The short version: I exported 14 months of invoices from my accounting software, uploaded them to an AI chatbot, and asked it to find patterns a human wouldn't bother looking for. In under an hour it found a client who was 22 percent of my revenue and a slow payer, prices I hadn't raised in three years while my costs went up, and a service I was quietly delivering at a loss. None of this needed a data scientist. It needed a spreadsheet and forty minutes.
Why I did this in the first place
I have an accountant. She is good at her job. She does my VAT return, my year end accounts, and she tells me when a payment is overdue enough to chase. What she does not do, because it is not what I pay her for, is sit and look for patterns across a year of invoices and ask "does this make sense as a business?"
That is a different job. It is the job of someone who has time to stare at numbers and get curious. I do not have that time, and neither do most small business owners I talk to. So I decided to see if AI could do it instead, using data I already had sitting in a folder.
This was not a clever prompt engineering exercise. It was me, on a Tuesday afternoon, exporting a CSV from my invoicing software and pasting it into Claude because I was procrastinating on something else.
The exact steps I took
- Exported 14 months of invoices from my accounting software as a CSV. Client name, invoice date, amount, due date, date paid.
- Stripped out anything I did not want a third party seeing (bank details, addresses) and kept client names as first initial plus industry, so I could still spot patterns without handing over full identifying data.
- Uploaded the file to Claude and separately to ChatGPT, because I wanted to see if they agreed.
- Asked specific questions rather than "analyse this," because vague prompts get vague answers. I asked things like "which clients pay late most often and by how many days on average" and "has my average invoice value changed month to month, and if so why might that be."
- Asked it to flag anything that looked odd, even if it could not explain why.
That last instruction did more work than anything else. It is the one thing most people skip, and it is where the useful stuff turned up.
What it found
Three things stood out, and I will be specific because vague case studies are useless.
First, one client made up 22 percent of my invoiced revenue over the period and paid an average of 19 days late against 30-day terms, every single time, for the whole 14 months. I knew they paid slowly. I did not know it was consistently 19 days, or that a fifth of my income was sitting with one client who treated my invoices as optional guidance rather than a deadline.
Second, my day rate for one type of consulting work had not moved since a certain point three years ago, while a rough calculation of my own overheads (software subscriptions, a part time assistant, travel) had gone up by roughly 19 percent over the same window. I had raised prices for new clients but quietly left legacy clients on the old rate, and never revisited it. The AI did not know my overheads went up. I told it that separately and asked it to do the maths against the invoice data. That combination, my messy real-world knowledge plus its pattern spotting, is where this works.
Third, and this was the uncomfortable one, a smaller service I offer as an add-on was invoiced at a flat fee that, once I mentally added up the hours it took me based on my own calendar, worked out at under minimum wage in effort per hour. Nobody flagged this because on paper it looked like "extra revenue." It was a loss leader I never chose to run, it just happened.
The uncomfortable part nobody likes admitting
Here is the bit that made me slightly annoyed at myself. None of this was hidden. It was sitting in a spreadsheet I already had. I could have found every single one of these three things myself with an hour and a calculator, months earlier, at any point in the last three years. I did not, because looking closely at your own numbers is boring and slightly frightening, and most small business owners, including me, will find almost anything else to do first.
That is the real story here, and it is not really about AI. AI did not find something invisible. It found something I was avoiding. The tool matters less than the fact that I finally sat down and asked the question. If you never ask, no amount of software helps you.
I think a lot of the AI-for-small-business content out there implies the tool does the thinking for you. It does not. It does the noticing for you, faster than you would notice yourself, but you still have to decide what to do with what it noticed. In my case that meant an awkward but overdue conversation with one client about payment terms, and a price review letter to five legacy clients.
Where AI got it wrong (and where it should not be trusted)
I want to be straightforward about the limits, because I have seen too many posts oversell this.
- It confidently suggested a "seasonal dip" in one month that was just me being on holiday for three weeks. It had no way of knowing that from invoice dates alone, and I had to correct it.
- It cannot see cash flow timing issues unless you also give it your outgoings, not just what is coming in. Revenue analysis without cost data tells you half a story.
- It has no concept of your actual relationship with a client. The 22 percent client who pays late is also the one who has referred me three other clients over the years. Any AI tool will tell you to fire that client. It does not know what that client is worth beyond the invoice column, and neither does any spreadsheet.
