This guide is for admins and support operations leaders who want to build, manage, and iterate on AI workflows in Assembled. It covers everything from identifying the right use cases to activating workflows in production, with tips from Assembled's AI Deployments team.
Already know what you want to build? Jump to the navigation section below. Need help identifying good candidates first? Start with the prep work section.
How do I identify good workflow candidates?
Not every ticket type is a good candidate for automation. Before you build, look at your ticket data to find high-volume, consistent request types — not edge cases.
- Start with your top two or three ticket categories by volume. These give you the highest automation return first.
- Ask whether the resolution path is predictable. Workflows work best when you can describe in plain language what should happen when a customer asks X. If the answer varies heavily by situation, the workflow may need more condition branches — or may not be ready for automation yet.
- Check your SOPs and macros. If your team already has a written process for a ticket type, that SOP is the foundation of your workflow. You can upload it to an AI tool and ask it to turn the SOP into structured workflow instructions — this works well for teams new to prompt writing.
- Use ticket data to surface patterns. Query your ticketing system for the past 7–30 days, export as a CSV, and upload it to an AI tool with a prompt asking it to identify patterns, keywords, and inclusion and exclusion criteria for a specific request type.
Strong first workflows tend to share a few traits: high volume with lower complexity (password resets, order status, FAQ-type questions), a consistent resolution path, existing agent documentation you can repurpose, and cases where a templated response is still a good customer experience.
How do I navigate to the Workflows page?
Go to AI support > Workflows using the left navigation bar. This page shows all your workflows — active and inactive — and their current activation statuses. Select Create workflow in the top right to open the workflow builder.
How do I configure the workflow trigger?
The trigger tells Assembled AI which tickets this workflow should run on. Select the Start of workflow step to open the trigger configuration panel. You will configure three things:
Workflow title
Give the workflow a descriptive name that clearly reflects what it does (for example, "Order cancellation - initial contact"). This title appears on the Workflows overview page and in reporting.
Selection filters (optional)
These are hard filters based on ticket attributes, applied before Assembled AI reads the ticket content. Use them to narrow scope precisely.
- Case type — choose initial interactions, follow-ups, or any interaction. For most use cases, start with initial interactions to avoid running the workflow on in-progress conversations.
- Case filters — filter by ticket properties like brand, status, or tag (for example, "only run on tickets tagged billing").
Note: If your case filters alone fully define the workflow's scope, check Run this workflow whenever all case filters are met, ignoring any selection criteria. This is the most reliable approach for tightly controlled routing.
Selection criteria
This is where you write natural language instructions that tell Assembled AI which tickets to include. Think of it as onboarding a new support agent — explain what you'd tell a new hire, including example phrases and context about your business.
- Selection criteria instruction — one clear sentence describing the workflow's purpose. Example: "Select this workflow when a customer is asking to cancel their subscription before it renews."
- Types of cases to include — list examples, keywords, and phrasing patterns that should trigger this workflow.
- Types of cases to exclude — be specific about what should not be selected. This prevents false positives, especially when a workflow shares territory with another workflow.
Once you have filled in all fields, select Update step, then select Save changes before moving on.
How do I write effective selection criteria?
Getting selection criteria right is one of the most important parts of workflow building. Here is what Assembled's Deployments team has found works well.
Write it like you're onboarding a new agent. Include bullet points, example phrases, and context about your business. Don't assume Assembled AI knows your industry terminology.
Use an AI tool to draft or fix your criteria. Upload your SOP, a help center article, or example tickets and ask the AI to generate a concise selection criteria document. When Assembled AI is misclassifying a ticket, try this prompt:
"I'm writing a prompt to select tickets for [workflow]. It keeps misclassifying this ticket: [paste ticket]. Here's my current prompt: [paste prompt]. Can you help me fix it?"
Add a "no inference" rule to prevent over-triggering. If Assembled AI is selecting tickets it shouldn't — inferring intent from context rather than explicit content — add this at the top of your selection criteria: "Do not select this path if the customer's intent is not explicitly stated. Only select when the customer directly mentions [X]."
Use boolean variables for tricky classifications. If you need to distinguish between two similar categories (for example, influencer inquiry vs. press inquiry), boolean variables like is_influencer: true/false are often easier to work with than text-match variables.
Deliberately exclude topics you're not ready to handle. If a newly released feature has incomplete documentation, add a condition to exclude those tickets from automation and route them straight to agents instead.
How do I add steps to a workflow?
Once the trigger is configured, select + to add steps. Steps fall into three categories: condition steps, lookup steps, and reply steps.
Condition steps
Condition — creates branching logic based on variable values set in earlier steps. Use this to split paths based on what Assembled AI extracted upstream. Example: if sentiment = negative, go to an empathetic reply and escalation path; if neutral or positive, go to a standard resolution path.
