← Back to the catalog
A1CUSTOMER COMMS

First-line support deflection

Reads incoming customer questions, answers the most common 60-70% of them automatically using grounded responses from the client's knowledge content, and routes anything uncertain to a human with the context already attached. The customer feels supported faster, the support team's queue stays focused on the genuinely hard questions, and answers stay accurate because the pattern grounds every response on real client documentation rather than its training data.

WHERE THIS FITS
BUSINESS SHAPES
Product companyDirect-to-consumerB2B services
VOLUME THRESHOLD
Below 200 inbound support requests per month a month, the payback rarely earns the build. Patterns this shape reliably pay back at 1,000+.
FITS BEST
E-commerce, SaaS, DTC with high ticket volume.
PAYBACK · 4-7 moBUILD · MediumVALUE · $40k-$180kWHEN · >200 tickets/mo
FAILURE MODE TO DESIGN AROUND
Out-of-date knowledge base → confidently wrong answers. Fix the KB first, or build maintenance into project scope.
REQUIREMENTS · 4 REQUIRED, 2 OPTIONAL

Requirements describe capabilities the pattern needs in your environment, not the vendors you must buy. Any system that fills a requirement satisfies it — that’s what makes the catalog portable across the long tail of SMB tooling.

  1. inbound_request_channel
    REQUIREDREADevent

    The channel where customers send their questions and the pattern listens for new ones.

    DATA SHAPE
    Free-text customer inquiries with sender identifier (email, account ID, or anonymous session) and timestamp. Optional metadata if available: topic tag, urgency level, prior thread.
    COMMONLY FILLED BY
    • ticket created in a helpdesk system
    • message arrives in a shared support inbox
    • chat session opened in a website widget
    • form submission from a contact page
  2. response_channel
    REQUIREDWRITErequest

    The channel through which replies reach the customer. Often but not always the same system as the inbound channel.

    DATA SHAPE
    Free-text reply tied to the original inquiry's thread or ticket identifier, plus the responder identity.
    COMMONLY FILLED BY
    • reply posted into the same helpdesk ticket
    • email sent from a shared inbox
    • message returned in the chat session
  3. knowledge_corpus
    REQUIREDREADcorpus

    The factual content the pattern grounds answers on. Without this, the pattern hallucinates and can't be trusted in production.

    DATA SHAPE
    Documents, FAQs, policies, product specifications, and procedural guides. Mix of structured (FAQ Q&A pairs) and unstructured (long-form articles, PDFs).
    COMMONLY FILLED BY
    • internal wiki or knowledge base
    • published help center articles
    • collection of policy documents in a file store
    • product documentation site
  4. human_escalation_route
    REQUIREDWRITEevent

    Where uncertain or low-confidence cases go so a human can take over without losing context.

    DATA SHAPE
    Escalation events including the original inquiry, the pattern's reasoning and confidence score, any draft response generated, and a suggested human owner or skill tag.
    COMMONLY FILLED BY
    • ticket reassigned within the helpdesk with internal notes
    • alert posted to a team chat channel with a deep link
    • email forwarded to a senior agent inbox
  5. customer_context_store
    RECOMMENDEDREADrequest

    Lookup of who is asking and their prior history with the business. Helps the pattern personalize and avoid contradicting prior decisions.

    DATA SHAPE
    Customer records keyed by identifier: name, account status, recent orders or interactions, support history, any active flags.
    IF MISSING
    Pattern still works but answers stay generic. Personalization and order-aware responses won't be possible. Deflection rate typically 10-15% lower without this role filled.
    COMMONLY FILLED BY
    • CRM with customer profiles and interaction history
    • e-commerce platform with order history
    • subscription billing system with account state
    • internal customer database
  6. supervisor_review_queue
    RECOMMENDEDWRITEbatch

    Periodic sample of pattern-handled cases that a human reviews to catch quality drift over time.

    DATA SHAPE
    Random or stratified sample of completed interactions with original inquiry, response sent, and metadata. Reviewer marks correct/incorrect/borderline.
    IF MISSING
    Quality drift goes unnoticed until customer complaints surface it. Strongly recommend filling this role even informally.
    COMMONLY FILLED BY
    • weekly digest emailed to a designated reviewer
    • dashboard with sampled interactions
    • ticket queue tagged 'AI-review-required' in the helpdesk
RUNTIME FLOW · 8 STEPS
  1. 01
    Listen for a new inquiry on the inbound channel and capture its content and sender identifier
    inbound_request_channel
  2. 02
    If sender identifier is recognized, fetch customer context to personalize the response
    customer_context_store
    DECISION Skip this step if customer_context_store is not filled or sender is anonymous.
  3. 03
    Classify the inquiry intent against the pattern's working taxonomy of 10-20 common topics tuned per client
  4. 04
    Retrieve relevant passages from the knowledge corpus using semantic search constrained by detected intent
    knowledge_corpus
  5. 05
    Generate a candidate response grounded on retrieved passages, with confidence score and citations
  6. 06
    Evaluate confidence and content against guardrails: minimum confidence threshold, no advice in regulated areas, no commitments on price or policy without source
    DECISION If confidence ≥ threshold AND guardrails pass, proceed to step 7. Otherwise jump to step 8.
  7. 07
    Post the response to the response channel as the configured responder identity, log the interaction for review sampling
    response_channelsupervisor_review_queue
  8. 08
    Route to a human agent with the original inquiry, draft (if any), reasoning, confidence score, and suggested owner
    human_escalation_route
EMISSIONS · 3

Structured outputs this pattern produces. Other patterns and client systems can subscribe to them, which is how the catalog composes over time.

  • labeled_inquiry

    Each customer inquiry classified by intent against the pattern's taxonomy.

    CONSUMED BY
    • analytics dashboards tracking topic trends
    • product teams looking for friction signals
    • knowledge base maintenance workflows that surface frequently-asked questions missing from docs
  • escalation_event

    Each escalation to a human, with the reason.

    CONSUMED BY
    • support team capacity planning
    • later patterns that may handle the escalated category
    • training feedback loops to expand the pattern's coverage
  • review_sample

    Sampled responses queued for human supervisor review, with the reviewer's eventual verdict.

    CONSUMED BY
    • quality dashboards
    • training data for prompt refinement and eval set growth
RELATED PATTERNS

Patterns that share the same category, business shapes, or capability needs as A1.

All CUSTOMER COMMS patterns →