Data Readiness: What Your Excel-and-ERP Landscape Needs Before AI Is Useful

The typical established SME runs on an ERP that holds orders, articles, and invoices, surrounded by a ring of Excel files that hold everything the ERP was too rigid for: the real price calculation, the capacity plan, the customer list with the notes that matter. Add a shared drive with twenty years of documents, an email archive that is the true system of record for what was agreed, and often a machine or two writing logs nobody reads. Nothing about this is shameful, it is what decades of working pragmatically look like.

The problem for AI projects is not that this landscape is messy, it is that the same fact lives in several places with several values. The customer's delivery address in the ERP, the corrected one in the Excel file, the newest one in an email from last Tuesday. A language model reading these sources will not tell you which is current, it will fluently use whichever it was given. Before asking what AI can do with your data, you have to be able to answer a simpler question: for each fact that matters, which source is the truth?

The expensive mistake is concluding that you must clean all data before any AI, which becomes a year-long consolidation project that exhausts the budget and goodwill before anything useful ships. Data readiness is not a company-wide state, it is a per-use-case question. AI-supported quoting needs current article data, price rules, and inquiry access. A service knowledge assistant needs manuals and service reports. Neither needs your accounting archive or the legacy CRM cleaned.

So invert the order: pick the use case first, then list the specific data it touches, then assess only that. For each source ask three things. Is it current, or does the real version live in someone's inbox? Is it accessible, meaning exportable or reachable through an interface, rather than locked in a system only one person can query? And is it consistent enough, meaning the same article or customer is identified the same way across the sources this one workflow joins? Three honest answers about one process beat a company-wide data audit every time.

Some fixes are prerequisites, and they are smaller than feared. Decide the leading system for each entity the use case touches, customers in the ERP, prices in the calculation sheet, and write that decision down so the AI integration reads from the right place. Stop the worst duplication going forward, even if you do not clean the past. Get export or API access to the systems involved, which for older ERP versions is often the single hardest step and worth scoping early. And move process-critical knowledge out of personal inboxes into something shared, at least for the process in question.

Other projects can wait, despite what a data consultant may tell you. You do not need a data warehouse, a company-wide data model, or perfect historical data to draft quotes or answer service questions. Historical cleanup matters when you get to analytics and forecasting, which is a later chapter. There is one caveat worth taking seriously now: access rights. The moment an AI system can search across sources, it will surface things that were effectively protected by being hard to find, salary lists on the shared drive being the classic case. Sort out who may see what before the search gets good.

For your chosen use case, run a small structured exercise. Have the people who do the process today walk through five real cases end to end, and note every system, file, and inbox they touch. For each touched source, record the three answers: current, accessible, consistent. Then attempt the smallest possible technical proof, exporting the relevant data and letting the AI process a handful of real cases, before signing any larger project. The gaps this surfaces in two weeks are the same ones that would otherwise surface in month four of an implementation.

Treat what you find as a punch list, not a verdict. Most SMEs discover the data for their first use case is closer to ready than the general mess suggested, once the question narrows from is our data good to is this data good enough for this. And every gap you close, one leading system named, one export unlocked, one inbox emptied into a shared store, pays off beyond the AI project, because it is the same foundation your future reporting, integrations, and eventual second and third AI use cases will stand on.