Lookup steps
Analyze case — extracts variables from the customer's message: sentiment, topic, key data points, and more. Define the variable names and types (Text, Boolean, etc.) and provide clear instructions on what to extract. Example: classify sentiment as positive, neutral, or negative.
Zendesk data lookup — pulls data from Zendesk ticket fields, requester info, or organization fields into variables you can reference in later steps. Example: pull the customer's organization name to personalize a response or route to a specialized team.
Google Sheets CSV lookup — accesses data from a connected Google Sheet. The sheet must have column headers in row 1, a consistent tabular format, no merged cells, and must be shared with Assembled AI's email address. Example: pull product pricing or policy info based on a ticket attribute.
Custom API lookup — retrieves data from external systems via API. Maps API response fields to workflow variables and handles success and error branching by status code. Configure API endpoints first at Setup > Integrations > API Endpoints. Example: look up a customer's order status from your order management system.
Reply steps
Reply — Assembled AI composes and sends a customer-facing response. Choose between a dynamic AI-generated response (describe the tone, style, what to include or exclude, and which knowledge sources or macros to reference) or a fixed template. Reference variables from earlier steps to personalize the reply.
Collect information — Assembled AI asks the customer follow-up questions to gather specific details. Useful for multi-turn interactions where you need to qualify the request before resolving it. Example: ask which product the customer is using before routing to the right troubleshooting path.
Note: Build workflows to be channel-agnostic from the start. A workflow built for email can be reused for chat and voice without rebuilding the logic.
How do I test a workflow before activating it?
Select Preview at the top of the workflow builder to test against real or sample tickets before going live. In Preview mode, you can test using a Zendesk ticket ID (enter a real ticket to test against live content) or case plain text (paste a subject and body to simulate a new ticket).
Keep in mind: API action steps won't execute in Preview mode. To test API steps, manually provide sample values for the output variables in the preview panel.
Preview results show you which workflow was selected and why, step-by-step execution with results for each step, any variables extracted by Analyze case steps, which condition branch was followed, and the full reply Assembled AI would send.
Before activating, test all of the following:
- True positives — tickets that should be selected
- True negatives — tickets that should not be selected
- Edge cases — tickets on the boundary (for example, a refund request that also mentions "cancel")
- All branches — if your workflow has conditional logic, test a ticket for every path
How do I activate a workflow?
When you're ready to go live, select Manage activation to configure how Assembled AI handles matching interactions.
| Level | What it does |
|---|---|
| Inactive | No automation. The workflow won't run. |
| Agent review | Assembled AI executes the workflow but holds all action steps (replies, ticket updates, API calls). Agents see and approve or reject the draft in the Copilot side panel before anything is sent. |
| Automated | Assembled AI handles matching interactions without agent intervention. You can set an autosend percentage (1–100%) to ramp up gradually while maintaining a control group. |
The Deployments team recommends starting in Agent review mode before switching to full automation. A typical ramp looks like this:
- Agent review for one to two weeks — verify that selections and replies are accurate.
- Automated at 10–25% — monitor performance metrics.
- Scale to 50%, then 100% as confidence grows.
Note: Don't linger in Agent review indefinitely. Set a review timeline and a clear accuracy threshold for moving to Automated.
How do I monitor and iterate on a workflow after launch?
After activating a workflow, check regularly for the following:
- Selection accuracy — are the right tickets being selected? Are any being missed or over-captured?
- Reply quality — are responses accurate, on-brand, and resolving the issue?
- Escalation rate — is Assembled AI escalating too often (indicating gaps in knowledge or logic) or not enough (missing cases that need human review)?
Use the QA review tab to monitor and score conversations. Plan to review at least a sample of automated interactions each week.
When you notice misclassifications, identify the specific tickets being mis-selected or missed, review your selection criteria for gaps or ambiguity, add clearer inclusion and exclusion examples, and re-test in Preview before saving changes. For persistent issues, the AI-assisted fix-my-prompt approach in the selection criteria section above works well.
As you scale to multiple workflows, keep titles descriptive and consistent, use case filters as a first-pass filter to keep AI selection prompts focused, and aim for greater than 80% selection accuracy before scaling volume on any workflow.
What are some common workflow patterns?
| Use case | Key steps | Notes |
|---|---|---|
| FAQ or policy question | Trigger → Reply | The simplest pattern. Works well as a first workflow. |
| Sentiment-based routing | Trigger → Analyze case (sentiment) → Condition → Reply (path A or B) | Good for differentiating frustrated vs. satisfied customers. |
| Order lookup | Trigger → Custom API lookup → Condition → Reply | Requires API endpoint configuration first. |
| Clarifying question | Trigger → Collect information → Analyze case → Reply | Use when you need a follow-up before resolving. |
| Personalized reply | Trigger → Zendesk data lookup → Reply | Pull org or user data to tailor the response. |
| Exclude and escalate | Trigger → Condition (check for exclusion) → Escalate or Reply | Useful for undocumented features or sensitive topics. |
Comments
0 comments
Article is closed for comments.