AI for Distributors: A Better Question Beats More Rules

AdobeStock Photos

Most bad AI answers start with a vague question. Fix the ask, decide how specific to be, and check the answer before it goes out.

Picture a purchasing coordinator at a distributor who drops a manufacturer’s new-item file into an AI tool and types, “Clean this up for our item master.” What comes back looks great. Descriptions trimmed, categories filled in, every row tidy. Then someone at the branch notices a line of ball valves now sells by the box instead of the each, because the tool guessed at what the unit column meant.

So she does what most people do next. She writes rules. Don’t change the unit of measure. Keep the manufacturer part number exactly as written. Leave the GTIN alone. Flag anything you aren’t sure about. The next prompt runs three paragraphs, the output gets a little better and a lot slower to get, and by the fourth round she’s arguing with the tool instead of working the file.

Anyone who has worked a supplier file through one of these tools will recognize the pattern. A vague ask, a wrong answer, and then a fence built around the wrong answer. The fence is where the frustration lives, and it’s aimed at the wrong end of the problem. The output wasn’t the defect. The question was.

“Clean this up” could mean a dozen things. Shorten the descriptions to fit the ERP field? Map the manufacturer’s categories to yours? Normalize units? The tool can’t see which one she meant, so it picks, and it does all of them with the same confidence. A better ask is one sentence: “Shorten the descriptions to fit our ERP description field, and leave every other column exactly as it is.” No fence required.

Quotes work the same way. Hand a tool a contractor’s takeoff and ask it to “build a quote,” and it will quietly supply whatever it doesn’t know, whether that’s a lead time, a substitute brand, or a freight line nobody asked for. Ask it instead to list every line on the takeoff that isn’t in your item master, with no substitutes suggested, and you get something an inside salesperson can check in a few minutes before deciding what to do about the gaps.

Some of the fencing habit is left over from how people learned these tools. Early on, getting a usable answer took long specifications, stacked rules, and guardrails, and plenty of people still write prompts as if they’re configuring a server. The tools have changed a lot since then. For most everyday work, what’s left for the person is a short, specific ask and a look at what came back.

Specificity isn’t a rule, though. It’s a dial. Turn it up when the stakes are yours: a quote going to a contractor, a substitution on a submittal, a unit of measure that feeds the item master. The model doesn’t know your line card, your customer, or what the engineer approved, so tell it. Turn it down when the machine is better at the shape of the thing than you are. Ask for a first draft of a line-card overview for a new inside salesperson, or a plain-English read of a manufacturer’s spec sheet, and you’ll often get a better result by leaving room than by dictating every sentence.

Nobody’s model is too dumb to follow a clear ask. The trouble is the gap. You hand it one, it fills it with a guess, and nobody looks. Or take a counter salesperson who asks what replaces a discontinued faucet cartridge and reads the answer straight to the plumber at will-call. It’s the same problem as the valve file. The question was loose, which was fine, and the answer went out unchecked, which wasn’t.

Vague is fine. Unchecked is not.

The Applied AI use cases ASA has collected from member companies point the same way. As ASA put it in September, “AI drafts, retrieves, flags, and assembles. A person validates it, prices it, quotes it, and has the conversation.” That check is what makes a loose question safe to ask. A rep agency building a line review deck and a manufacturer answering a distributor’s product question are in the same spot. Ask plainly, then read before it goes out.

For owners and branch managers, that changes what AI training should look like. A shared folder of long prompt templates teaches people to build fences. The better habit fits on an index card taped next to the counter terminal. Say exactly what you want changed and what you don’t. Decide how tight the ask needs to be by asking who gets hurt if it’s wrong. Read the answer before it leaves the building. People who work that way spend less time fighting the tool and more time on the file, the quote, and the customer in front of them.

By Nils Swenson