Straight answer: if you export your last two or three years of invoices and feed them into an AI tool with the right questions, it will show you patterns in your own pricing that you have been too close to see, usually within an afternoon. It will not fix the pricing for you. You still have to make the call, put the number up, or fire the client. That part is still entirely yours.
Why I turned my own invoices into a data project
Five years ago my business fell apart in the way businesses do when the world changes and you don't change fast enough with it. I have written about the rebuilding elsewhere, but this year I stopped rebuilding on gut feel and decided to look at the actual numbers instead of the story I was telling myself about my pricing. So one wet Tuesday in January I pulled every invoice I had raised between 2021 and the end of 2025 out of Xero. 187 invoices. I stripped out client names, kept the scope, the fee, whether it was a retainer or a one-off project, and pasted the whole lot into a spreadsheet.
Then I fed it into Claude in batches of twenty rows at a time, alongside my Toggl time-tracking export for the same period, and asked it to look for patterns in profitability. Not "am I charging enough" in the vague, LinkedIn-poster sense. I wanted to know which specific client types, project sizes, and scopes were quietly costing me money once my real hours were counted.
What I did, step by step
- Exported invoices from Xero as a CSV, going back three tax years.
- Anonymised client names to "Client A," "Client B" and so on, so I wasn't tempted to defend anyone I liked.
- Added a column for actual hours worked, pulled from Toggl, next to the hours I had quoted for.
- Pasted the data into Claude in chunks, with the prompt: "Here is fee data and time data for a UK marketing consultancy. Identify which project types and fee bands show the biggest gap between quoted hours and actual hours, and flag any patterns by month or client size."
- Cross-checked the flagged patterns against my own memory of each project, because the AI has no idea which clients were a nightmare and which were a joy.
The whole thing took about six hours across two evenings. Most of that was cleaning the data, not analysing it.
The numbers that made me wince
My average project fee had dropped from £4,800 in 2021 to £2,300 by 2024. I knew things had been tighter, but I hadn't sat with the actual figure until it was on the screen in front of me.
The bigger discovery was in the retainers. Clients paying under £1,500 a month were taking up more of my actual hours than clients paying £3,000 a month or more. One retainer client, paying £900 a month, was generating an average of 11 hours of real work when I had originally quoted the package at 6. That's not a rounding error, that's a client I had been subsidising for over a year without noticing, because the invoice going out every month felt like income, and income feels like progress even when it isn't.
Across the whole data set, 40 percent of what started as a "quick call to help you out" turned into unbilled scope creep of six hours or more within three months. Every single one of those started with me saying yes to something small because I liked the person and didn't want to be the one who made it awkward.
The bit the AI wouldn't sugarcoat
Here's the part most posts about AI and pricing skip, because it's not a flattering story. I expected the analysis to tell me I was undercharging the market. Instead it kept pointing at the same pattern: I undercharged people I liked. The clients who paid full rate without argument were, more often than not, the ones I had less warmth for personally, the ones where the relationship stayed professional and slightly distant. The clients I discounted, extended deadlines for, and did "just one more thing" for free, were the ones I got on with best.
That is an uncomfortable thing to see written back to you by a chatbot with no stake in your feelings. It's also, I suspect, true for a lot of small business owners reading this. The advice everywhere is "know your worth" and "raise your prices," which is true but useless, because it treats pricing as a market problem. For a lot of us it's a people-pleasing problem wearing a spreadsheet as a disguise. I've read Dale Carnegie enough times to know the business lessons about liking people and being liked cut both ways: the same warmth that wins you clients is the exact thing that lets them quietly underpay you.
AI is very good at finding that pattern in cold data. It is no good at all at making the phone call where you tell someone you like that their retainer is going up by 60 percent in March. That conversation is still entirely on you.
How to run this on your own business this week
You don't need anything fancy. Here's what works, in order:
- Pull two to three years of invoices from whatever you use, Xero, QuickBooks, FreeAgent, even a folder of PDFs if that's what you've got.
