# Assist suggestions
URL: /response-workflows/assist-suggestions
Type: concept
Description: How Stylo pre-generates responses and surfaces them as draft suggestions in the Zendesk sidebar.
Keywords: assist suggestions, copilot suggestions, sidebar, drafts
Assist suggestions are draft responses that appear in the Stylo sidebar when an agent opens a ticket. Instead of starting from scratch, agents see ready-to-use responses that have already been generated based on the ticket context, customer data, and your workflow's instructions.

Suggestions can come from two sources:

* **Response Workflows** — AI-powered response templates that are automatically matched to tickets based on their "when to use" description
* **Workflows** — multi-step automations (data lookups, API calls, conditional logic) that produce a draft at the end of their execution

Both types surface in the same place in the sidebar, and agents interact with them the same way.

How suggestions work [#how-suggestions-work]

From Response Workflows [#from-response-workflows]

1. **Customer sends a message** — a new message arrives on a ticket
2. **Stylo runs in the background** — all response workflows set to "cache" mode are evaluated against the ticket
3. **Responses are pre-generated** — workflows that match the ticket generate their responses and store them as drafts
4. **Agent opens the ticket** — the Stylo sidebar shows the best matching suggestions, ranked by confidence
5. **Agent reviews and sends** — the response is pasted into the Zendesk composer for review and sending

Suggestions are generated proactively — by the time the agent opens the ticket, the response is already waiting.

From Workflows [#from-workflows]

1. **A trigger fires** — a webhook, schedule, or manual action starts the workflow
2. **The workflow executes** — steps run in sequence (fetch order data, check eligibility, call an API, generate a response with AI)
3. **The workflow pauses for review** — when the workflow reaches a human review step set to surface a draft, it pauses and shows the draft in the sidebar
4. **Agent reviews and sends** — the agent reviews the draft, optionally edits it, and approves. The workflow then resumes and completes any remaining steps (update ticket status, add tags, log to CRM).

Workflow-sourced drafts carry structured context data from the steps that ran before the review — agents can see what data was gathered and what decisions the workflow made.

What agents see [#what-agents-see]

Each suggestion shows:

* **Source name** — which response workflow or workflow generated this response
* **Confidence** — how confident Stylo is that this suggestion matches the ticket (for response workflow suggestions)
* **Generated response** — the full response text, ready to review
* **Context data** — structured data from the workflow's execution (for workflow-sourced suggestions)

Agents can:

* **Use the suggestion as-is** — paste it into the composer and send
* **Edit before sending** — modify the response in the Zendesk composer
* **Dismiss** — skip the suggestion and write their own response

Setting up suggestions [#setting-up-suggestions]

For Response Workflows [#for-response-workflows]

1. Set the workflow's automation mode to **Cache**
2. Choose a trigger type (typically **Customer reply**)
3. Write a clear "when to use" description so Stylo matches it to the right tickets
4. Publish the workflow

The workflow will now run in the background whenever a matching ticket event occurs, and the generated response will appear as a suggestion in the sidebar.

For Workflows [#for-workflows]

1. Build your workflow with the steps needed to gather data and generate a response
2. Add a **Human Review** step and turn on its draft-to-sidebar option {/* TODO(docs): the Workflow Builder still labels this toggle with the retired agent-assist term; name it here once that label is updated. */}
3. Configure the draft text field to reference the upstream node that contains the response
4. Set the visibility mode to **Cache**
5. Publish the workflow

When the workflow's trigger fires, it will execute up to the human review step, then surface the draft in the sidebar for agent review.

Multiple suggestions [#multiple-suggestions]

If more than one source matches a ticket, agents see multiple suggestions ranked by confidence. For example, an order-related ticket might have a pre-generated response from a "Refund Policy" response workflow and a data-enriched draft from a "Order Investigation" workflow — the agent sees both and picks the most relevant one.

On-demand generation [#on-demand-generation]

Some suggestions appear with a "generate" option instead of a pre-rendered response. This happens when Stylo identified a matching response workflow but didn't pre-generate the response (to save resources). The agent clicks to generate the response on demand, using the current ticket state.

Suggestions on messaging channels [#suggestions-on-messaging-channels]

For messaging channels (WhatsApp, Web Widget, chat), suggestions work the same way from the agent's perspective — they appear in the Stylo sidebar in Zendesk Agent Workspace. The difference is in how the response is delivered once the agent approves it.

Progressive automation [#progressive-automation]

Assist suggestions are the starting point of Stylo's progressive automation model:

1. **Start with suggestions** — agents review every response before sending. This builds confidence in Stylo's quality and surfaces edge cases.
2. **Monitor acceptance rates** — track how often agents use suggestions as-is, edit them, or dismiss them. High acceptance rates indicate the workflow is ready for more automation.
3. **Graduate to automatic sending** — once you're confident in a workflow's quality, switch it from cache mode to internal note or public reply mode. The AI sends responses automatically, and agents only handle escalations.

This progression lets you build trust gradually — every suggestion that an agent reviews is a data point that informs whether full automation is appropriate.

Tips [#tips]

* **Cache mode is risk-free.** Cached suggestions are never sent automatically — an agent always reviews first. This makes it a great starting point for any workflow.
* **Use "when to use" descriptions to reduce noise.** If agents see too many irrelevant suggestions, tighten the response workflow's "when to use" description so it only matches the right tickets.
* **Track usage.** Monitor how often agents use, edit, or dismiss suggestions. High dismissal rates indicate the workflow needs improvement — either the matching is off or the generated responses aren't helpful.
* **Combine sources.** Response workflows are great for straightforward responses. Workflows are better when the response requires multi-step data gathering or conditional logic. Use both together for comprehensive coverage.

Related [#related]

* [Assist](/assist)
* [Background Automation](/response-workflows/background-automation)
* [Response Workflows](/response-workflows)