88% Are Using It, 6% Are Making Money: The AI Marketing Books of 10 Companies

I saw a number recently that really stung.

88%. That's the share of marketers now using AI tools in their day-to-day work.

Almost everyone's jumped in.

But guess how many are actually making money from their AI investments?

6%.

That's from McKinsey's 2025 "State of AI" report. In other words, 94% of companies spent the money and got nothing back.

Where does this gap come from? I dug through the companies actually getting results, and I found something. The ones making money are doing something fundamentally different from the ones who aren't.

The ones not making money are mostly using AI to write copy, generate images, and blast emails. Faster, sure. But nothing fundamentally changed.

The ones making money are asking a different kind of question: Which customer is going to leave next month? Which of these leads will actually convert? How much inventory should we stock next month?

The former is automation. The latter is prediction. Between them sits an entire chasm of ROI.

Let me open the books on 10 companies. Once you've seen them, you'll understand exactly where the gap is.

Starbucks: Getting Personalization Right Beats Opening New Stores

Let me start with one that made me go "wow."

Starbucks has an internal AI platform called Deep Brew. What it does sounds straightforward: it analyzes each member's purchase history, location, the weather, and whether there's anything happening locally — then pushes a coupon through the app at the exact moment you're most likely to want a drink.

That's it. One thing.

The result? Members who received these personalized pushes spent 3x more than those who didn't. App engagement in the US market jumped 23%. Member engagement in China climbed 35%. All told, AI drove a 30% lift in ROI globally, and per-customer transaction values rose 14%.

Even more jaw-dropping is the supply chain side. AI-optimized inventory and replenishment saved $125 million a year.

Think about that. A coffee brand captured this much incremental value just by getting "the right coupon, to the right person, at the right time." Compared to opening new stores or running ad campaigns, the bang-for-buck is absurd.

Sephora: Customer Lifetime Value Up 29% — Without Discounting

Beauty retail is fiercely competitive, and Sephora's playbook deserves its own section.

They built a system called Beauty OS that stitches together several things: predicting how much each customer will be worth over time (CLV, or Customer Lifetime Value), AI-powered product recommendations, virtual try-on, and skin diagnostics. The app, website, and physical stores all share the same data pipeline.

The virtual try-on feature alone: users who engaged with it converted at 3x the rate of those who didn't. Cross-category purchases jumped 47%.

But the headline number is this: across the entire membership program, customer lifetime value rose 29%.

What does 29% mean? It's not selling them one extra item. It's the same customer, over the entire arc of their relationship with the brand, contributing nearly 30% more revenue.

No discounting involved. Sephora uses CLV predictions to segment customers: who's worth investing in heavily, what messaging works for whom, where to shift the budget. In plain terms, they put their best resources where they matter most — instead of treating every customer the same.

On top of all that, AI also cut their content production costs by 38%.

Netflix: Saving $1 Billion a Year by "Knowing You're About to Leave"

Netflix is probably the most-cited AI marketing case study. But every time it comes up, people only talk about the recommendation algorithm and miss the part that actually makes money.

The recommendation algorithm is impressive, sure. It drives over 80% of viewing activity on the platform. But that's just the appetizer.

The real money is in churn prediction.

Netflix has a dedicated model watching every single subscriber, judging whether they're "about to cancel." The moment someone is flagged as high-risk, targeted retention actions kick in immediately. In the most recent round of targeted interventions for high-risk users, they cut cancellation rates by 6% within three months.

Their monthly churn rate holds steady at around 2.5%, which is remarkably low for the streaming industry.

Do the math: every subscriber who doesn't churn is one less acquisition cost to pay. Netflix's system saves an estimated $1 billion+ per year.

Here's a detail that's equally ruthless. Netflix uses AI to generate multiple versions of thumbnails for every title, then picks the one most likely to make you click, based on your viewing habits. That single move can boost engagement by up to 30%.

Even what they greenlight for production is guided by predictive models. Their original content success rate sits at around 93% — the industry average is roughly 35%.

Grammarly: Cutting 400 Bad Leads Down to 200 Good Ones

This company changed how I think about the power of predictive lead scoring.

