The cost of missing urgent feedback
The majority of customer feedback can wait. A suggestion about a minor UI preference, a positive note about a feature, a low-priority question — none of these need immediate attention. But some feedback represents an actively unhappy customer, a critical failure, or an imminent churn risk. That subset has a short window. A customer signaling that they are about to leave may be reachable today and gone next week.
Manual review at volume does not catch these reliably. By the time a weekly feedback review happens, the urgent items are buried under everything else that came in that week. Automatic urgency detection is an attempt to surface the subset that should not wait — so it can be routed to someone who can act on it quickly.
The signals Rereflect looks for
Urgency is not a single thing. Rereflect's urgency detection looks for a combination of signals, each of which increases the probability that an item is high-priority:
- Strong negative sentiment — a compound VADER score in the very negative range is correlated with urgent feedback, though not deterministic. Most very negative feedback is urgent; not all urgent feedback is maximally negative.
- Churn language — explicit phrases that signal intent to leave: "canceling," "switching to," "looking at alternatives," "not worth it anymore," "going to ask for a refund." These are strong signals when present.
- Escalation language — words and phrases that indicate the customer feels the issue is unresolved and escalating: "unacceptable," "this has been going on for weeks," "already contacted support," "need to speak to someone."
- Critical failure vocabulary — terms indicating that something is broken in a way that blocks the customer's work: "can't access," "lost data," "completely broken," "production is down."
- SLA or compliance references — mentions of contractual obligations, SLA terms, or compliance requirements often indicate that the impact of a failure is not just inconvenience.
When an LLM is configured, Rereflect uses it to reason about urgency more holistically — considering the combination of signals and the overall context of the feedback, rather than checking for individual keywords. The LLM can identify urgency in items that do not use the exact phrases on a keyword list but clearly describe a critical situation.
How the detection pipeline works
Urgency detection runs as part of the standard analysis pipeline on each new feedback item. The process:
- Sentiment pre-filter — items with strong negative sentiment are weighted more heavily as urgency candidates. Items with neutral or positive sentiment can still be flagged, but face a higher bar.
- Keyword scan — the item is scanned for urgency-related vocabulary across the signal categories described above. Matches are weighted by signal type, with churn language and critical failure vocabulary scoring highest.
- LLM reasoning (if configured) — the LLM receives the feedback text and a description of urgency criteria, and returns a binary flag plus a brief reason. This catches cases that keyword matching misses.
- Final flag — items meeting the threshold are marked as urgent in the database and surfaced in the urgent feedback dashboard.
The threshold for flagging is intentionally calibrated toward sensitivity (catching more) over specificity (fewer false positives). The cost of missing a genuinely urgent item is higher than the cost of reviewing an item that turns out to be fine. The trade-off means your urgent feedback queue will contain items that turn out not to need immediate action — but it will catch most of the ones that do.
Where urgency detection fails
Honest enumeration of failure modes:
- Polite frustration — some customers describe serious problems in calm, measured language. An item that says "I am becoming concerned about the reliability of this feature as it has failed three times this week" is urgent, but contains no keywords that trigger detection and reads as mildly negative in sentiment.
- Domain-specific criticality — a phrase that signals urgency in your product may be neutral in a general model. If "retry limit exceeded" means something is broken in your product, that needs to be in your urgency category description.
- Non-English items without LLM — keyword lists are English-centric. Non-English urgent feedback will be missed by the keyword layer if the LLM is not configured.
- Delayed urgency — some items describe slow-building problems ("this has been getting worse over the past few months") that are urgent in the sense that they represent a churning customer, but do not contain acute language. These are harder to catch without understanding the customer's history.
Urgency detection is a filter that makes the urgent item review process faster — it does not replace it. The goal is to reduce the fraction of feedback you need to review immediately to a manageable subset, while catching most of the genuinely urgent items.
Tuning urgency for your product
The default urgency signals are reasonable starting points, but "urgent" is product-specific. A data loss event is always urgent. Whether a feature request marked as blocking is urgent depends on who the customer is and your support policies.
Rereflect's custom urgency configuration lets you describe what urgent means for your business. The description feeds into the LLM classification prompt, which means you can include product-specific signals ("any mention of data export failure is urgent") and exclusions ("billing questions are not urgent unless the customer mentions cancellation").
Review your false negatives — items that should have been flagged but were not — periodically and use them to refine the urgency description. A few targeted additions to the description usually cover the systematic gaps.