Knowella

The AI Agent — ask the question, get the fix

A copilot session that diagnoses 861 failures, finds the 90% cause, and closes its own loop.

5 min read Real product screens
app.knowella.com Live walkthrough
The AI Agent is the copilot that sits over your document pipeline and answers questions about it in plain language. What makes it worth a walkthrough is not that it answers — it is that a single session moves from a vague worry to a specific fix to a verified result, without anyone opening a report. The filmed session diagnoses 861 failures and closes its own loop.
In this article

The empty state is the tutorial

Opening the agent shows four suggestion cards rather than a blank prompt. Each is a question the agent knows how to answer well, which is a better introduction than any help text — you learn the shape of a good question by clicking one. Notably, one of those four cards turns out to be the fix for the problem the session is about to find.

Diagnose: what is actually failing?

The first question is the broad one — what are the most common failure reasons? The agent comes back with four, and the distribution is the finding.

  • unknown_carrier777 failures · 90.2%
  • load_number_missing63 failures · 7.3%
  • missing_pod13 failures · 1.5%
  • pod_incomplete8 failures · 0.9%

Nine of every ten failures are one thing

777 of 861 is 90.2%. That single number reframes the whole backlog: this is not a document-quality problem, an OCR problem, or a carrier-behaviour problem. It is a configuration gap. No carrier profile matched the sender email, so 777 perfectly readable documents were never read.

Fix: the answer was already on screen

The fix is "add a new carrier profile" — which is the third of the four suggestion cards the agent offered on its empty state before any question was asked. The copilot's default suggestions already pointed at the highest-leverage action in the tenant. Following it opens the Carriers flow covered in its own guide.

Verify: ask again, differently

The session closes the loop by asking which documents need review right now. Twenty rows come back, all carrying the same reason. That confirms the diagnosis from a second direction — the same cause seen through a different question — rather than taking the first answer on trust.

How to run your own session

The pattern is worth copying whole: ask for the distribution before asking about any single document, act on the largest slice rather than the loudest one, then re-ask from a different angle to confirm. Three questions. The agent is a diagnostic instrument, and diagnosis beats triage when 90% of a queue shares one cause.

Ready to build your own?

Start free and ship your first app in minutes — or have us set it up with you.