Optimize
Every conversation teaches you something.
Optimize reviews your recent conversations, finds recurring patterns, and suggests practical improvements for your agents. Every recommendation is backed by real Guest interactions, and nothing changes until you approve it.
Summary
On this page you'll learn how to:
- Understand the different types of optimization suggestions.
- Run a new analysis and review recommendations.
- Improve prompts and your Knowledge Base using real conversations.
- Identify recurring business trends.
- Apply changes safely and only when you're ready.
- Read the Technical details if you need the pipeline, eligibility rules, and apply flow.
What you can do
Optimize helps you continuously improve your agents using real conversations.
You can:
- Discover better ways to improve prompts.
- Spot missing knowledge before it becomes a recurring issue.
- Identify trends across Guest conversations.
- Review the evidence behind every recommendation.
- Decide which changes to apply and which to ignore.
Why it matters for your business
Your agents don't get better by chance.
Every conversation reveals opportunities to improve how your business communicates with Guests. Instead of manually reviewing hundreds of conversations, Optimize highlights what matters most.
With every improvement you can:
- Save money: reduce repeated Guest questions and follow-up work.
- Work faster: spend less time reviewing conversations by hand.
- Standardize service: keep responses consistent across every interaction.
- Delight every Guest: close knowledge gaps before they frustrate Guests.
Who can use Optimize
Optimize is available only when it has been enabled for your workspace. If you don't see it in the navigation, contact your Voicetta representative.
Permissions
Viewers
- View suggestions and analysis status.
Members, Admins, and Owners
- Start a new analysis.
- Apply or ignore suggestions.
- Acknowledge trends.
If there are pending suggestions, your Home dashboard may also display an Optimization opportunities card.
Suggestion types
Optimize groups recommendations into three categories.
Prompt improvements
Suggestions to improve your agent's prompt or greeting.
Every suggestion includes:
- what should change
- why it matters
- conversations that support the recommendation
When you apply a suggestion, Voicetta updates the agent configuration and synchronizes it automatically.
Knowledge gaps
When Guests ask questions your agent can't answer, Optimize detects the missing information and prepares a Knowledge Base draft.
Applying the suggestion creates a draft for review. Nothing is published until someone approves it in the Knowledge Base.
Trends
Some conversations reveal patterns rather than problems.
Optimize highlights recurring topics, Guest behavior, and emerging trends across recent conversations to help you understand what is changing.
Trends don't modify your agents. Simply acknowledge them or ignore them.
How it works
- Open Optimize.
- Review any pending suggestions.
- Select Analyze now to start a new analysis whenever you want.
- Open a suggestion to review:
- the recommendation
- why it was created
- the supporting conversations
- Apply the change or ignore it.
- For trends, acknowledge the insight or dismiss it.
You can also open any supporting conversation in History to see the full Guest context behind a recommendation.
What happens after you apply
Prompt improvements
The updated prompt is saved and synchronized with your voice provider.
If someone has already modified the prompt since the suggestion was created, Voicetta blocks the update to prevent overwriting newer changes.
Knowledge gaps
Applying a suggestion creates a new Knowledge Base draft.
The draft still follows the normal review and approval process before it becomes available to your agents.
Trends
Acknowledging or ignoring a trend simply marks it as reviewed.
No prompts or Knowledge Base content are changed.
Requirements
Optimize works best when it has enough real conversations to learn from.
In most workspaces you'll need around 20 completed conversations before analysis produces meaningful recommendations.
Only completed conversations with a usable summary and intent are included. Test conversations, incomplete conversations, and recovery traffic are automatically excluded.
Manual analysis can be started approximately once every 60 minutes.
Privacy
Optimize analyzes existing conversation summaries, detected intent, current prompts, greeting messages, and Knowledge Base titles.
It does not analyze raw conversation transcripts.
Technical details
For teams that want to understand how Optimize works under the hood.
