Home / architect / case-deflection-strategy
case-deflection-strategy
Use this skill when designing or assessing a Salesforce case deflection program — identifying high-frequency contact reasons, selecting deflection channels, and defining KPIs such as deflection rate and goal completion rate. Trigger keywords: case deflection, self-service, Einstein Conversation Mining, chatbot containment, knowledge-led deflection, contact reason analysis. NOT for the self-service portal UX, search experience and case-form design — use admin/self-service-design. NOT for end-to-end Service Cloud channel and routing architecture — use architect/service-cloud-architecture.
Open the skill folder on GitHub
Case Deflection Strategy
Use this skill when a practitioner needs to design, assess, or improve a Salesforce case deflection program. It activates when the goal is to reduce live contact volume by routing customers to self-service channels — primarily through Einstein Bots, knowledge articles, and Experience Cloud self-service — and when decisions about which topics to deflect, which channels to deploy, and how to measure success need architectural guidance.
Before Starting
Gather this context before working on anything in this domain:
- What is the org's current case volume by contact reason? Pull from Case reports or Einstein Conversation Mining. Without topic-level volume data, deflection prioritization is guesswork.
- Does the org have Salesforce Knowledge enabled and populated? Deflection without a serviceable knowledge base produces bot conversations that end in agent escalation — defeating the purpose.
- What is the org's self-service channel inventory? Experience Cloud portal, Messaging for In-App and Web, Einstein Bots, and Email-to-Case all have different deflection profiles.
- Are data categories set up correctly on Knowledge articles? Miscategorized or uncategorized articles surface in wrong contexts and produce low article-solved rates.
Questions to Ask Before Configuring
Each question traces to a gotcha in references/gotchas.md.
| Question | Why it matters | What a good answer adds | What proper configuration adds over just doing it |
|---|---|---|---|
| "What share of last quarter's cases came from each Case Origin, and how much chat or messaging transcript history exists?" | Conversation Mining covers conversational channels only (Gotcha 2) | A topic list weighted by the channels that actually carry volume | Wave 1 targets the real top reasons, not the chat-visible ones |
| "Which numerator and denominator define deflection, and which event proves a customer's goal was met?" | Containment, case reduction, and goal completion diverge (Gotchas 1, 6, 8) | One written formula per channel, reported per topic | Bot goal steps and article votes record resolution instead of inferring it |
| "Which audiences must see each candidate article: guest, authenticated customer, partner, high-volume portal users?" | Visibility needs a category in every group on the article, and high-volume users have no role (Gotchas 3, 7) | A visibility matrix of category group by audience | Articles reach the audience the topic was chosen for |
| "Which channels is each candidate article published to?" | Channel flags hide categorized articles (Gotcha 4) | Channel flags checked in the readiness audit | No "published but invisible" articles at launch |
| "Where will the bot run, and will we test knowledge answers there as a guest?" | Results depend on the running context (Gotcha 5) | A test script per audience through the deployed channel | A designed KnowledgeFallback path instead of a dead end |
What a proper deflection design adds over "just launching a bot": the topic list matches real volume, the metric proves resolution rather than silence, and every target article is reachable by the audience it was written for.
Core Concepts
Einstein Conversation Mining for Topic Identification
Einstein Conversation Mining (ECM) analyzes historical chat and messaging transcripts to surface the most frequently discussed contact reasons. It clusters similar intents together and ranks them by volume, making it the primary tool for identifying deflection candidates. The output is a report showing topic clusters, volume share, and a recommended automation potential score. Topics with high volume and low complexity (password resets, order status, return requests) are the best first-wave deflection targets. ECM requires existing transcript history and a Messaging or Chat channel (UNVERIFIED (2026-10-03): the often-quoted 90-day minimum is Help-only). Orgs without transcript data must use case subject/description classification or manual categorization as a proxy.
