1
00:00:06,960 --> 00:00:11,970
A ticket arrives. Jev reads it and routes it.

2
00:00:11,970 --> 00:00:20,990
Venn, with no history, is given plain descriptions of the 77 queues and routes the same ticket.

3
00:00:20,990 --> 00:00:28,990
Scored on the same 3,080 tickets: Jev 79.7%, Venn without history 69.0%. On day one, use Jev.

4
00:00:28,990 --> 00:00:35,600
Neither has seen one of this desk's tickets yet.

5
00:00:35,600 --> 00:00:41,700
Attach the desk's closed tickets. Nothing is trained; the store is pointed at.

6
00:00:41,700 --> 00:00:51,220
Every answer now comes with the precedents it used, and how each was resolved.

7
00:00:51,220 --> 00:01:05,750
The counter moves to 93.6%.

8
00:01:05,750 --> 00:01:15,910
Five resolved tickets per queue already clear the decision model's interval on this desk; on CLINC150 it takes fifty.

9
00:01:15,910 --> 00:01:28,590
One embedding of the ticket. From it: the queue, the three precedents, and whether it belongs in the queue it landed in.

10
00:01:28,590 --> 00:01:39,700
A ticket a legacy rule filed in the wrong queue: the same vector, one more question, and set_membership flags it.

11
00:01:39,700 --> 00:01:52,840
The registered question, affected by the FX-fee change or not, at threshold 0.9: 5% precision and 9,505 tickets sent to review.

12
00:01:52,840 --> 00:02:02,840
Re-tag the whole backlog with the 77 queues and file the two FX queues: 361 tickets, 98% precision, 99% recall, 1.9 seconds in process.

13
00:02:02,840 --> 00:02:13,990
A sweep is as good as its question. The queues the desk already has are the better question.

14
00:02:13,990 --> 00:02:18,990
“I was NOT charged twice. One charge, but my statement shows two lines.”

15
00:02:18,990 --> 00:02:24,990
Venn's precedents say duplicate charge. Jev reads the negation.