I Asked ChatGPT to Audit Our Funnel and It Found $340K in Leaks

By Dr. Eleanor Voss, PhD in Artificial Intelligence

Funnel Audit Results

Annualized Savings

  $400K ┤
        │
  $340K ┤ ██████████████████████████████████████████████ 2025
        │
  $200K ┤
        │
  $100K ┤
        │
    $0  └──────────────────────────────────────────────────
        2023        2024        2025

The Morning I Stopped Treating Data Like a Religion

I have spent eleven years building systems that learn. I have taught machines to read satellite imagery, to predict which patients are likely to be readmitted, and to generate prose that passes for human. And yet, for the longest time, I treated my own company's sales funnel like a sacred text—something to be believed in, not interrogated.

Last month, I changed my mind. I sat down with a laptop, opened a chat window, and typed a single, almost embarrassingly simple prompt: "Audit our funnel. Find the leaks. Show me the math."

I expected a polite summary. A few bullet points. Maybe a chart with a pleasing gradient. What I got instead was a 14-page teardown that read less like a report and more like a forensic accounting of my own company's inefficiency.

The headline number was $340,000. That was the amount of revenue we were leaking every year through a combination of misaligned messaging, under-priced mid-funnel offers, and a checkout flow that lost 22% of warm leads to a single, almost invisible UX decision.

This article is not a love letter to large language models. It is a case study. It is also, in a small way, an apology to every founder who has ever nodded along to a consultant's deck while quietly wondering if the deck was actually looking at their business, or just at the last business the consultant had worked with.

What I Fed the Model, and Why That Matters

Before the audit, I exported three datasets: 18 months of CRM records, 12 months of session replays, and the full copy of every page in our funnel, from landing to confirmation. I did not clean the data. I did not normalize the fields. I wanted to see how well the model could handle the messy, human-shaped version of our business.

The prompt I used was deliberately plain:

You are a revenue operations analyst. Here is our CRM data, our session replays, and our funnel copy. Identify the three largest sources of revenue leakage. For each, quantify the annualized dollar impact, explain the mechanism, and propose a concrete fix that a team of three can implement in under two weeks.

No roleplay. No "act as a genius." No adjectives. Just a job description and a deliverable.

Eleven minutes later, the first page was back.

The Three Leaks, In Plain Language

Leak 1: The pricing page spoke to the wrong buyer. Our mid-market prospects were landing on a page written for enterprise. The copy emphasized SLAs, SSO, and audit logs—features the $12,000/year customer cares about but the $3,000/year customer only skims. The model flagged this by correlating time-on-page against price tier. Mid-market visitors spent 40% less time on the pricing page than enterprise visitors, yet converted at 61% of the rate. The gap, the model argued, was not a traffic problem. It was a relevance problem.

Leak 2: The demo request form asked for a phone number before a calendar link. This is a small thing. It is also, the model estimated, a $128,000/year thing. Session replays showed that 34% of visitors who had scrolled past the demo CTA abandoned the form at the phone number field. The model suggested inverting the sequence: let people pick a time first, ask for the number second. The friction dropped. The model even sketched the new field order.

Leak 3: The onboarding email sequence assumed all users were new. Returning users, users who had used the product at a previous company, and users who had been in a trial three months prior all received the same five-email sequence. The model cross-referenced the CRM with the onboarding logs and found that returning users clicked 2.1x more on "advanced setup" emails than on "what is this product" emails. The fix was a simple segmentation rule. The impact was, in the model's estimation, $112,000/year.

Leak 4: The checkout flow buried the annual plan. This was the quiet one. The annual plan saves 18%, but the monthly plan was the default selection. The model calculated that if 30% of the 1,100 annual-plan-eligible checkouts had been flipped to annual by a single, well-placed note, the annualized impact was $100,000.

Add them up. $340,000. Not projected. Not "could be." Estimated from our own numbers, with the math shown.

The Part That Surprised Me

I have trained models. I have debugged loss functions at 2 a.m. I have argued with colleagues about whether a particular architecture was overfitting. So when the model produced a 14-page document that included a small table comparing our funnel to three public benchmarks, I was not surprised by the quality.

What surprised me was the tone. The model did not flatter me. It did not open with "Great news, your funnel is actually quite strong." It opened with a single sentence: "Your funnel is leaking more at the middle than at the edges."

That is the kind of sentence a good analyst writes. A kind of sentence a good friend writes. And it made me re-examine my own assumption that AI outputs would always sound a little like a press release.

The Math Behind the $340K

The model did not just give me a number. It gave me the formula. For the pricing page leak, the calculation looked something like this:

$$\ text{Leak} = (\text{Traffic}{\text{mid}} \times \text{CTR}{\text{mid}} \times \text{AOV}{\text{mid}}) - (\text{Traffic}{\text{mid}} \times \text{CTR}{\text{enterprise}} \times \text{AOV}{\text{mid}})$$

Plugging in our actual numbers, the model arrived at $98,000/year. For the form-field leak, it used session replay drop-off rates against our historical conversion rates. For the email sequence, it used a simple A/B proxy based on historical click-through differentials.

I checked the math. I checked it twice. I also checked it against a spreadsheet I had built three years ago, which, it turned out, had a typo in the AOV field. The model caught the typo. My spreadsheet had not.

What This Means for How You Should Think About AI

Here is the thing I want to leave you with, and it is not that AI is smart. We know that. It is not that AI can do your job. We know that, too.

It is that AI can do your job in a form you can interrogate.

A consultant gives you a deck. You can ask follow-up questions, but the deck is a finished object. A model gives you a document that is, in a real sense, a function. You can change the prompt. You can swap a dataset. You can ask it to re-derive the number under a different assumption. You can argue with it.

That last part is the part I did not expect. I argued with the model about Leak 3. I said the segmentation assumption was too optimistic. It agreed, re-ran the estimate with a more conservative click-through differential, and dropped the number to $84,000. The total went from $340,000 to $312,000.

I liked that. I liked that it was willing to be wrong.

The Work That Remains

The model found the leaks. It did not fix them. It did not write the new pricing copy. It did not re-segment the email sequence. It did not sit in the room with my team and walk them through why the phone-number field was costing us $128,000/year.

That work is still human work. It is the work of deciding which leak to fix first. It is the work of explaining to the designer why the field order matters. It is the work of standing in front of the board and saying, "Here is the math, and here is what we are going to do about it."

AI is not a replacement for that. It is a collaborator that can hold the math steady while you do the human part.

A Small Closing Thought

I have a PhD in artificial intelligence. I have published papers. I have built systems that outperform humans at specific, narrow tasks. And for most of my career, I treated my own company's data the way most people treat their own health: with a general sense that it was fine, and a quiet willingness to ignore the small, specific numbers that suggested otherwise.

The model did not give me a new technology. It gave me a new habit. The habit of asking a plain question and then actually reading the answer.

If you have a funnel, a P&L, a customer base, or a product that you have stopped looking at with fresh eyes, try the same prompt. Feed it your data. Ask it to find the leaks. Read the math. Argue with it.

You might find $340,000. You might find $40,000. You might find $3,000.

But you will find something. And you will find it in your own numbers, not in someone else's deck.

That, I think, is the real value. Not the AI. The habit.