Rules before AI: A rule card for a reliable process

The decision comes before the model

An automated workflow becomes reliable when its inputs, rules, permitted actions and exceptions are explicit. Anthropic recommends starting with simple systems and comprehensive evaluation, adding complexity only when it serves a clear purpose.

That is a practical order for business processes: describe the decision first, then automate the repeatable parts. The rule card download gives you two working pages to define the rules and check a first case.

Hofmann as a concrete example

In the Hofmann workflow, a prospect selects a service, enters deployment details, has distance calculated, sees surcharges and receives a quote. The automation is explainable because each step has an input, a rule and an output.

The case material records 38 qualified enquiries in 7.5 weeks, 21 of 38 outside office hours and a cycle time under 60 seconds. The approximately 25-minute manual baseline is estimated. Conversion and the resulting net benefit are modelled or still pending; this article does not present them as measured revenue. Human handling time is not inferred from the cycle time.

Your rule card

Complete the fields in this order:

1. Trigger: What starts the workflow? What data exists at that moment?
2. Required fields: What must be present before a decision is possible?
3. Decision rules: Write three rules in the form “If ..., then ...”.
4. Permitted output: What may the system confirm, calculate or send by itself?
5. Stop condition: Which input must be handed to a person instead of guessed?
6. Measurement: Which metric shows whether the workflow works? Set a baseline and period.

For example: “If deployment details are complete and the surcharge matches a documented pricing rule, a quote may be generated.” A counterexample is: “If distance is outside the defined range, stop and hand the case to a person.”

A concrete boundary test

If the current pricing rule defines a night surcharge from 22:00 to 06:00 and the first 20 km without travel cost, the register should include 21:59, 22:00, 05:59 and 06:00, plus 19 km, 20 km and 21 km. This is a proposed boundary test, not a result report. Before execution, the accountable person must confirm the current rule version.

Test before production

Use at least one normal and one incomplete case. Write the expected output before running either. A suggested quoting test set can include a documented surcharge, night and Sunday requests, missing deployment details and a distance outside the normal range; these are proposed test cases, not reported Hofmann results.

Assess four things separately: time to output, completeness of captured data, share of human handoffs and later conversion. A fast answer alone does not prove economic impact.

What the card makes visible

The card separates work that is suitable for automation from work that needs judgement. It also shows where a model might help later and where a fixed rule is enough. That makes it harder to hide an undefined process question behind a model.

Working prompt: Choose a process with recurring inputs. Complete the card in 20 minutes. If you cannot name a stop condition and a measurement, the process is not ready for automation.