NATURAL LANGUAGE PROCESSING ROUTING WORKSHEET Space Office | https://www.spaceoffice.ai/blog/natural-language-processing Fictional one-message examples. B = booking, S = service question, R = review. A request for a visit or a change goes to booking. A service question with no request goes to service. Unresolved cancellations and unclear requests go to review. Case-insensitive baseline: suggest booking if the text contains book, appointment, or slot. Otherwise give no booking suggestion. Candidate outcomes are hand-authored. ID | Human | Candidate | Message 01 | B | B | Can I book a service early Saturday? 02 | B | B | Can we move my repair appointment to Friday? 03 | B | B | Is there a slot for a brake adjustment this week? 04 | B | B | Can I book a tune-up, and what does it cost? 05 | B | B | Could I bring my bike in Tuesday? 06 | B | R | Can I drop off my bike for a tune-up Thursday? 07 | S | S | I do not want an appointment. Do you sell inner tubes? 08 | S | S | Do I need to book just to buy a pump? 09 | R | R | Has my booking actually been canceled? 10 | S | B | What does a tune-up cost? 11 | S | S | How long does a puncture repair usually take? 12 | R | R | Did that change go through? I have two different emails. 13 | S | S | Do you work on hydraulic disc brakes? 14 | S | S | Do you carry toddler helmets? 15 | S | S | Is a tire patch included in a standard repair? 16 | S | S | Which derailleur brands do you service? Booking is positive. Count review as no booking suggestion for the binary comparison. Baseline TP 4, FP 3, FN 2, TN 7. Precision 4/7, recall 4/6. Candidate TP 5, FP 1, FN 1, TN 9. Precision 5/6, recall 5/6. Candidate sent 3 examples to review, including 1 actual booking. For a real test Queue policy and approved input: Development and final case selections: Baseline and candidate versions: Allowed actions and review owner: Reason for each misroute: Maximum review backlog and stop condition: Human decision: