Guide 03 · Step 4 of 5

Implement small

Ship the narrow version, on real inputs, with a human checking everything. Expand afterwards.

Independent guidance from Synergy. ISO/IEC 27001 and ISO 22301 are published by ISO/IEC, which has no affiliation with Synergy and has not reviewed this material. Always work from the published text of the standard.

The failure mode here is building the complete solution before anyone has seen it work on real data. It feels efficient. It produces a system that is confidently wrong at scale.

Run it in shadow mode first

Automate the workflow but do not let its output go anywhere. The human does the task as before, and the automation does it alongside. Compare the two. This costs almost nothing and it surfaces the exceptions your runbook missed — which will be several.

• Start with the real inputs, including the ugly ones. Test data is always cleaner than reality and it flatters every system.
• Log every case where the automation disagreed with the human — and read them. The pattern in those disagreements is the actual specification.
• Set the threshold in advance. You wrote acceptance criteria in step 3. Do not move them afterwards to match the result.

Design the failure, not just the success

Decide now what happens when the automation is uncertain or the input is unrecognised. The options are: stop and route to the human, produce a best guess marked as such, or refuse and log. The wrong answer is producing a confident output with no signal that it was uncertain.