A 40-person logistics firm once asked me to automate their customer complaint responses. The volume was there. Around 30 a month, all arriving by email, all needing a reply inside a day. On paper it was a clean candidate.
Someone changes a process in March. The page describing that process still says what it said in January. Nobody decided to let it drift. It drifted because updating the page was a separate job, and separate jobs get dropped.
Someone asks what discount a reseller qualifies for. Three people answer. All three sound certain, two are wrong, and nobody can point to where the rule actually lives.
AI sounds like a developer tool. Read the coverage and you would think the whole point of it is shipping code faster.
The automation did not break because the AI got dumber overnight. It broke because the one person who understood it was on vacation.
Your team can read a messy document and still get it right. They fill the gaps from memory. An AI tool cannot. It only has the words on the page.
Most critical tools do not arrive with a launch. They creep in. Someone tries an AI tool on a Tuesday to save twenty minutes, it works, and a year later three people cannot do their jobs without it.
A customer disputes an invoice. You pull the file and find the figure came from an AI draft nobody checked. Now you have a question you cannot answer: who approved this, and on what basis?
Most AI mistakes do not happen while someone is watching.
Series: Building an AI-Ready Business — Part 3 of 3 Part 1: Process problem · Part 2: Writing SOPs · Part 3: Decision logs · Series hub
SOPs tell you what to do. They never tell you why you do it that way. That gap is where onboarding stalls, handoffs leak, and AI agents misfire when conditions change. The fix is a decision log.