The idea
The AI tools most businesses are looking at today are language models. A language model produces a likely answer based on patterns in what it was trained on and in what you give it. It is making a judgment, not looking something up and not calculating.
That one fact explains most of what these tools do well and most of what goes wrong with them.
Where it fits
- Drafting. A first version of an email, a report, a summary, or a reply, for a person to finish.
- Summarizing and extracting. Pulling the key points, dates, or figures out of long documents.
- Sorting. Putting requests, tickets, or documents into categories.
- Answering questions over your own material. When the system is given your documents to draw on, it can answer from them.
- First-pass review. Flagging what a person should look at more closely.
Where it does not, on its own
- Facts it was never given. Asked about your orders, your customers, or last week, it has nothing to go on unless it is connected to that information. It may answer anyway.
- Exact arithmetic and record keeping. These belong to ordinary software. An AI system can call that software, but that is engineering work, not a setting.
- Decisions someone must answer for. It can inform a decision. It cannot be accountable for one.
- Tasks where a wrong answer is costly and looks right. A confident wrong answer is the characteristic failure. Where that matters, a person has to check.
Three questions that sort any task
- Is the task mostly about language or patterns, or about exact records and numbers?
- If the answer is occasionally wrong, will someone catch it before it does harm?
- Is the information it needs available to it, or would it have to guess?
A made-up example
Picture a distributor whose staff spend much of the day answering "where is my order?". An AI assistant that is connected to the order system and drafts the reply is a good fit: language-heavy, checkable, and working from real data. The same assistant with no connection to the order system is a poor fit. It would have to guess, and it would guess fluently.
Exercise · about 10 minutes
List three tasks in your business that take real time each week. Put each through the three questions and sort it: good fit, poor fit, or fits with a person checking.
One thing to take to your team
Ask of any AI proposal: where does it get its information, and who checks the answer?