condition, one response.
Simple, clear‑cut scenarios.
Multi‑Tree Rule
Several if branches beneath one trigger.
Same guest request with multiple possible conditions/outcomes.
Workflow‑Style Rule
Step‑by‑step guidance with checks in sequence.
Troubleshooting or multi‑step support flows.
Templated Response Rule
Picks a ready‑made reply template based on conditions.
Fast, consistent messaging with variables (e.g., guest name, code).
Key points
Focus on the reply. Describe what the AI should say or whether it should escalate to human not back‑office tasks.
Stay conditional. Use clear if clauses so the AI knows when to apply the rule.
Keep each rule self‑contained. No need to reference other Procedures.
💬 Where to Find Procedures
To create or manage your AI agent’s decision-making logic, go to:
Agent Hub → Wiki → Procedures
From here, you can define specific conditions and instructions that guide how your AI should respond based on message type, reservation details, or other triggers. Procedures ensure your agent responds with the right message at the right time.
All Procedures defined here apply only to the selected inbox.
Procedure Templates
Conduit has preset templates for common procedures in each industry. As a user you will be able to fully customize each template in your workspace, so it fits your business' needs.
To find to Procedure templates
Navigate to Agent Hub > Knowledge > Wiki > Procedures > Add Procedure
Select Import template and choose between Conduit's predefined procedures
Good vs Bad Procedures
❌ Vague Condition
“If needed, let the guest check out late.”
The AI doesn’t know what “needed” means.
✅ Clear Condition
“If a guest requests late check-out and no new guest is arriving the same day, approve check-out until 13:00.”
❌ Internal Action
“If the guest reports damage, notify the operations manager and create a task.”
The AI cannot perform back-office actions.
✅ Guest-Focused Reply
“If a guest reports damage, thank them for letting you know and tell them the team has been informed and will follow up if needed.”
❌ Missing “If” Clause
“Offer a 20% discount for repeat guests.”
No context is given for when to use this.
✅ Conditional Statement
“If a guest mentions they’ve stayed before, offer a 20% discount on their next booking.**”
❌ Rule That’s Actually Knowledge
“Pets are not allowed in the apartment.”
This is just a static policy — no condition, no guest message trigger. It belongs in Knowledge.
✅ Correct Rule
“If a guest informs you they’re bringing a dog and it’s a service animal, reply that certified service animals are allowed despite the no-pet policy, and ask if any special arrangements are needed.”
❌ Rule That’s Actually a Guardrail
“Do not disclose any pricing differences between platforms.”
This is a permanent restriction on what can be said, not a situational reply.
✅ Correct Rule Version
“If a guest asks why the price on another site is lower, explain that prices may vary across platforms due to different fees or promotional procedures.”
✅ Correct Guardrail Version (belongs in Guardrails)
“Never mention price differences between platforms in guest communicatio”ns.”.”*
❌ Ambiguous Outcome
“If the place is cold, fix the heating.”
No clear instruction on what to tell the guest.
✅ Guest-Facing Response
“If the guest says the apartment is too cold, ask if the windows are closed and the thermostat is set above 21°C. If not resolved, tell them the team has been alerted.”
❌ Non-verifiable Condition
“If a guest asks to check in early on the day of arrival, and the cleaners are done, let them check in.”
This assumes the AI knows the real-time cleaning status, which it doesn’t — so the rule cannot be followed reliably.
✅ Trigger Based on Guest Input
“If a guest asks to check in early, let them know you'll check with the cleaning team and confirm as soon as the apartment is ready.”
How the Reasoning Engine Uses Your Procedures
How Procedures are processed
The AI’s reasoning engine reads all your Procedures, matches the if conditions against reservation data and chat context, and selects the rule(s) that fit.
It performs any time or price calculations needed (e.g., “hours until arrival”, “percentage refund”).
The engine then drafts a reply that follows the chosen policy, applies Style, and finally passes through Guardrails before sending.
Sample Internal Trace – Early Check-In Question
FAQ
1. What information can I refer to in Procedures?
Procedures can reference reservation details (dates, payment status, guest count), chat history, your Knowledge base, and calendar availability. We’re adding more data sources soon—feel free to suggest priorities.
2. When should I use templated replies?
Templated replies should be used carefully and only in specific situations where an exact response is required. Since these responses are sent verbatim to users, they need very precise "if conditions" to prevent misuse or inappropriate context.
Key considerations for templated replies:
Only use them when specific wording is mandatory (legal requirements, company policies)
Always include detailed context for when the template should and shouldn't be used
Provide clear "if conditions" that must be met before using the template
Consider adding fallback options for cases where the template might not fully apply
Example of a well-defined templated rule:
"If a guest requests a VAT invoice AND has completed their stay AND has paid in full, use template: 'Thank you for requesting a VAT invoice. I'll prepare this for you within 24 hours. Please confirm the company details to be included: company name, address, and VAT number.'"
3. How do I test that the AI follows a Procedures?
Open Sandbox, adjust reservation details in Settings, and impersonate the guest. Send the triggering question and verify the reply.
Last updated
Was this helpful?