Recommendations analyzes your conversations and surface actionable suggestions to improve deflection. It tells you exactly which knowledge articles to add or edit, and quantifies the projected impact of each recommendation.
To access Recommendations, navigate to Settings > Analytics > Insights.
Understanding Deflection Rate
Deflection Rate measures how effectively Atom resolves employee questions without human intervention.
A conversation is considered deflected when three conditions are met:
No request was created — the conversation did not result in a Service Request, Incident, or General Request.
Not smalltalk — the conversation was a genuine support interaction, not casual chatter or an ambiguous message.
AI evaluation confirms deflection — an AI evaluation reviews the conversation and marks it as successfully deflected, meaning Atom provided a meaningful answer that resolved the employee's question.
All three conditions must be true for a conversation to count as deflected.
Qualifying conversations are all conversations where an employee asked a genuine support question—the total pool minus smalltalk and ambiguous messages. This is the denominator.
Deflected conversations are the subset of qualifying conversations where no request was created AND the AI evaluation confirmed Atom successfully resolved the question. This is the numerator.
Deflection Rate = Deflected conversations ÷ Qualifying conversations
How deflection connects to recommendations
Recommendations specifically target subthemes where the deflection rate is below 80% and there are at least 10 interactions. These are areas with enough volume to matter and clear room for improvement—the system has identified that adding or improving knowledge would close the gap.
Understanding Hours Saved
The Hours Saved metric quantifies the value Atom delivers by deflecting conversations that would otherwise require manual handling by your team.
How Hours Saved is calculated
The calculation works per subtheme, then adds up across all subthemes:
Get MTTR per subtheme: Every subtheme has a Mean Time To Resolve (MTTR)—the average time it takes your team to manually resolve a request in that category. If there isn't enough data for a subtheme, a default MTTR of 6 hours is used.
Count deflected conversations per subtheme: The system counts how many conversations in each subtheme were deflected by Atom (using the deflection criteria described above).
Multiply: For each subtheme, Hours Saved = MTTR × Number of deflected conversations.
Sum across subthemes: Add up the hours saved across all subthemes to get your total Hours Saved.
For example, your "Outlook Signature" subtheme has an MTTR of 1 hour and 27 deflected conversations last month. Your "VPN Troubleshooting" subtheme has an MTTR of 3 hours and 12 deflected conversations. That's (1 × 27) + (3 × 12) = 63 hours saved across just two topics.
How Recommendations work
Recommendations are generated automatically by analyzing patterns across your Atom conversations and request data.
Conversation analysis: The system reviews all conversations handled by Atom and groups them by subtheme—the topic clusters already generated by Insights (e.g., "VPN Troubleshooting", "Password Reset", "Okta MFA Issues").
Deflection assessment: For each subtheme, the system identifies conversations that could have been deflected with knowledge but weren't. This assessment uses AI-based grading to determine whether a knowledge article would have resolved the employee's question without creating a request. Only subthemes with a deflection rate below 80% and at least 10 interactions are considered—this ensures recommendations focus on areas with meaningful volume and clear room for improvement.
Learning from requests: For conversations that resulted in requests, the system learns from those requests—pulling resolution context to understand what information would have helped the employee self-serve. This directly informs the content of each recommendation.
Refresh cadence: New recommendations are generated every 15 days. The system avoids creating duplicate recommendations across analysis cycles, so you won't see the same suggestion repeated. Recommendations are sorted by highest projected hours saved.
Types of recommendations
Every recommendation falls into one of two action types:
Add a new article
An Add recommendation appears when the system identifies a high-volume subtheme where no relevant knowledge article exists. Employees are asking about this topic, but Atom doesn't have the content to answer—so conversations are escalating to requests unnecessarily.
When you open an Add recommendation, you'll see:
Pre-generated content: A draft article based on resolution notes from past requests in that subtheme. This gives you a starting point you can review and publish.
Rationale: Why this recommendation was created—typically citing the volume of conversations in this subtheme and the absence of matching knowledge.
Sources: The specific conversations and requests that informed the recommendation, so you can verify the suggested content.
Edit an existing article
An Edit recommendation appears when a knowledge article exists for a subtheme, but it isn't fully resolving employee questions. This could mean the article is incomplete, outdated, or unclear—employees are reading it but still creating requests afterward.
When you open an Edit recommendation, you'll see:
Suggested changes: What to update in the existing article, based on patterns in the conversations where the article fell short.
Rationale: Why the edit is recommended—often citing conversations where the article was surfaced but didn't prevent a request.
Sources: The conversations and requests that show where the current article is falling short.