Deflection Rate and Goal Completion Rate
Deflection rate measures the percentage of contacts that were resolved in a self-service channel without transferring to a live agent. It is the headline KPI for a deflection program. An "average 27% deflection rate" is often attributed to Salesforce (UNVERIFIED (2026-10-03): no fetched Salesforce source states it). Goal completion rate (GCR) is the more meaningful operational metric: it measures the percentage of self-service sessions where the customer achieved their stated goal (e.g., found an article, completed a transaction, got an answer from the bot) regardless of whether they subsequently escalated. A bot with 60% GCR and 30% deflection rate is performing well: the escalations are genuinely complex. A bot with 5% GCR and 30% deflection rate is suppressing contacts without resolving needs, which degrades CSAT. Always track both. A "198% average ROI" figure also circulates (UNVERIFIED (2026-10-03): no fetched Salesforce source states it); build the business case from the org's own avoided handle time instead.
Knowledge Article Quality as a Prerequisite
Deflection channels that surface knowledge articles — bots, self-service portals, search — are only as good as the articles behind them. Common knowledge quality failures that kill deflection programs: articles are written at an agent reading level not a customer reading level; articles are too long and not scannable; articles answer edge cases instead of the top 80% of contact reasons; data category assignments are inconsistent so articles appear in wrong channel contexts. Before launching a deflection channel, audit article coverage against the ECM topic list and grade each article for customer readability. The knowledge readiness assessment is a required input to the deflection program design, not a nice-to-have.
Data Category Hygiene
Data categories and article channel flags together decide which articles each audience sees. Three documented rules drive most "we have articles but customers can't find them" complaints:
| Rule | Consequence for deflection |
|---|---|
| Making a category visible exposes its whole family line, ancestors and descendants | Re-tagging articles at every level is unnecessary |
| An article is visible only if the user sees at least one category in each of its category groups | Adding a second category group without guest visibility hides every article tagged in it |
| Once visibility is configured, a user with none in a group sees only articles not classified in that group | A missing grant fails silently as "no results" |
Article channel flags (IsVisibleInPkb, IsVisibleInCsp, IsVisibleInPrm) are a separate gate on top of categories. Validate both end to end as each audience before launch; details and sources are in references/gotchas.md Gotchas 3, 4, and 7.
Common Patterns
Pattern: Einstein Bot as First-Door Triage
When to use: Org has Messaging for In-App and Web or Live Agent Chat, has a knowledge base with 50+ articles, and can identify 5–10 high-frequency contact reasons from ECM or manual analysis.
How it works: Deploy an Einstein Bot as the entry point for all inbound messaging. The bot handles a menu-driven or NLU-driven triage that routes customers to: (1) article self-service for informational topics, (2) transactional bot dialogs for automatable requests (order status via API, password reset via Flow), or (3) Omni-Channel queue transfer for genuinely complex issues. Configure containment — the bot resolves the intent without human transfer — as the primary success metric. Use session end surveys to measure goal completion rate.
Why not the alternative: Routing all contacts directly to agents and using agents to recommend articles (reactive deflection) is consistently less effective than front-door triage. It increases agent handle time even for conversations that could have been self-served, and it does not reduce contact volume.
Pattern: Knowledge-Led Self-Service Portal with Federated Search
When to use: Org has Experience Cloud and wants to reduce email and phone contact volume for informational queries (how-to questions, policy lookups, product documentation).
How it works: Build an authenticated or guest Experience Cloud site with Salesforce Knowledge as the content layer. Enable federated search so customers can find articles, community answers, and Trailhead content from a single search bar. Gate case submission behind a search-first flow: before the Web-to-Case form renders, the customer sees top-matched articles for their typed subject. This is the Salesforce-recommended "article deflection before case creation" pattern. Track article-solved rate (sessions ending in article view with no case submitted) as the deflection metric.
Why not the alternative: Embedding a case form on the landing page without a search-first gate guarantees that high-frequency informational queries generate cases, regardless of how good the knowledge base is.