Grammarly has 30 million daily active free users. The problem? With that many free users, the sales team couldn't tell who was genuinely ready to upgrade and who was just messing around.

Their old approach: push 400+ "leads" to sales every month. The result? A huge chunk were bots, or accounts that registered and never logged in again. Sales wasted enormous amounts of time on people who were never going to buy.

Then they brought in Salesforce Einstein. The AI analyzed each free user's activity patterns — which features they used, what types of documents they edited, whether they collaborated with others. Then it scored every account. And here's the key: this score didn't measure "how active are they." It measured "how likely are they to upgrade."

The impact was immediate.

Account upgrade rates jumped 80%. MQL (Marketing Qualified Lead) conversion rates rose 30%. The sales cycle shrank from 60-90 days down to roughly 30 days. Monthly leads pushed to sales dropped from 400+ to around 200 — but the quality more than doubled. Email unsubscribe rates held at just 0.04%, versus an industry average of around 2%.

The person running this at Grammarly said something that nails it: "We went from chasing volume to chasing precision. Precision is worth far more than volume."

Progressive: One Predictive Feature, $2 Billion a Year in Premiums

What Progressive does is, hands down, the most stunning of the 10.

This insurance company has a telematics device called Snapshot that collects driving behavior data. Progressive took over 10 billion miles of driving data and trained a machine learning model on it.

The model does one very specific thing: it judges how strong a potential customer's intent to buy insurance is right now. At the moment of peak intent, a "Buy Now" button pops up in the mobile app.

That single feature generates $2 billion in new premiums every year.

The model identifies high-intent customers with roughly 90% accuracy.

Let that sink in for a second. This is fundamentally different from wedging an ad into someone's social feed. The AI calculates "this person is most likely to pay right now" — and at that exact moment, puts the purchase path right in front of them.

The value of a prediction isn't in generating a report. It's in triggering an action. Which channel, which person, what moment. Progressive took this to the extreme.

They later partnered with Claritas to use AI for optimizing ad creative. In controlled A/B tests, performance improved by 197%.

U.S. Bank: Lead Conversion Up 2.35x

When banks do AI marketing, the biggest bottleneck is usually the wall between departments. Wealth management, retail banking, mortgages — each runs its own show, and customer data doesn't flow between them.

U.S. Bank rolled out Salesforce Einstein across the entire bank. The AI scores and ranks leads in real time, and it does something many banks can't: cross-departmental sharing. A retail banking customer showing signals of loan interest is instantly visible to the mortgage team.

The result: lead-to-conversion ratios improved 2.35x. That's 235%. Manual lead screening workload dropped by over 60%.

They also use Einstein Discovery for churn prediction, and they did something I think is brilliant: they don't just tell the team "who might leave" — they explain "why."

A churn score sitting on a dashboard tells the team nothing about what to do. But "this customer is likely to leave because their service response times have slowed over the past three months, and a competitor has been in contact"? Now the team knows exactly how to intervene.

Stitch Fix: 75% of Outfit Selections Are AI-Driven

Stitch Fix is, at its core, a machine learning company that happens to sell clothes.

Their algorithms analyze your style preferences, purchase history, body measurements, seasonal trends, and return behavior — then predict the probability that you'll keep each item in the box they send you.

About 75% of box selections are now driven by AI.

The financials are solid: AI-driven selection pushed average order value up 9% year-over-year, and it's been climbing for seven consecutive quarters. Revenue per active customer rose to $542, up 3.2% year-over-year.

They also acquired two companies — one doing CLV prediction, the other demand sensing — and integrated both into their styling engine. Later they added virtual try-on and an AI style assistant.

In Q3 of fiscal year 2025, they returned to growth.

L'Oreal: Virtual Try-On Surpasses 100 Million Uses a Year

L'Oreal is the world's largest cosmetics company. But you may not know they also employ over 2,000 IT and beauty tech experts, plus 800 data analysts.

Their virtual try-on technology, ModiFace, is now used more than 100 million times a year — a 150% year-over-year increase. Users who engage with the feature convert at 3x the baseline rate. Marketing campaigns with AR (augmented reality) experiences see conversion lifts of 20% to 80%. After deploying ModiFace in physical stores, sales in related categories rose about 30%.

