Detection rule types reference

A detection rule tells Sidekick what structured data to look for in conversation messages. When a match is found, the value gets highlighted inline and (above the configured confidence threshold) auto-saved to a custom field.

Three methods to define what counts as a match. Pick the one that fits your data shape.

Pattern builder

A no-code editor for structured identifiers. You define:

• Prefix — what comes before the value (e.g. ORD-, SN:, Account ID)
• Suffix — what comes after (often empty)
• Allowed characters — digits, letters, alphanumeric, custom set
• Min length / max length
• Priority — when multiple rules could match, higher-priority rules win

• Order numbers with consistent prefix — ORD- + 6+ digits → matches ORD-123456
• Account IDs — ACC- + 8 alphanumeric → matches ACC-A1B2C3D4
• Internal SKUs — digits-only, length 8-12

The pattern builder produces deterministic matches at zero AI cost. Use this whenever your data has a structured shape.

Pre-built templates

Atender ships with ready-made templates for common data types:

• Serial Number — Alphanumeric product identifiers
• Email — Email addresses mentioned in messages
• Phone — Phone numbers in various formats
• Order ID — Order or transaction identifiers (configurable)
• Product Code — SKU or product reference codes

Templates are starting points — duplicate one and tweak it for your specific format.

AI detection

Describe what to look for in plain language and let the LLM extract it. Use when the data is fuzzy enough that regex doesn’t fit.

• Affected product when customers describe it variously — The product the customer is asking about. Could be referred to by name, SKU, or descriptively (e.g. "the blue one")
• Sentiment or intent — The customer's primary intent: complaint, inquiry, refund request, cancellation
• Structured data in unstructured language — The deadline date the customer is asking about

AI detection is more flexible but has costs:

• LLM call per detection — adds latency and per-conversation AI spend
• Lower confidence by default — the AI is less certain, so values often fall below the auto-save threshold and get flagged for agent review instead
• Worse at exact identifiers — AI is great at “what product” and bad at “extract this exact 12-character order number”

For exact identifiers, use Pattern builder or RegExp. For fuzzy concepts, use AI detection.

RegExp

For when you want raw regex. Atender accepts a standard regex pattern and runs it against message bodies.

• Data that doesn’t fit the pattern builder’s shape — Mixed prefix/suffix patterns, alternation
• Existing regex you already maintain elsewhere — \bSO\d{6,}\b for sales orders
• Migrating rules from another system — Direct port

RegExp gives you full control, but:

• Test thoroughly — bad regex can cause false positives that pollute custom fields
• Use word boundaries (\b) to avoid matching substrings
• Atender provides a built-in test tool — use it before saving (see Create a detection rule)

Picking which method

• Is the data structured (consistent prefix/length/charset)? — Pattern builder
• Is the data fuzzy or context-dependent? — AI detection
• Do you already have a working regex? — RegExp
• Is the cost of false positives high (it gets auto-saved to a critical field)? — Pattern builder or RegExp (deterministic)
• Is the data described differently each time customers mention it? — AI detection

Confidence threshold

Each rule’s match gets a confidence score. Sidekick’s global confidence threshold (Settings → Sidekick → General) decides:

• Above threshold → auto-save to the mapped custom field, highlight green
• Below threshold → flag for agent review, highlight amber, don’t auto-save

Pattern builder and RegExp matches typically score very high (deterministic). AI detection scores vary — tune the threshold based on what’s noise vs. signal in your data.

See also

• Create a detection rule
• Auto-populate a custom field with Sidekick
• Custom Fields