- It is not a substitute for an accountant on anything involving tax, VAT thresholds, or compliance. Ask it those questions and you will get an answer that sounds confident and is sometimes wrong. I checked everything tax-related with my accountant before acting on it.
So the honest framing is this: treat it as a very fast, very patient junior analyst who has read your spreadsheet but has never met your clients and does not understand your industry. Useful. Not in charge of decisions.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
How to do this yourself this week
You do not need anything expensive for this. Here is the version I would run if I were starting from scratch today.
- Export 12 months of invoices from whatever you use, Xero, QuickBooks, FreeAgent, or even just a Google Sheet you have been keeping by hand. Columns needed: client, invoice date, amount, due date, paid date.
- Add a rough monthly overheads figure if you have one, even an estimate. This is the piece that turns "here is your data" into "here is what your data means."
- Anonymise anything sensitive if you are using a free or consumer AI tool. Use initials for client names, drop addresses and bank details.
- Upload the file to Claude or ChatGPT and ask three specific questions rather than one vague one. Something like: which clients consistently pay late and by how many days, has my average invoice value or day rate changed over this period, and what looks unusual in this data even if you cannot explain why.
- Cross-check anything before you act on it, especially anything involving tax or contracts. Ask a second AI tool the same question and see if the answers agree, then check the important ones with your accountant.
- Give yourself a follow-up task, not just a report. My follow-up tasks were: send a payment terms email to one client, send a price increase letter to five clients, and quietly retire the loss-making add-on service.
The whole exercise took me one afternoon, start to finish, including writing the follow-up emails. If you have never done anything like this, that is a small time cost for what it turned up.
If you are not comfortable doing this alone
If spreadsheets make your eyes glaze over, or the idea of uploading business data anywhere makes you nervous, that is a completely reasonable place to bring in help rather than muddle through it yourself. I wrote a separate piece about how I use AI daily without being remotely technical, and the honest answer is that most of what I do is exactly this kind of thing: giving a tool a specific, boring, real task rather than expecting it to be magic.
If you want someone to sit with your actual numbers and set this up for your business, that is a fair reason to bring in an AI consultant for a few hours rather than an ongoing retainer. It is a one-off project, not a subscription, and it should not cost you what you think it will. I have written about what AI consultants charge if you want real numbers before you have that conversation.
Why this matters more than another content idea
I spend a lot of time writing about AI and content, including some hard lessons, like the time I published 569 blog posts in 45 days and Google only indexed 5 percent of them. That experiment taught me something that applies here too: AI is extremely good at producing volume and pattern recognition, and completely indifferent to whether any of it is useful unless you point it at a specific, real problem with real stakes.
Invoice data is about as unglamorous as business data gets. Nobody writes a LinkedIn post about their invoice CSV. But it is one of the few datasets every small business owner already has, sitting there, unexamined, that tells you the truth about who pays you, what you charge, and where your time goes. That is worth an afternoon far more than most of the AI experiments I have run, and I have run a lot of them since I started rebuilding this business in public, the same way I tracked the numbers when my website started paying me again after a long dry patch. Numbers you look at change decisions. Numbers sitting in a folder change nothing.
Frequently asked questions
Is it safe to upload my invoices to ChatGPT or Claude?
Anonymise client names and remove bank details and addresses before uploading anything to a free or standard consumer AI tool. If you use a business or enterprise tier with a data processing agreement, check the specific terms, but as a habit, strip identifying detail out first regardless of which tool you use.
Will AI replace my accountant for this kind of analysis?
No, and it should not try to. AI is good at spotting patterns across a spreadsheet quickly. It has no authority on tax rules, VAT thresholds, or compliance, and it will sometimes give a confident wrong answer on those topics. Use it for pattern spotting, then check anything financially significant with a qualified accountant before acting.
What if I don't have clean invoice data going back very far?
Six months of data is still useful, though 12 to 14 months is better because it shows seasonal patterns rather than a snapshot. If your records are messy, start today: export whatever you have now and set a reminder to run this same exercise again in six months once you have more history.
What questions should I ask the AI tool?
Be specific rather than asking it to "analyse" your data. Ask which clients pay late and by how many days on average, whether your average invoice value or day rate has changed over time, and what looks unusual in the data even if it cannot explain why. Vague prompts produce vague, forgettable answers.
Related reading: The AI Subscription Stack: What I Pay For (and What I Cancelled) in 2026 and I Got AI to Chase My Late Invoices. Here's What Happened.