On the media side, they have a tool called Tidal that automates ad buying optimization. In a Nordic pilot, media efficiency improved 22% and ad performance improved 14%.

Then there's TrendSpotter, an AI that analyzes over 3,500 information sources to predict beauty trends 6 to 18 months in advance. Those predictions feed simultaneously into product development and marketing strategy.

Notice: L'Oreal isn't using AI as a single tool. They've built a web. Try-on, recommendations, ad buying, trend prediction — every node feeds data to the others.

Zara: 85% of Products Sold at Full Price

In fashion retail, on average only 60% of merchandise sells at full price. The rest gets marked down for clearance.

Zara's number is 85%.

That 25-percentage-point gap is worth billions in profit.

How? Zara uses RFID across 6,000+ stores to collect real-time sales data, layered with customer behavior analysis and social media trend signals — then predicts demand at the SKU level for every product.

From design to shelf, Zara takes just 10 to 15 days. The industry average is 3 to 6 months.

MIT research confirmed that this system reduced forecast error from 21% to 17%, cut potential lost sales by 24%, and lowered inventory management costs by about 20%. Zara stores see 17 customer visits per year — almost unmatched in fashion retail.

For marketing teams, the takeaway is this: demand forecasting isn't just a supply chain concern. If you can predict what customers will want, your promotions get sharper, your discounting shrinks, and your margins get fatter.

Amazon: 35% of Revenue Comes from Recommendations

Finally, Amazon.

Their recommendation engine processes over 150 billion customer behavior data points every day, covering more than 600 million products. This engine drives roughly 35% of Amazon's total revenue.

At Amazon's scale, that's over $70 billion.

Users who engage with recommendations have average order values 31% higher. Amazon's bounce rate is about 35%, compared to 50% for Walmart and 45% for Target. Recommendations on the checkout page reduce cart abandonment by 4.35%.

Amazon's approach is blunt and direct: homepage, product pages, emails, Alexa — every single touchpoint gives you a different, personalized experience.

This kind of infrastructure isn't something most companies can replicate. But the underlying principle applies to everyone: when you can accurately predict what a customer wants next, everything else gets easier.

What These 10 Companies Got Right

Looking across all these books, I noticed that the ones making money share four things.

First, they ask "what will happen next," not "what already happened."

Every single one uses AI for prediction. Will this customer churn? Will this lead convert? What will demand be next month? Reports can wait. Predictions can't. Gartner's early-2026 assessment: by 2028, 60% of brands will use AI agents for one-to-one customer interactions. First movers feast; latecomers get the scraps.

Second, predictions must connect to actions.

A churn score sitting on a dashboard won't reduce churn by a single dollar. Progressive pops a buy button at the moment of peak intent. Netflix auto-triggers retention campaigns. Sephora reallocates marketing budgets based on CLV. The value of a prediction equals the value of the action it triggers.

Third, the business problem comes first. Then the technology.

None of these 10 companies started by evaluating tools and then figuring out what to do with them. They all started with a specific problem — "how do we retain more customers," "who should sales prioritize" — and then found the technology to solve it. Gartner's 2025 CMO survey backs this up: organizations getting results from AI are the ones tying AI to specific business outcomes. Nobody's running AI experiments for fun.

Fourth, they restructured their processes instead of patching old ones.

McKinsey calls that 6% who are actually making money "AI outperformers." These companies are 3x more likely to have restructured their workflows. They don't treat AI as a plug-in. They redesign entire processes around what AI can do.

In Plain Terms

Each of these 10 companies spent years and invested tens of millions — sometimes hundreds of millions — to build their predictive capabilities.

Until recently, only they could afford to play that game.

That's the problem Pecan set out to solve. They make it possible for marketing, operations, and growth teams without a data science function to ask a business question in plain language and get back a prediction they can act on. No code. No waiting for a data scientist's availability. No external consultants.

Their customers' average results: churn down 12%, CLV up 10%, ROAS (Return on Ad Spend) up 15%, and models go live 32x faster than the traditional approach.

The story across all 10 companies is really the same story: whoever gets predictive capabilities into the hands of frontline decision-makers first will be the one to capture what the other 94% of companies still haven't.

As for when that 94% catches up — that's on them.