From spreadsheet chaos to a usable data base: preparing your numbers for AI
AI agents are only as good as the data they work with. How a small business can get its sales, costs, customers and stock into a shape that is actually useful.
- Start with four data sets: sales, costs, customers and stock.
- Keep one source of truth per topic and use clear, consistent names.
- Update on a fixed rhythm; a slightly older but reliable data set beats a fresh but messy one.
Why data comes before AI
Ask an agent for a margin analysis and it will produce one. Whether the result is useful depends almost entirely on the data underneath. If sales are in one spreadsheet, costs in another and half the customer list lives in an e-mail inbox, the agent has to fill the gaps, and that is where errors come from.
The good news: you do not need a data warehouse. A handful of clean, regularly updated tables is enough for most small businesses.
Which data actually matters
Sales: what you sold, when, to whom (if known) and for how much. A shop exports this from its till system or online shop, an agency from its invoicing tool.
Costs: rent, wages, purchasing, software, marketing. An export from your bookkeeping or from your tax adviser is usually the best source. Sensible categories matter more than detail; an agent can work well with a manageable list of cost categories.
Customers and stock: a customer list with name, contact, first and last order and status; for businesses that sell goods, a stock list with article, quantity, minimum stock and supplier. A hotel might replace stock with room types and occupancy, a practice with services and appointment types.
Clean CSV exports
Most tools can export CSV, and CSV is the format agents read most reliably. A few rules make a big difference: one row per record and one column per field, a single header row with clear column names, no merged cells, no subtotals in between and no colour coding that carries meaning.
Use one date format throughout, ideally year-month-day. Enter amounts as plain numbers without currency symbols in the cell and name the currency in the column header. Watch the decimal separator: German exports often use a comma, which can be misread if the file is opened with English settings. Check a few rows after every export.
One source of truth and simple naming
For each topic, decide where the valid version lives. If the customer list exists in your CRM, in a spreadsheet and in your newsletter tool, choose one as the master and treat the others as copies. Otherwise you will spend more time reconciling lists than working with them.
Use names a stranger would understand. A file called sales_2026-09 is better than final_v3_new. Column names like net_amount_eur or invoice_date remove ambiguity. Write customer and article names the same way everywhere; the same customer under three spellings counts as three customers in any analysis.
How often to update
Match the rhythm to the decision. Stock that can run out within days needs frequent updates. Sales and costs for a monthly review need a monthly export on a fixed day. Customer lists can often be updated whenever something changes.
A slightly older but reliable data set is more useful than a fresh one full of gaps. Note when each data set was last updated, so that you and the agent know how current an answer is.
What agents can derive from it
With clean sales and cost data, an agent can show which products or services carry your margin, which months are weak and how fixed costs relate to revenue. With a customer list, it can find customers who have not ordered for a while and draft a friendly follow-up for your approval. With stock data, it can flag articles that are about to run out.
With all of it together, it can prepare a first financial plan, a business plan draft or the figures for a funding application. In NEXUM, this data lives in one place on the platform and the agents work on it directly. When something is missing or contradictory, they ask instead of guessing.