I Fed a Year of My Invoices Into AI. It Told Me Three Clients Were Costing Me Money.

The short version: Export your invoices and your logged hours, feed both into ChatGPT or Claude, and ask it to calculate your effective hourly rate per client. Most small business owners have never done this and are quietly working for less than they'd pay someone else to do their job. It took me one wet Tuesday afternoon to find three clients who were, on paper, losing me money every single month.

What happened when I finally ran the numbers

I'd been telling myself a story for about eighteen months. The story was: my consulting work is going well, I'm busy, the calendar is full, therefore the business is healthy. Busy and healthy are not the same thing, and I knew that in theory, but I hadn't checked it in practice for one simple reason: I didn't want to know.

Then, on a slow afternoon in late 2025, I pulled twelve months of invoices out of FreeAgent as a CSV, pulled twelve months of logged hours out of Toggl, and dropped both files into ChatGPT with a fairly blunt prompt: work out my effective hourly rate for each client, in order, lowest to highest.

The output took about forty seconds. Reading it took a lot longer, because I had to keep stopping.

Here's what it found, with client names swapped for letters:

• Client A: paid me £14,400 across the year, logged at 172 hours. Effective rate: £83.72 an hour.
• Client B: paid £6,000, logged at 61 hours. Effective rate: £98.36 an hour.
• Client C: paid £21,000, logged at 108 hours. Effective rate: £194.44 an hour.
• Client D: paid £9,600, logged at 96 hours. Effective rate: £100 an hour, exactly.

My target rate, the number I use to price new work, is £150 an hour. Client A, who I'd have described as one of my "good, easy" clients, was running at just over half that. Client D was sitting right on the edge. Both of them were clients I liked personally, which is exactly why I'd never checked the maths.

Why almost nobody runs this number

I've asked around since. I put a version of this question to a room of about forty small business owners at a workshop in Manchester last autumn: who has calculated their actual hourly rate, per client, in the last twelve months? Two hands went up. Two, out of forty people who all run businesses that live or die on their time.

The reason isn't laziness. It's that the answer might be bad news, and most of us would rather stay busy than find out the busyness isn't paying. There's also a quieter reason: this kind of audit used to take a proper afternoon with a spreadsheet, formulas, and a headache. Now it takes about fifteen minutes, which removes the excuse entirely.

I write a fair bit about using AI without needing to be technical, and this is one of the clearest examples I've hit on. You don't need a formula. You don't need a pivot table. You need two exports and one plain-English question.

There's also a psychological trap worth naming here: revenue feels like proof of success, so a client paying you £14,400 a year sounds great until you realise what it cost you to earn it. Turnover is vanity, effective hourly rate is closer to the truth.

How to run this audit yourself, step by step

You need two things before you touch any AI tool.

• Twelve months of invoices. Export from Xero, QuickBooks, FreeAgent, or wherever you invoice from, as a CSV. Client name, invoice date, amount, that's the minimum.
• Twelve months of hours, per client. If you use Toggl, Harvest, or Clockify, export the same period as a CSV. If you don't track time formally, use your calendar and your memory to estimate a rough figure per client per month. Rough is fine. Wrong by ten percent still tells you a true story.

Then, before you upload anything, strip out anything that would identify the client if the file ever leaked. Replace client names with codes. If you use a paid, business-tier account with ChatGPT or Claude your data generally isn't used for training, but I still don't put real names in. Habit, not paranoia.

The prompt I used, more or less word for word:

"Here are two CSVs: one is invoices paid per client over the last twelve months, the other is hours logged per client over the same period. Match them by client code. Calculate the effective hourly rate for each client (total paid divided by total hours). Rank clients from lowest to highest effective rate. Flag any client below £[your target rate] an hour. Then tell me which clients had the most billing activity relative to their rate, since those are costing me the most time for the least return."

That last sentence matters. A client at £70 an hour who takes four hours a year is a rounding error. A client at £70 an hour who takes 200 hours a year is a structural problem in your business.

Once you have the ranked list, ask a follow-up question the AI can't answer for you, because it doesn't know your business: which of these clients bring you referrals, testimonials, case study material, or work you enjoy enough that the money isn't the whole story? Write that answer yourself, on paper, before you decide anything.