Decision Guidance
| Situation | Recommended Approach | Reason |
|---|---|---|
| High chat/message volume, ECM data available | Einstein Bot with NLU + ECM topic coverage | NLU handles natural language variance; ECM gives validated topic list |
| Low transcript history, but case data available | Manual topic clustering from case subjects + Bot menu dialog | Menu-driven bots do not require NLU training and are reliable for known topic lists |
| Primarily email and web form contact volume | Search-first portal with article deflection gate before case form | Reduces case creation for informational queries without requiring a bot |
| Knowledge base is sparse or poorly categorized | Knowledge readiness sprint before any deflection channel launch | Launching deflection without serviceable articles produces escalation-only sessions |
| Org has <50 articles but wants bot | Bot with transactional automation only (order status, case status lookup) | Transactional bots do not depend on knowledge quality |
| Deflection rate plateau after initial launch | ECM re-analysis to find next topic cluster; article quality audit | Initial deployment typically covers 3–5 topics; ECM identifies the next wave |
Recommended Workflow
Step-by-step instructions for an AI agent or practitioner working on this task:
- Audit current contact volume by topic. Run Einstein Conversation Mining on the available transcript history, and pull Case volume grouped by Origin and Reason. Rank contact reasons by volume. Identify the top 10 topics and tag each as: informational, transactional, or complex. Deflection candidates are informational and transactional.
- Assess knowledge readiness. For each deflection candidate topic, check whether a customer-readable knowledge article exists, is assigned to the correct data categories, and is visible on the target self-service channel. Flag gaps. A topic with no article cannot be deflected via knowledge-led channels.
- Select deflection channel per topic. Use the Decision Guidance table above. Informational topics go to search-first portal or bot + article. Transactional topics go to bot + Flow automation. Complex topics get escalation paths, not deflection targets.
- Define KPI baseline. Before launch, record baseline with the queries in
references/examples.md: inbound volume by Case Origin, case creation rate, average handle time. Derive the target deflection rate from the volume share of wave-1 topics, not from a published average. Set a goal completion target and a containment target for bot deployments. - Build knowledge gaps. Author missing articles for deflection candidates. Write at a 6th–8th grade reading level. Use numbered steps for how-to content. Keep articles under 500 words. Assign data categories aligned to the target channel's visibility rules.
- Launch deflection channel and instrument. Deploy bot dialogs or portal search-first flow. Tag sessions with the deflection topic so you can report deflection rate per topic, not just overall. Set up a session end survey for goal completion rate.
- Run monthly ECM refresh. Deflection programs stagnate when the topic list is treated as fixed. Re-run ECM every 30–60 days to find emerging topics and retire topics where contact volume has dropped.
Review Checklist
Run through these before marking work in this area complete:
- [ ] ECM report or equivalent topic analysis completed with at least 10 topics ranked by volume
- [ ] Each deflection candidate topic mapped to informational / transactional / complex
- [ ] Knowledge article coverage assessed against deflection candidate list; gaps documented
- [ ] Data category assignments validated for target channel and user profile (guest and/or authenticated)
- [ ] KPI framework defined: deflection rate baseline, GCR target, containment rate target (bot deployments)
- [ ] Bot or portal deflection channel configured with per-topic session tagging for reporting
- [ ] Post-launch review scheduled within 30 days with ECM re-run planned at 60 days
Salesforce-Specific Gotchas
Full detail and sources in references/gotchas.md. The short list:
- An abandoned bot session looks like containment; measure goal completion with bot goals.
- Einstein Conversation Mining covers conversational channels, so email-heavy orgs need case text analysis.
- Category visibility is broad (whole family line); the trap is the at-least-one-category-per-group rule.
- Article channel flags hide correctly categorized articles.
- High-volume portal users have no role, so grant category visibility by profile or permission set.
Output Artifacts
| Artifact | Description |
|---|---|
| Deflection candidate topic list | Ranked list of contact reasons with volume, complexity tag, and recommended channel |
| Knowledge readiness gap report | Table mapping deflection topics to article status: exists / needs update / missing |
| KPI framework document | Baseline metrics and targets for deflection rate, GCR, containment rate |
| Deflection program roadmap | Phased rollout plan: wave 1 (top 3–5 topics), wave 2 (next 5), quarterly ECM refresh cadence |
| Data category audit output | List of data category assignment issues affecting article visibility on target channel |
Related Skills
- architect/service-cloud-architecture — use for end-to-end Service Cloud channel and routing decisions that frame the deflection program within a broader service strategy
- architect/multi-channel-service-architecture — use when deflection spans multiple channels (web, mobile, messaging) and channel-specific configuration decisions are needed
- admin/einstein-activity-capture-setup — not directly related but often confused; ECM is a separate product from Einstein Activity Capture