Four AI Systems Built to Show Their Working

Most of what gets sold as an AI bot is a prompt with a text box around it. It answers confidently, it cannot explain how it reached an answer, and it will cheerfully do the one thing you never wanted automated — apologise on your behalf to somebody threatening legal action, or pass a misread invoice total into your accounts.

Three failures cause almost all of the damage, and none of them are model quality:

1. It cannot be audited. A lead is scored 82 and nobody, including the people who built it, can say which answer produced the 82. Sales stop trusting the number, and the system quietly falls out of use.
2. It does not know its own limits. A refund demand, an allergy, a solicitor's letter — these get the same generated politeness as a compliment about the parking.
3. It is confidently wrong. An extracted total that is off by two hundred looks exactly like a correct one. A field the system admits it could not read is far cheaper than a number it invented.

We wanted a house pattern for this work rather than a new judgement call on every build.

Four systems, one shape. Each takes unstructured input, produces a structured judgement, shows the reasoning next to the answer, and hands over to a person when the stakes outgrow the machine.

Lead Qualifier. Four questions, then a score out of 100 — scored on budget, timeline, authority and service fit. The score is never shown alone: every signal appears beside it with the points it contributed, so a salesperson who disagrees can see exactly which answer moved the number. A perfectly funded, urgent enquiry for something we do not sell scores zero on fit and does not reach a consultant's calendar.

Booking Agent. It resolves availability for the chosen day before it offers a single time. A taken hour renders as taken; a closed day offers nothing at all. The agent structurally cannot promise a slot that is already gone, which is the failure that turns a booking bot into a complaint.

Document Extractor. An invoice becomes structured fields, and the figures are then re-added independently. When subtotal plus tax does not equal the stated total, the system says so and names the discrepancy rather than passing a clean-looking number downstream. Fields it could not read say so plainly instead of guessing.

Review Responder. Sentiment and aspect classification, then a drafted reply — except when the review mentions money, health, or legal action. Those leave automation entirely: nothing is posted, and the review is routed to a person who owns both the reply and the follow-up. Knowing when not to answer is the feature.

Rules first, model second. The scoring, extraction and classification logic is deterministic and covered by assertions, so the same input always produces the same output and any failure can be reproduced. A hosted model slots in behind the same interface where it genuinely adds something. The rules are the floor: with no model reachable, the system degrades to explainable scoring rather than to nothing.