Before you buy AI, look at your spreadsheets

Before you spend anything on AI, open your spreadsheets. AI tools can only work with the information you give them, and if that information is scattered, duplicated or inconsistent, you’ll get answers that sound confident and are quietly wrong. The good news is that getting your data ready is mostly unglamorous tidying, and it pays off whether or not you ever buy an AI tool.

Find out where your information actually lives

Ask a few people on your team where they’d look to answer three questions. How many active clients do we have? What did we invoice last quarter? Who’s our main contact at our biggest customer? You’ll probably hear about the accounting software, a CRM (customer database), a shared drive, three spreadsheets with “final” in the name, someone’s inbox and a notebook. That’s normal.

Write it down. A simple list of each source, what’s in it, who updates it and how often is worth more than any tool demo. It’s also where you’ll often find that the real problem isn’t one AI can fix.

Duplicates and inconsistent naming

The same customer appears as “J Smith Ltd”, “J. Smith Limited” and “Smith, John”. One sheet says 03/04/2026, another says 3 April 2026, and a system set to American format reads the first as 4 March. One column holds “N/A”, “tbc”, “0” and blanks, all meaning different things. People cope because they know the history. Software doesn’t, and AI will happily count three customers where there’s one.

Google’s own guidance for analysing data with Gemini in Google Sheets is: “For the most accurate results, ensure your data is in a consistent format, with clear headers and no missing values.” That’s good advice for any tool.

Pick one source of truth

For each kind of information (clients, contacts, invoices, stock, volunteers), decide which system holds the master copy. Everything else either reads from it or gets retired. If the CRM holds the master contact details, nobody keeps a private list of phone numbers. This is a decision, not a technology. It costs nothing, and it removes a lot of duplication at the source.

Simple tidy-up wins

You don’t need new software for this. Most of the gain comes from a few habits:

• One row per thing (one client, one invoice) and one column per fact. No merged cells, and no subtotals in the middle of the data.
• One header row, with clear, unique column names.
• Dates as real dates in one format, and amounts as numbers rather than text like “£1,200 approx”.
• Drop-down lists (data validation) for columns with set answers, such as status or region, so “Edinburgh”, “edinburgh” and “Edinbugh” stop being three places.
• A separate notes column for free text, so comments don’t end up in the date column.
• Colour as decoration only. Software can’t read “the red ones”.
• A “read me” tab saying what each column means and who owns the sheet.

Let AI help with the clean-up, then check it

AI is genuinely useful for the tidying itself. It can suggest which rows are probably the same customer, standardise names and addresses, split a “full address” column into parts, pull dates and amounts out of free-text notes, or write the formula you couldn’t remember. It’s quick, and it’s also exactly the kind of work where small errors hide. So:

• Work on a copy, never the original.
• Ask it to flag likely duplicates for a person to merge, rather than merging them itself.
• Compare row counts and totals before and after. If 1,204 rows became 1,187, you should know which 17 went and why.
• Spot-check a sample by hand, including the awkward cases.

Asking questions in plain English

Tools such as Copilot in Excel and Gemini in Google Sheets let you ask questions of a spreadsheet in plain English: “which clients haven’t booked since March?” or “what’s our average invoice by region?” With tidy data, this is one of the most useful things AI can do for a small team.

But the tool answers the question you asked, not the one you meant. Does “revenue” include VAT? Is an “active” client someone who booked this year, or someone with an open contract? Which financial year? The tool will pick a definition without telling you, and the answer will look just as authoritative either way.

So you still need someone who knows the numbers: a person who knows roughly what the answer should be, and notices when a figure is twice what it ought to be. Check new answers against a figure you already trust, such as last year’s accounts, before you rely on them.

A short checklist

• Do you have a list of where each kind of information lives, and who owns it?
• Is there one agreed source of truth for clients, contacts and money?
• Does each important spreadsheet have one header row, one row per record and no merged cells?
• Are names, dates and categories consistent, with drop-downs where they help?
• Do you know roughly how many duplicates you have?
• Could a newcomer understand each column from a “read me” tab?
• Is there a person who knows what the numbers should look like?

If you answered no to more than a couple, that’s your first project, and it costs time rather than software. Try the AI readiness self-check to see where you stand. If the tidy-up is the job, data pipelines is how we would build it.