Every ticket the AI couldn't answer is telling you something

The number nobody reports

Every AI support vendor shows you what percentage of tickets got deflected. Almost none show you a useful view of the rest — the escalations arrive as a list of failures, or as a single "could not handle" bucket.

That is backwards. A high deflection rate tells you the past went well. The escalations tell you what to do next, and they are the only list of documentation work you will ever get that is ranked by real customer demand rather than by whoever shouted loudest in the last team meeting.

Missing and unclear are different problems

This is the distinction that makes the list actionable, and it falls straight out of how the agent works.

Collapsing those two into "the AI couldn't handle it" is common and it destroys the signal. One sends you to a blank page; the other sends you to a page that already exists and is contradicting itself or burying the answer. Glassdesk keeps the counts separate per category and says which one dominates, so the suggestion reads like "your help centre appears to be missing warranty content — 9 tickets found nothing to cite" rather than a bar chart.

It works with content logging switched off

Worth stating plainly, because most conversation-analytics products need to read and store your customers' messages to do anything.

The gap ranking is derived from non-content columns only — the ticket's category, whether it escalated, and the confidence score. No message text is aggregated, and escalation reasons are never pooled, because a model-written reason can quote the customer. Content logging is off by default. If you turn it on, example questions appear next to each gap to make the list easier to act on; if you leave it off, you still get the full ranked list.

So the privacy-preserving configuration is not the degraded one. That is deliberate — a feature that only works once you loosen your data settings is a feature with a hidden price.

Sentiment, without a survey nobody fills in

Alongside the gaps, each ticket carries a frustration score from 0 to 1. Aggregated, that gives you an average, a count of high-frustration tickets, and — the useful one — a frustration average per category.

Per-category is where it earns its keep. "Customer sentiment is 0.31" is a number for a slide. "Returns runs at 0.62 while everything else sits near 0.2" is a Monday morning task: your returns policy, or your returns process, is making people angry, and now you know which one to read first.

One thing we deliberately tied together: the high-frustration threshold shown in your dashboard is the same constant the escalation gate uses. If those two ever drifted apart, the sentence "high-frustration tickets always go to a human" would be false for every ticket in the gap — so they are the same value in the code, not two numbers that happen to match.

How to actually use it

• Read the escalation list before the deflection number. One is a report card, the other is a work queue.
• Start at the top and write three articles. The list is ranked by volume, so the top of it is where a single afternoon of writing moves the most tickets.
• Re-index and watch the same categories. The gaps you filled should drop out of the list within days. If a category stays, the article you wrote is not answering the question your customers are actually asking — which is itself worth knowing.
• Treat a rising category frustration score as an early warning. It usually moves before the ticket volume does.