Practical AI for distributors, built on the data you already have.

Most distributors do not know what they do not know

The value is rarely in a report somebody already runs. It is in the question nobody has thought to ask.

Consider one that comes up constantly: which of your customers are not buying something that customers just like them buy routinely? There are only two answers. Either that customer is genuinely unique, which is less common than people assume, or it is a missed opportunity that has been sitting there quietly for years. And if nobody looks, nobody asks, and it stays missed. That is not a sales problem or a data problem. It is a question-asking problem.

What this looked like for one distributor

A distributor wanted to understand the blind spots in its product mix, sales and customer usage. The raw material was ordinary: sales history already sitting in Prophet 21.

I built a system that analyzed buying patterns across the whole customer base, grouped customers whose purchasing behaved alike, and then identified what most of a group bought that a particular customer did not. Assembling and shaping the data used ordinary data engineering. The analysis itself was done by AI, which is what made it possible to work across the entire history rather than a sample somebody had time to look at.

The output was not a report that lands in an inbox and dies. Sales reps got an interactive dashboard where they could view or print notes for every customer in their territory, and prepare for a conversation or a sales meeting without an arduous manual dive through the data.

What changed is the shape of the conversation. Reps walk in with something useful and fact-based to discuss, instead of relying on whatever comes to mind or letting the customer set the agenda. The talking points cover both what the customer is already spending, explained back to them clearly, and specific areas where their business could grow. It took about eight weeks.

What I do

• Work out what is worth asking. The most expensive mistake in AI is building something impressive that answers a question nobody needed answered. This comes first, and it is most of the value.
• Build AI into the systems you already run. Analysis against your ERP data, and AI inside the applications and integrations I build, rather than a separate tool somebody has to remember to open.
• Get your data ready. Surfacing the right tables and files, and writing the ETL work that lets AI reach them cleanly. On an on-premises system this is the bulk of the effort.
• Roll tools out to your people, and train them. Deployment is the easy half. Knowing what to use it for, and where not to trust it, is the half that decides whether it sticks.
• Design ways to use AI to grow the business. Serving customers better and finding revenue that is already within reach, rather than automating cost out of the back office and calling it a strategy.

Cloud makes this easier, and on-premises does not stop it

Having your data in the cloud changes the AI landscape, and the reason is mundane: access. It is trivial to let an AI read from a cloud system securely. Walling data inside a local security perimeter and still letting AI reach it is harder — not impossible, just considerably more involved.

In practice that means an on-premises distributor pays for more piping. I surface the specific data points, files and tables that a given question needs, and write ETL procedures where the data is not shaped for it. That work is real, and it is worth knowing about before you budget for a project.

It also means the move off on-premises P21 is worth more than the migration itself. Distributors planning that move are, without necessarily intending to, removing the main obstacle to using AI on their own data.

Where AI does not pay off

Every AI page on the internet is a list of promises. Here is the other half.

Someone who is good at ChatGPT is not an AI capability. It is a reasonable place to start, and it is not the same skill. Production work requires knowing how to frame a question, where a question can introduce false information, and how to specify a result that is repeatable and can run across mass amounts of data. Most distributors do not have time to become experts in what is, realistically, another specialty. They need the tool, not another discipline to master.

Expectations fail in both directions, and both are expensive. Some owners are certain AI cannot help because it “doesn’t know my business” — I have been told more than once that they would rather have actual intelligence. Others assume they can throw information at the wall and AI will sort it out. The second is the more costly error. Information that is not understandable by the humans in the room is unlikely to be processed well by AI either.

Too little data is a real limit. This is about breadth rather than company size. A distributor with only dozens of customers may not have the range for the kind of comparison that makes recommendations meaningful. Disorganized data is a different matter: that is a first phase, not a disqualification, and AI is good at helping organize it before it analyzes it.

Readiness is relative to the objective. There is no general answer to “is our data good enough”. If you want detail about product specifications you have never recorded, the first project is creating that data. If you want to know which items tend to sell together, you almost certainly already have what you need. This is the part where an outside professional earns the fee: getting clear on what result you actually want and why, then staging the work so each phase gets closer to it.

If your experience of AI so far is a promising prototype that got less reliable the more you asked of it, that has its own explanation — see why AI projects stall after the demo.

Questions distributors ask about AI