- Put them in a simple spreadsheet with columns for date, client (use initials, not names, so you're not tempted to argue with the data), fee, project type, and quoted hours if you tracked them.
- If you have any time-tracking data at all, add real hours worked next to quoted hours. This one column is where the useful truth lives.
- Paste it into ChatGPT, Claude, or Gemini in manageable chunks and ask it directly: "Which fee bands and project types show the biggest gap between quoted time and real time? Which clients or client sizes are most profitable per hour?"
- Read the output with your own memory sitting right next to it. The AI knows the numbers. You know why Client B was worth keeping even at a loss, because they sent you three referrals worth £14,000.
If your bookkeeping is too messy to even get this far, that's a sign the problem is bigger than pricing, and it's worth getting proper help rather than guessing. That's the exact gap an AI implementation coach is meant to fill, someone who sits with your actual data rather than giving you a generic pricing framework off a slide deck.
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.
Where the AI got it wrong
It flagged one long-standing client as my least profitable relationship, based purely on hours versus fee. On paper it was right. In practice, that client is the reason two of my biggest projects last year existed at all, both from referrals she made without being asked. The AI had no way of knowing that, because referrals don't show up on an invoice. This is the limit worth remembering every time someone tells you AI will run your business for you: it reads what's on the page, not what the relationship is worth beyond the page. Treat the output as a very good first draft of the truth, not the whole truth.
What changed after I saw the numbers
I raised my minimum project fee to £3,200. I ended two retainers that were quietly costing me money every month, politely, with notice, and no drama. I built a simple pricing framework on the website with tiers instead of one bespoke quote for everyone, closer to the way Revolut built a value ladder that lets people self-select into the right level rather than negotiating from scratch every time. I also thought harder about positioning the higher tier as the calm, obvious choice rather than the expensive one, in the way Calm built simplicity into its whole brand instead of competing on features.
One small addition that's made the pricing conversation easier from the first email: I added a rough project calculator to the site so prospects can see a fee range before we even speak, rather than me having to be the one who first says a number out loud. If you've never built one, the logic behind interactive calculators for converting visitors applies just as well to pricing pages as it does to lead magnets. It takes the awkwardness out of the first conversation, because the number is already sitting there before anyone has to be brave about it.
None of this needed a big software rollout or weeks of consultancy. It needed one honest afternoon with a spreadsheet, an AI tool that doesn't care about your feelings, and the willingness to sit with a number I didn't like. That's the whole method. It's not glamorous. It works.
Frequently asked questions
What AI tool is best for analysing invoice and pricing data?
Claude and ChatGPT both handle this well because they're strong at spotting patterns across messy pasted data. Neither connects directly to your accounting software for this kind of ad hoc analysis, so you'll still need to export a clean CSV or spreadsheet first. Gemini works too if you already use Google Sheets, since you can run the analysis inside the sheet itself.
How much invoice history do I need before this is useful?
Twelve months will show you something, but two to three years shows you real patterns rather than one unusual quarter. If your business changed significantly in that time, a rebrand, a pivot, a big client leaving, note those dates so you don't mistake a one-off event for a permanent trend.
Is it safe to put client financial data into an AI tool?
Strip out client names, company names, and anything identifying before you paste data into any AI tool, and use initials instead. The numbers and patterns matter far more than knowing whose invoice each one was, and this protects you from any confidentiality issue with your client agreements.
What should I do if the AI flags a client as unprofitable?
Check the flag against what you know that the data can't see, referrals, reputation, how much the relationship costs you in stress versus what it earns you elsewhere. Use the AI's output as the start of a decision, not the decision itself. Then, if the numbers don't work, have the pricing conversation directly rather than letting the relationship quietly drain you for another year.
Related reading: How to Get More Views on Instagram Reels (What Moved the Needle for Me) and Why Instagram Reach Dropped and How to Fix It in 2026.
