Artificial intelligence has an uncomfortable property: it amplifies what it finds. If a company's data is in order, AI speeds up decisions. If it is scattered, duplicated or contradictory, AI produces fast, confident answers… about the wrong information. That is why, in almost every project, the first job is not AI but order.
Symptoms of messy data
You don't need an audit to know whether there is a problem. A few familiar signs are enough:
- There are several versions of the same spreadsheet, and nobody is sure which one is the latest.
- Sales and accounting report different figures for the same month.
- The weekly report depends on one specific person being available.
- To answer a simple question, three files have to be cross-checked by hand.
If several of these situations sound familiar, any automation built on top will inherit the disorder.
A single source of truth
The goal is not to have more data, but to have a reliable version of the data that matters. Getting there takes three steps.
Inventory the sources. Where the information lives today: the ERP, spreadsheets, email, the billing platform. Often the map alone already reveals duplicates.
Agree on definitions. What is an "active customer"? When is an invoice "overdue"? If each area answers differently, the indicators will never match. Writing these definitions is a short exercise with an enormous impact.
Assign owners. Every data set needs someone who is accountable for its quality. Without an owner, the order lasts only as long as the initial enthusiasm.
Getting things in order is not an endless project
A common mistake is to want to order everything before moving forward. It works better to start with the indicators that are actually used to decide: cash flow, receivables, inventory, sales. Those sources are put in order, centralized, and the work is built on top of them.
The point is not to have perfect data, but reliable data for the decisions that matter.
Once that core is working, the rest of the information is added in stages, as specific needs appear.
Small rules, big difference
Data quality is decided, above all, at the moment it is recorded. A few simple rules prevent most problems:
- Validate data when it is entered, not when it is reviewed at the end of the month.
- Use one standard format for dates, currencies and codes.
- Detect duplicate customers and products before they multiply.
- Record who changed what, and when.
None of these practices requires sophisticated technology, but all of them make what comes next work.
Then, AI
With the data in order, AI starts to deliver for real. A copilot can answer "how is receivables doing this month?" by querying a reliable source, instead of guessing between versions. A document search system can cite the right contract. And the indicators update themselves, without anyone having to build Monday's report.
In summary
Ordering data is not the most visible part of an AI project, but it is the part that determines whether it works. A single source of truth, shared definitions, clear owners and a few quality rules are enough for decisions, human or automated, to be made on information you can trust.