AI, Meet ERP

Keep your system of record and your system of inference as two separate machines. Engineer the boundary between them.

Most AI-in-ERP projects fail by asking one machine to do another's job. Your ERP is a system of record – deterministic, arithmetic, auditable. A large language model is a system of inference – fast, linguistic, and wrong quietly. This series keeps them apart and engineers the boundary between them, one real manufacturing workflow at a time, with every number computed from a published, sourced model company. New post every week.

Here is how we make that concrete.

The series follows one fictional company – Spiese Fluid Power, an $85M hydraulic-power manufacturer that also distributes – through ten real workflows. Spiese is a composite, but it is not made up: every figure we publish about it is computed from a single published set of operating assumptions, each one sourced to a public benchmark or flagged as a stated choice you should replace with your own. A series arguing that arithmetic belongs to the system of record should not run on invented numbers, so it doesn't.

Each post works a single problem all the way through the same spine: what it costs today, where the boundary sits, what the AI does and what it must never do, how you would know it worked, and the smallest version you could start on Monday.

We should be plain about our own position. HarrisData is family-owned, with no outside investors to impress – so when the honest answer is "this technology isn't worth buying yet," we can say so. Our own ERP, like most of our customers, runs on the IBM midrange; where a post notes that something is easier there, that is our home turf talking, and we have kept the advice itself platform-neutral.

One more disclosure, and a fitting one for a series about using AI well: we used AI to help make it – as a research assistant, an editor, a graphics consultant, a fact-checker, a plagiarism checker, a consistency checker, and more. But the ideas are ours, the words are ours, and the responsibility for publishing them is ours. That is the boundary this series argues for, applied to the series itself: the machine reads, drafts, checks, and explains; the people decide what is true, and answer for it.