A few years ago a consumer fintech gave me open-ended access to its transaction history: look around, tell us what you find.

The revenue distribution came back with something odd. A small percentage of accounts were producing a transaction volume no ordinary consumer produces. A first pass at the data led to better questions, so I kept digging. The data revealed a segment nobody knew existed: entrepreneurs reselling the company's service through their own accounts, recruiting their own customers and charging for the convenience. The company had a small-business channel living inside its consumer base and didn't know it.

I presented the findings to the board. The board asked for a progress report the following quarter. The program built around that segment became one of the company's better growth stories.

That was a month of specialist work, and almost none of it was the analysis anyone remembers.

First I had to get the transaction history out and join it to the marketing attribution data. Then came the part with no glamour in it: going variable by variable to see how each one was distributed, what the categories actually contained, and how those variables moved against each other. Only after that could I run the whole thing through the frameworks I'd built up over the years for growth accounting, borrowed in part from Jonathan Hsu's work: cohort analysis, growth efficiency for users and for revenue, revenue distribution, engagement.

That sequence is standard practice. It gives me enough orientation to find the insight later..

What it buys is the right to start asking real questions. I'd write an ad hoc query, look at what came back, and let it point me at the next one, over and over, until I had something I believed. Then more work again to turn what I'd found into something a board could take in at a glance.

That's the loop: analysis begets analysis. Every pass leads to a better question than the one you started with, and the good questions sit downstream of a lot of plain, unremarkable work. Having run that sequence hundreds of times is what let me tell a real finding from an artifact of the data.

AI changes the economics of that pre-work. Much of it can now be compressed. What remains is judgment and the quality of your questions.

That's the change I want to talk about, and it starts with the test I posed last week: everyone at your leadership meeting writes the company's customer count on an index card before anyone speaks, and then you read the cards out.

So, did the numbers match?

If they didn't, don't fret, because this piece ends where your problem starts. If they did, you're standing on something rarer than you might think, and the question becomes what that trusted count is for.

The count was never the prize. It's the ticket. The interesting questions are behind it: which of those customers are worth acquiring more of? Which ones are carrying the business? Which are about to leave, and which would respond if you gave them a reason to stay?

The discipline that answers those questions is customer analytics.

The cleanest way I know to explain it to someone who's never heard the term: the value of your business is the sum of the value of your customers, the ones you have plus the ones you'll acquire. At the risk of sounding impersonal, when breaking down a business, I think of each customer as a bond. You pay something to acquire it, and it pays off over some period of time. The analogy works even better in groups: these hundred customers cost $10K to acquire and will pay back $40K over their lifetime.

Your company is that portfolio of bonds.

That's a different animal from the P&L. The P&L is the top-down picture, revenue at the top and costs beneath. What it doesn't do well is break the business into the parts that produce that picture. Customer analytics gives the bottom-up version of the same company: the small group producing most of the revenue, the new customers who may or may not turn into loyal ones, the customers who refer other customers, the occasional spenders who show up twice a year.

The health of the business is how those groups are distributed and which direction each one is moving.

The discipline isn't new, least of all from me. I spent six years inside a venture investment firm reading the transaction histories of hundreds of companies, and in May 2024 I wrote the whole thing up in a piece called "Look at the Stars, Look How They Shine for You" (https://www.linkedin.com/pulse/look-stars-how-shine-you-david-smith-u7cje/). I would make the argument again, with one exception.

The exception matters.

That article has a section called Costs and Risks, and its job was to slow you down: check whether your data is even worth mining before you spend real money mining it. I quoted Benn Stancil's description of the problem, still the best one I know: a mid-sized company's data is not an oil field, it's a peat bog, and no matter how good the tools get, "it'll never burn as hot or as bright."

In 2024 I meant that as a reason for many companies to think hard before investing in customer analytics. I was right to say it.

Two years later I need to revise my own advice, because both halves of the ROI calculation moved: what the data is worth, and what it costs to get the answers out.

The first move is the one I didn't expect: the peat bog got richer.

I assumed the value locked in a company's data was fixed, and the only question was the cost of getting it out. But your data has learned a second job. It used to describe the business to you. Now, more than ever before, it can also run parts of the business, feeding AI-driven systems that act on it directly instead of waiting for a person to read a report.

What it takes to trust data with that job is next week's subject. Today's point is smaller: the same records are worth more than they were when a report was the only thing they could become.

The second move is the one this piece is about: the analysis got cheaper.

The 2024 article named four costs that gate any customer-analytics effort. Grading my own homework two years later:

  • Security and compliance. Unchanged, if anything heavier. I would not revise this part.
  • Organizational alignment. Didn't move an inch, and be suspicious of anyone selling software that claims otherwise. Definitions and trade-offs still get made by the people running the company, because the numbers have to match how you've decided to run it. AI can't make those calls for you. It can run parts of the business; it can't decide what the business means by its own numbers.
  • Good data. Still the gate, and a big enough subject that it gets its own piece next week.
  • Technical skills. This is the one that collapsed. The cohort table, the lifetime-value model, the concentration analysis: each of those used to mean a dedicated analyst, which meant a hire, which meant the project started as a headcount conversation. A few weeks ago I watched a working dashboard (filters, big numbers across the top) get built in half an hour of plain-English conversation with an AI tool. That was a specialist's full day of work not long ago. It is also most of the month I opened with.

So the threshold moved. The company I'd have told to pass in 2024 looked like this: not that many customers, modest contract values, a few years of history at most, and a price of answers that started with hiring someone. For that company, the juice wasn't worth the squeeze, and I said so at the time.

Today there's more juice and much less squeezing. The bar dropped from "worth a headcount" to "worth an executive's attention for a few afternoons," and far more companies clear it than know they do.

Why they don't know it is the objection I hear in different words at different companies.

  • "You're getting too technical on me."
  • "We can't seem to get out of our own way on this."
  • "That's been a hard area for us to address, and frankly I don't see the benefit."

Different words, same sentence underneath: we cannot do any of this, because our data is a mess.

That objection is real, and it deserves a serious answer. It gets one next week.

I'm David Smith. I set AI strategy with leadership teams, build the data and systems underneath it, and ship the highest-leverage pieces myself. If you're looking at a version of this, I'd like to hear about it.