You Wrote a Thousand AI Emails, and Not a Single Customer Replied — Why?

Let me tell you something that happened.

Last year, a friend of mine who works in SaaS complained to me: he'd used a large language model to batch-generate a thousand personalized marketing emails, each one customized based on the customer's industry, company size, and past purchase history.

He hit "send" with high hopes.

The result? A 12% open rate and a 0.3% reply rate. Worse than the mass emails he used to write by hand.

He came to me with the data: "I'm using the latest LLM, I did the personalization thing — why is nobody responding?"

Why?

Look, the problem isn't that AI isn't powerful enough. The problem is that his definition of "personalization" and the customer's definition of "personalization" are two completely different things.

1. What Is "True Personalization"?

A lot of people think too narrowly about "personalization."

They think — adding "Dear Mr. Zhang" at the start of an email and "Hope you have a great week" at the end is what personalization means.

Wrong.

That's "mail merge." Word could do that 20 years ago.

True personalization means the entire content, the entire timing, the entire channel is custom-built for "this one specific person."

Let me give you two real examples.

Amazon — in 2021, 35% of their sales came from their recommendation system. But their recommendations aren't the crude "people who bought A also bought B" collaborative filtering. Their recommendations look like this: if you bought maternity clothes last week, this week the homepage pushes strollers at you; if you searched for camping gear two weeks ago, today your homepage shows a "great weather for camping this weekend" gear bundle.

Netflix — 80% of their watch time comes from AI recommendations. But Netflix's recommendation isn't just "you like sci-fi." It's "you like to watch sci-fi shorts under 30 minutes after 9pm on Wednesdays, because you have work the next day."

See the difference?

True personalization knows who you are, knows what you want right now, and knows when to give it to you.

2. Generative AI Took This to a Whole New Level

Old AI personalization was essentially "classification" — slapping a "30-year-old male office worker" label on you, then pushing content tagged with that label.

Generative AI personalization is "creation" — generating a unique piece of content for you, and you alone.

These are fundamentally different things.

Here's the most direct example.

Old AI recommends products to an "outdoor enthusiast" — it pushes "hiking boots, hardshell jackets, tents" — because in the tag database, all of these are tagged "outdoor."

Generative AI recommends products to the same person — it pushes "for your camping trip to the Qinling Mountains this weekend, 5–15°C, chance of rain, your last pair of boots had a slippery sole — we recommend this pair of hiking boots with stronger grip, plus this lightweight rain jacket."

The former is tag matching. The latter is writing a paragraph for one person, on the spot.

L'Oréal, in 2023, used generative AI for content and saved 120,000 hours of manual labor, while SEO performance also improved. They didn't use AI to write generic copy — they generated a different landing page for every customer, for every search intent.

What does 120,000 hours mean? It's roughly a year of full-time work for 60 people.

Wow.

3. The Three Layers of Personalization "Magic"

Generative AI can exert its power across three layers in the customer journey.

Layer One: Content Generation

The most basic, and the most easily abused.

AI can batch-generate emails, product descriptions, social media copy, landing page copy — everyone knows this.

But the gap between using it well and using it badly is worlds apart.

Companies that use it badly generate a thousand roughly-similar pieces of copy, mass-send them, and get flagged as spam by customers.

Companies that use it well generate a thousand completely different pieces of copy — each one tied to a specific customer's actual need.

What's the difference? It's in the "input" — what data you feed the AI determines the quality of what comes out. Feed it only the customer's "name and company," and it can only write a "Hello Mr. Zhang" opening. Feed it the customer's full journey data, and it can write genuinely personalized content.

Layer Two: Recommendation

Old recommendations were "product-product" associations. New recommendations are "person-scenario-product" matching.

For example, Amazon no longer pushes "people who bought A also bought B." It pushes "Mother's Day is coming up, your mom likes this kind of thing, and you bought her something at this time last year."

The granularity of recommendation has shifted from "category" to "scenario."

How big an impact can this kind of recommendation have?