Pipeline overview
flowchart LR Convos[Eligible conversations] Trigger[Manual or scheduled analyze] Corpus[Build corpus from summaries and intent] Model[Optimization model] Suggest[Pending suggestions] Human[Human review] Apply[Apply or ignore] Convos --> Trigger Trigger --> Corpus Corpus --> Model Model --> Suggest Suggest --> Human Human --> Apply
What data is analyzed
Optimize analyzes:
- Conversation summaries
- Detected intent
- Current agent prompts and greeting messages
- Knowledge Base titles (and a limited coverage check against approved Knowledge content when verifying gaps)
- Conversation metadata needed to scope and link evidence (for example conversation id, time window, assistant)
Each corpus record is a short, bounded text block such as summary + intent — not the full call transcript.
Raw conversation transcripts and encrypted conversation payloads are not analyzed.
How suggestions are generated
- Select an eligible conversation corpus for the workspace (and per assistant where needed).
- Build a structured analysis payload: current configuration, Knowledge titles, and corpus records.
- Run the optimization model with a fixed suggestion budget and category allowlist.
- Store suggestions as pending, each linked to one or more supporting conversations as evidence.
Suggestion categories:
| Category | What it targets | Apply behavior |
|---|---|---|
| Prompt improvement | Prompt or greeting (inbound / outbound) | Writes agent config, then syncs to the voice provider |
| Knowledge gap | Missing Knowledge content | Stages a Knowledge Base draft only |
| Trend | Recurring topics / behavior | Informational — acknowledge or ignore |
flowchart TB pending[pending] applying[applying] applied[applied] failed[apply_failed] ignored[ignored] pending -->|Apply prompt| applying applying -->|Provider sync ok| applied applying -->|Sync error| failed pending -->|Apply knowledge gap| applied pending -->|Ignore| ignored
Safety
Optimize never changes your workspace automatically.
Every suggestion requires explicit approval before any change is made.
Additional safeguards include:
- Prompt version / hash validation so Apply cannot overwrite newer edits made after the suggestion was created
- Knowledge Base draft workflow before publication (same review path as other staged knowledge)
- Idempotent Apply for prompt changes (safe retries)
- Fingerprint suppression so ignored or recently applied ideas are not re-suggested immediately
- Read-only access for Viewer roles
- Platform and workspace feature flags so Optimize can be enabled per environment
Analysis requirements
Only eligible conversations are included in an analysis.
A conversation is typically eligible when it:
- has ended (completed)
- is not a test conversation
- has a usable summary and/or detected intent
- is not recovery / retry follow-up traffic
- belongs to an inbound or outbound agent in the workspace
Corpus sizing (typical production defaults):
- About 20 eligible conversations in the business before analysis is meaningful
- About 10 eligible conversations per agent before that agent is included
- Recent window of about 7 days for prompt and knowledge analysis
- Trends compare the current period against a prior baseline window
- Caps limit how many conversations and suggestions are processed per run
Analysis schedule
flowchart TD Manual[Analyze now] Cool[60 minute cooldown] Sched[Scheduled analyze] Day[At most once per 24 hours] Queue[Queued or running run] Worker[Background worker] Done[Completed suggestions] Manual --> Cool Manual --> Queue Sched --> Day Day --> Queue Queue --> Worker Worker --> Done
- Manual: start from Analyze now. Rate-limited to about one manual run every 60 minutes. If a run is already queued or running, the request reuses that run instead of starting a duplicate.
- Scheduled: when scheduling is enabled for the environment, Optimize can enqueue a workspace run about once per day when enough eligible conversations exist.
- Analysis runs asynchronously in the background; the UI polls run status until suggestions appear.
Applying changes
Prompt improvements
Applying a prompt suggestion updates the selected agent configuration and synchronizes the new version with the connected voice provider. The suggestion moves through applying until sync finishes (applied or apply_failed).
Knowledge gaps
Applying a Knowledge suggestion creates a Knowledge Base draft. The draft follows the standard approval workflow before becoming available to agents.
Trends
Trend recommendations are informational only. Acknowledging or ignoring a trend never modifies prompts or Knowledge Base content.
Designed for production
Optimize is designed to improve agents continuously without interrupting live traffic.
It follows the same approval and deployment workflows as the rest of Voicetta — human review, staged knowledge, and provider sync — so your team keeps full visibility and control over every change.