What the AI got wrong, and why that's the important part

Here's the bit that most people writing about "AI for business insights" skip over, because it's less flattering to the tool: the AI told me Client A was underperforming, and it was right about the maths, but it had no idea Client A had referred me two other clients worth £30,000 between them. It couldn't know that, because I never told it, and it had no way of asking.

AI is excellent at arithmetic across a spreadsheet you'd never sit down and do by hand. It is useless at judgement, loyalty, reputation, or the fact that a client you've worked with for six years might be worth keeping at a lower rate simply because they trust you enough to say yes to bigger projects without a lengthy negotiation. It will happily hand you a ranked list and let you make the wrong decision faster than you would have made it on your own.

The uncomfortable bit isn't the AI's output. It's that most of us go looking for this kind of tool hoping it will make the decision for us, so we don't have to feel bad about firing a client we like. It won't. It just removes your last excuse for not knowing the number.

The three types of client this exposes

Once you've run the numbers, clients tend to sort into three groups.

• Underpriced but efficient. Low rate, low hours. Not urgent, but worth a renegotiation at the next contract renewal.
• Underpriced and heavy. Low rate, high hours. This is where your actual damage lives, because these clients quietly absorb months of your year at half what they're worth. This is where I found Client A.
• Scope creep in disguise. Fine rate on paper, but the hours have crept up over the year through "quick calls" and "one more small favour" that never got billed. This one's sneaky because the invoice looks healthy right up until you add up the unbilled hours sitting outside the system entirely.

That third category is the one I'd underestimated most. I had roughly nineteen hours across the year, spread across four clients, that never made it onto an invoice at all. Free calls, a favour rewrite, "just a quick look" at a landing page. Nineteen hours at £150 an hour is £2,850 I simply gave away, and I only found it because the AI flagged a gap between logged time and billed time that I'd have never spotted by eye.

What I did with the answer

I renegotiated with Client A, moved them from an hourly retainer to a fixed monthly fee that reflected the true scope, and their rate went from £84 an hour to £142 within one conversation. I let Client B go at contract renewal, not because I disliked the work, but because the effective rate had been sliding for two years and I'd been too polite to raise it. I kept Client D and simply started billing every call, including the "quick ones," which alone lifted their effective rate above target within three months.

I didn't touch Client C. Client C was already the model I should be selling to more people, and running this audit told me that too, which is its own kind of useful.

This is the same instinct I've written about before when I looked at what happened after my own website started paying me again: you can't fix what you haven't measured, and most of us are running our pricing on vibes rather than numbers. The vibes said Client A was easy money. The numbers said Client A was the most expensive relationship in my business.

If you'd rather not run this yourself, or you want someone to sit with you and work through pricing, positioning, and where AI earns its keep in your business, that's a fairly typical starting conversation with an AI consultant for small business. It's a two-hour job for someone who does it regularly, and it tends to pay for itself inside the first renegotiated contract.

One last caution, because I've written before about trusting AI output too readily after publishing 569 blog posts and watching Google index five percent of them: check the AI's arithmetic against your own invoice totals before you make a single decision based on it. It made a small error the first time I ran this, double-counting one invoice that had been part-refunded. I caught it because the total didn't match my bank statement. Always reconcile against something real.

Free resource: The Annual Contract Negotiation Swipe File.

Frequently asked questions

How do I get my invoice data ready for an AI audit?

Export twelve months of invoices from your accounting software (Xero, QuickBooks, and FreeAgent all do this in a couple of clicks under reports or invoices) as a CSV with client name, date, and amount. Replace client names with codes before uploading anywhere, and match that against a CSV of hours logged per client from your time tracker or your calendar.

Is it safe to put client financial data into ChatGPT or Claude?

Use a paid business-tier account, which generally isn't used for model training, and strip out anything identifying, such as real client names or contact details, replacing them with codes like Client A and Client B. That gives you the useful maths without any real data leaving your control in a recognisable form.

What if I've never tracked my hours ?

Estimate. Go through your calendar for the past twelve months and rough out hours per client per month, rounding to the nearest half hour. Being ten or fifteen percent off still surfaces the true pattern, because the gaps between clients are usually far bigger than your margin of error.

How often should I run this audit?

Twice a year is enough for most small businesses, once at the halfway point and once before annual contract renewals, so you're pricing your next year on real numbers rather than on how busy last year felt.

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.