Everyone's AI is the same. Your ERP is the moat.

There is a strange anxiety in mid-market companies that they are "behind on AI" because a competitor has a better model. Nobody has a better model. Everybody is calling the same three or four providers, and whichever one is ahead this quarter will not be ahead next quarter.

What is not evenly distributed is the thing that makes a model useful: your operational data, and your ability to let something act on it.

What you have that cannot be bought

• Forty thousand resolved support tickets. Every one is a labelled example of how your company talks to your customers, what your actual policy is in practice, and which exceptions you really make. This is the highest-quality training and grounding material in your business and it is sitting in your helpdesk being treated as an archive.
• Twenty years of transactions. Who buys what, when, in what combination, with what seasonality and what failure patterns.
• Your exception history. Every time somebody overrode the standard process and why. This is the tacit knowledge of the company, accidentally written down.
• The systems of record themselves. The ERP everyone complains about is the reason an AI can do something rather than say something. A competitor with a better model and no integration has a chatbot.

Which is why the integration is where the work is

People are consistently surprised by where the effort goes in these projects. It is not prompting and it is not model selection. It is:

• Getting clean, current data out of systems that were not designed to be read by anything but their own UI — and out of the vendor portal with no API, and the fixed-width text file that lands at 4am
• Building a typed, validated, reversible action layer over systems that assume a human is clicking
• Finding the account-level overrides and exceptions that live in fields nobody mentioned, or in somebody's head
• Reconciling the documentation with what the company actually does

That is unglamorous, it is most of the timeline, and it is also exactly why the result is defensible. Nobody can copy it, because it is a map of how your specific company works.

The practical consequence for how you buy

Three things follow, and they should change your shortlist:

1. Prefer systems where you keep the layer you built. The action taxonomy, the policy configuration, the evaluation set, the accuracy history, the ledger — all of it should be yours in a portable format. Ask what you keep if you cancel in eighteen months. If the answer is "your data" but not your configuration and evaluation history, you are renting the part you paid to create.

2. Do not let anyone talk you into replacing the ERP first. "We'll do AI after the migration" is how three years disappear. The integration work is the same either way, and doing it now produces a working system and a much better understanding of what you actually need from the next ERP.

3. Stop worrying about which model. Build so that swapping it is a config change and a re-run of your evaluation set. Then genuinely stop thinking about it.

The uncomfortable corollary

If your data is genuinely bad — no clean product data, no structured ticket history, an ERP nobody trusts — then you do not have an AI problem, you have a data problem wearing an AI costume, and the honest sequence is to fix some of that first.

We say this in assessments and it has cost us work, which is roughly how you know it is true. The good news is that "some" is doing real work in that sentence: you need clean data for the one process you are automating first, not across the enterprise. Pick a process whose data is already decent, ship it, and use the win to fund the rest.