There is a moment in most operations meetings now where someone says "could AI do this?" It is a fair question. It is also usually the wrong first question.

The right first question is: could a rule do this?

Rules first

Traditional automation runs on rules. When an order comes in, create a task. When inventory drops below a threshold, send an alert. Every weekday at seven, assemble the report and post it. Rules are cheap to run, they behave the same way every time, and when they break, they break in ways a log file can explain.

If you can write the instructions for a process on one page, and the instructions do not contain the word "usually," you probably want rules, not AI. Examples from systems we run:

  • Posting low-stock and restock alerts from live POS inventory to a team channel
  • Retiring expired promotions from a website on schedule
  • Reconciling records between a loyalty platform and a POS
  • Assembling a weekly analytics report and delivering it automatically

None of that needs a model. All of it produces daily value.

Where AI earns its keep

AI belongs where the input is messy and human judgment is currently the only bridge. The tell is language: reading, writing, extracting, summarizing, classifying. If the process involves a person looking at unstructured stuff and producing structured stuff, AI is a candidate:

  • Pulling fields out of documents that arrive in twenty formats
  • Drafting content that a person then reviews and approves
  • Answering questions from a pile of SOPs and product data
  • Triaging inbound email into categories a rule could then route

Note the shape of these: AI in the middle, structure on both ends, and a person nearby. The reliable pattern is AI proposes, human disposes, especially anywhere the output faces a customer or a regulator.

The hybrid truth

Most valuable systems we build are hybrids, and the AI part is smaller than people expect. A document intake system might be ninety percent plain software: queues, validation, routing, audit trail. The AI does one job, extraction, inside a machine that rules built.

This matters for cost and reliability. Rules are deterministic and nearly free. Model calls cost money and are probabilistic. Pushing every decision through a model when a rule would do is how AI projects end up expensive, slow, and weirdly flaky.

A decision test you can run in a meeting

  1. Write the process as instructions. If they fit on a page without "usually" or "it depends," build rules.
  2. If the instructions require reading or writing natural language, or judgment across messy inputs, AI is a candidate for that specific step.
  3. Ask what happens when the AI is wrong once per hundred runs. If the answer is "a person catches it in review," proceed. If the answer is "a customer or regulator sees it," redesign until a person is in the loop.
  4. Estimate volume. Automating something that happens four times a month rarely pays either way.

The best outcome of this exercise is often boring: a plain automation, shipped in a week, saving hours immediately. Boring is underrated.

Not sure which side your process lands on?

Describe the workflow. We will tell you whether it needs AI, a plain automation, or neither.

Discuss Your Project