Where AI actually belongs in your operations (and where it doesn't)

Every business I talk to right now is under pressure to “add AI” and short on answers about where. The honest answer is that most processes should not get a model anywhere near them, and a few are transformed by one. After two years of building these systems for real operations, here is the test I actually use.

Sort the work into three piles

Take any recurring process and break it into steps. Each step lands in one of three piles:

• Mechanical: moving data, formatting, math, filing, notifying. Rules cover it completely.
• Judgment: summarizing, interpreting, drafting, classifying the messy middle cases. A capable person does this with context, not a rulebook.
• Accountability: approving, committing money, signing, telling a client something is true.

The mistake behind most failed AI projects is putting a model in the wrong pile. Mechanical work does not need a model; a script is faster, cheaper, and never hallucinates. Accountability work must not get a model; someone’s name has to be on the decision. The judgment pile is where AI earns its keep.

Scripts for the mechanical, models for the judgment, humans for the accountability. Systems that respect those lines ship. Systems that blur them get turned off within a quarter.

What this looks like in practice

A drafting step is a perfect example. In a reporting system I built for an agency, scripts pull every number and run every audit, a model interprets the results and writes the narrative, and a person reviews the draft before a client ever sees it. The model never touches the data collection, and it never sends anything.

Compare that with the request I decline most often: “have the AI just handle the client emails.” Sending is accountability. The workable version is a model that drafts the reply in your voice with the relevant history attached, and a human who reads it and hits send. You keep ninety percent of the time savings and all of the trust.

Signs a process is ready

Not every judgment step is worth automating this year. The ones that pay off share a profile:

1. It recurs. Weekly beats quarterly. Automation is rent, and the process has to pay it.
2. The inputs are reachable. The data lives in systems with APIs or exports, not in someone’s head.
3. A wrong draft is cheap. A human catches it in review and the cost is a minute, not a refund.
4. Someone owns it. A named person wants the hours back and will tell you when the output drifts.

Score your most annoying recurring process against those four. If it clears them, the mechanical steps become scripts this month, the judgment steps get a model with a template and a reviewer, and the accountability stays exactly where it is.

That is the whole strategy. It is less exciting than “AI runs the business” and dramatically more likely to survive contact with a real one. If you want a second opinion on which of your processes clear the bar, get in touch.