~/build-log

Where AI actually pays off in operations (and where it doesn't)

Every second conversation I have these days starts with "we should add AI to this." Sometimes the answer is yes. More often, the honest answer is: you don't have an AI problem, you have a process problem - and AI will just help you do the wrong thing faster.

After ten years of building automation for ecommerce, HR and back-office operations, here's the mental model I use.

The quadrant that matters

Score any candidate process on two axes: volume (how often it happens) and judgment cost (how expensive a wrong decision is).

  • High volume, low judgment cost - this is AI territory. Ticket triage, invoice matching, product data enrichment, order anomaly flagging. A 95% accurate model with a human fallback beats a human doing it 100% of the time, every time.
  • High volume, high judgment cost - automate the preparation, keep the human decision. AI assembles the context, drafts the answer, highlights the risk; a person clicks approve.
  • Low volume, low judgment cost - don't bother. The integration will cost more than it saves. A checklist is fine.
  • Low volume, high judgment cost - this is where experienced people earn their salary. Leave it alone.

What actually breaks AI projects

It's almost never the model. It's:

  1. No ground truth. If nobody can tell you what a "correct" output looks like, the model can't either.
  2. Undocumented process. If the process lives in three people's heads and disagrees between them, you'll automate the disagreement.
  3. No fallback path. The 5% of cases the model gets wrong need somewhere to go. If that path doesn't exist, the 5% becomes support tickets and refunds.

Automate the boring, prepare the important, and never let a model make a decision nobody is accountable for.

Where I'd start in 2026

If you run an operations-heavy business and want one place to start: take your highest-volume manual workflow, write down what "done correctly" means, and measure how much time it eats per month. If it's more than one full-time salary and the failure cost is low - that's your pilot. Everything else can wait.

Working on something in this space? I take a small number of consulting and development engagements.

Let's talk