• Coca-Cola's "Share a Coke" campaign, printing names on bottles — social media engagement jumped 8.7x, and sales rose 2%.
• Nike's Nike Fit uses AI to scan feet and recommend sizes — digital channel conversion rate rose 40%.
• Ferrari's car configurator uses an LLM to help customers customize every detail — the customization process got 20% faster.

These aren't sci-fi stories. They happened between 2023 and 2025.

Layer Three: Dynamic Interface

This is the coolest layer.

The same website shows different people different interfaces.

• A customer who often buys outdoor gear sees "great weekend for hiking" on the homepage.
• A customer who just moved sees "10 essentials for your new home" on the homepage.
• A customer scrolling the app at 2am sees "late-night calming series."

The entire UI rearranges itself for "the you in this moment."

The AI marketing industry is projected to reach a market size of $107.5 billion by 2028. This isn't because everyone's chasing a trend — it's because this stuff actually drives conversion rates up 50% and customer retention up 7%.

4. But 80% of Companies Die on "Data"

At this point, you might be tempted — "I need to get on generative AI too!"

Hold on.

I have to tell you a counterintuitive fact: the success or failure of generative AI is 90% about data and 10% about the model.

• 80% of companies have data scattered across "silos" in different systems — one copy in CRM, one in the customer service system, one in the e-commerce backend, one in social media — completely disconnected.
• With disconnected data, AI sees "five fragmented portraits of the same customer," and the content it generates is inevitably a Frankenstein mess.
• It's like handing a top-tier chef half a wilted carrot and a piece of moldy bread, then asking him to make a Michelin three-star meal. Even the best chef can't cook without ingredients.

So, before you buy any AI tool, answer three questions first:

1. Can I pull up a customer's complete journey data in under 30 seconds?
2. Is my customer data unified and managed in one place?
3. Can I delete all of a customer's data with one click (compliance requirement)?

80% of companies can't answer these three questions.

Then don't rush into generative AI. Go back and lay the foundation first.

5. Three Plain Execution Principles

If you're set on doing this, I'll give you three plain execution principles:

Principle 1: Start small. Don't try to "AI-ify everything" all at once.

Pick one specific customer journey (like the win-back email after cart abandonment) and do it thoroughly. Once one scenario works, copy it to the next. Don't try to AI-ify all customers, all channels, and all content at once — you'll die of complexity.

Principle 2: Human in the loop.

Generative AI will "hallucinate" and produce content that doesn't fit your brand voice. There must be humans reviewing, editing, and setting the rules. A real example: a financial company let AI automatically generate investment advice — and the AI recommended a high-risk product to a conservative client. That kind of mistake costs you the customer, instantly.

Principle 3: Measure the right things.

Don't just look at "open rate" and "click rate" — those are vanity metrics.

What you should really look at:

• How much did Customer Lifetime Value (LTV) go up?
• How much did conversion rate go up?
• How much did Customer Satisfaction (CSAT) go up?
• How much did customer acquisition cost come down?

These are what AI should actually deliver.

6. One Final Judgment

Having written this far, I want to give you a contrarian take.

Over the next three years, "knowing how to use AI" will no longer be a competitive advantage. "Knowing when NOT to use AI" will be.

AI drives the marginal cost of personalization toward zero, which means everyone will use it. When everyone's using it, "personalization" itself stops being a differentiator.

So where's the real differentiator?

It comes down to this — whether you can, in certain key moments, NOT use AI, and instead use real people, take it slow, and embrace imperfection.

Like when a customer complains for the first time — it's not AI customer service replying in seconds, it's a real person calling them back within 5 minutes. Like on a customer's birthday — it's not a gorgeous AI-generated card, it's a handwritten, ugly, but heartfelt postcard.

AI lets efficiency take off, but makes the human touch scarce.

Whoever can do both of these things well will truly win customers' hearts over the next three years.

Back to my SaaS friend.

I had him change his 1,000 AI emails into 50 — but for each one, after AI wrote it, a human reviewed it, edited it, and added something like "I noticed your company launched a new product last week — congratulations," a real human observation.

Two weeks later he told me: reply rate went from 0.3% to 4.7%.

15x.

Wow.

That's the power of AI + the human touch.