AI Features That Earn Trust
In the sales demo, Loomwork's AI model read a clean PDF invoice perfectly. Harbourline's real inbox holds scanned photocopies, handwritten corrections and twelve-page invoices covering forty shipments. When Meera asks, "How accurate is it?", the answer "very good" will not do. She needs to know what it gets right on her invoices, and what happens when it gets something wrong.
The rule
Measure on their data, gate by confidence, keep a human on costly calls.
- Measure on their data. Build an evaluation set: 300 real Harbourline invoices, with the correct values filled in by clerks. That is your answer key. Score the model field by field and carrier by carrier, and re-run the whole set after every change to the prompt or the model.
- Gate by confidence. Let the system act alone only when it is confident and the figures pass independent checks: the line items add up to the total, and the LR number exists in the TMS. A model's own confidence score is useful, but a sum that adds up is much harder to fool.
- Keep a human on costly calls. Mistakes are not equal. Misreading a free-text note costs nothing; misreading an amount can pay the wrong money. Costly fields that fail a check go to a clerk.
Every clerk correction is a new labelled example. Add it to the evaluation set, and the measurement keeps getting more honest.
Data rules still apply: the model runs only where Kavitha has approved, and Harbourline's invoices are not used to train anything without their written consent.
The shortcut
Route by the cost of being wrong. For each field the AI reads, ask what a wrong value would cost. High cost: check it, or send it to a person. Low cost: let it through.
Measure
On 300 labelled invoices, the model reads invoice totals correctly 97% of the time and extra charges 88% of the time. On scanned photocopies, only 71% of fields are right.
The trap is one proud number. "95% accurate overall" can hide a segment, such as scanned invoices, where the system is wrong almost a third of the time.
Drills
Answer fast — trust the rule.
Meera asks, "How accurate is it?" What is the best answer?
Overall accuracy is 95%, but scanned photocopies score 71%. What should you do?
Which cheap, independent check catches a misread amount?
You changed the prompt to fix one carrier's invoices. What should you do next?
Which field most needs human review when its check fails?
Remember
- Measure on their data, gate by confidence, keep a human on costly calls.
- Shortcut: route by the cost of being wrong.
- Trap: one overall accuracy number that hides a weak segment.