Marketers: How Should You Actually Be Using AI?

A while ago, I ran into a friend who's been in marketing for over a decade.

He told me his company had just rolled out ChatGPT. The marketing team was using it to write WeChat Official Account posts, Xiaohongshu (RED, China's leading lifestyle social platform) posts, and ad copy — they were having a field day. And the output? At first glance, it looked pretty decent.

But then he asked one question:

"After you published those pieces, did the conversion rate go up?"

The whole office went silent.

Nobody had an answer.

That set off a long chain of thinking for me. Look — Generative AI (Gen AI) has been taking the marketing world by storm over the past couple of years. Everywhere you turn, there's a story about "10x efficiency gains." But the people who've actually figured out how to use it correctly? Precious few.

That's what I want to talk about today.

Let Me Show You Some Numbers First

Vanguard used Generative AI to write ad copy and ran it on LinkedIn. The result? Conversion rates jumped 15%.

Emirates NBD used AI for personalized credit card recommendations. Lead volume surged 177%. You read that right — 177%.

Unilever used AI to analyze customer messages — not just what customers said, but the emotions behind their words. Customer service response time was slashed by nearly 90%.

Walmart went even further. They built an AI negotiation bot specifically to negotiate prices with suppliers. The result? Costs dropped 3%, and most suppliers said they actually preferred negotiating with the bot.

Bloomberg predicts this market will exceed $1.3 trillion by 2032.

Tell me — can marketers afford to sit still?

But There Are Two Types of AI. Can You Tell Them Apart?

Let me clear up something fundamental first, because a lot of people genuinely haven't grasped the difference.

What is Analytical AI?

It looks at past data and helps you predict the future. Banks use it to assess whether a customer will default on a loan. Spotify uses it to recommend playlists. E-commerce platforms use it to guess what you'll buy next. Its core capability is prediction.

And Generative AI?

It doesn't predict. It creates.

Give it a prompt, and it can write a blog post, paint a picture, edit a video, or generate a code snippet. It brings something into existence from nothing.

Think about it — could these two types of AI possibly have the same impact on marketing?

Analytical AI helps you do existing things more accurately. Generative AI helps you create entirely new things. In a field like marketing, where content and communication are the lifeblood, the latter's disruptive power is obviously far greater.

Amazon is already using Generative AI to write summaries of product reviews. Coca-Cola has publicly stated that their future marketing roadmap is a fusion of "AI and human creativity."

Good grief — if even Coca-Cola is being this aggressive, how can you sit on your hands?

Two Choices That Determine Whether AI Makes or Costs You Money

Alright, enough excitement. Let's get into something truly useful.

I've been studying how various companies are putting Generative AI into practice, and I've found that success or failure really comes down to two choices.

Choice One: What data do you feed it?

Tools like ChatGPT run on general-purpose large language models. They've ingested nearly all publicly available information on the internet — they know a little bit about everything. Ask them to draft a social media post or whip up an ad concept, and they'll do it fast and well.

But if you're Walmart and you need to tell a customer which aisle a product is on, when an out-of-stock item will be replenished, or what alternatives are available — ChatGPT can't help you. That information lives inside Walmart's own stores. It's not on the public internet.

That's when you need customized inputs. You feed the model your company's own data.

Morgan Stanley took a middle path. They used a technique called RAG (Retrieval-Augmented Generation) to essentially "tutor" a general-purpose large model using their own internal documents. This way, the output had both breadth and specialized depth.

Choice Two: Should a human review the AI's output?

This choice is more subtle — and more critical.

A Fortune 500 company's sales team used AI to generate proposal drafts. After the drafts were written? The sales team would review every single line, revise, personalize, and only then send it to the client. The human is the last line of defense.

But Amazon's AI-generated review summaries go live without any human eyes on them at all.

Why the stark difference?

Because the cost of getting it wrong is different.

If a review summary is slightly off, the worst that happens is a few downvotes. But if a sales proposal has errors? You could lose a million-dollar deal and damage your brand.

So here's the principle: the higher the cost of error, the deeper the human involvement needs to be.

Four Quadrants — Find Your Seat

Cross-reference those two choices, and you get four quadrants. Let me break each one down.

Quadrant One: General Data + Low Human Intervention.

Fast and cheap, but also the riskiest. Best suited for scenarios where the cost of errors is low. For example, using ChatGPT for internal brainstorming, having employees do quick research, or generating summaries. If something's wrong, no big deal — it never leaves the building.

Quadrant Two: General Data + High Human Intervention.

Fast, but with human guardrails. Suited for scenarios that require broad information but can't afford mistakes. The classic example is social media content. You don't need proprietary company data to write a social post — but you absolutely need a human to review it before publishing, because one ill-considered post can trigger a PR disaster.

Quadrant Three: Custom Data + Low Human Intervention.

Fast and precise — provided your data quality is good enough. The retail AI that helps Instacart shoppers locate products on store shelves falls into this category. Store data is updated in real time, so the AI's answers are almost never wrong — no human intervention needed.

But this quadrant is the hardest to implement. You have to integrate your internal systems with the AI and ensure the data is continuously updated. Many companies can't pull this off.

Quadrant Four: Custom Data + High Human Intervention.

The slowest, the most expensive, but also the safest. Bloomberg GPT helping financial firms draft SEC (U.S. Securities and Exchange Commission) filings falls here. The data is financial and proprietary, and every draft gets human review — because if those documents contain errors, the penalties can be astronomical.

You see, choosing a quadrant isn't about picking the most advanced option. It's about understanding the nature of your task: How costly is an error? How precise does it need to be? Do you actually have the data?

Pick the wrong quadrant, and AI won't just fail to help you — it could actively harm you.

Enough About Benefits — Let's Talk About the Pitfalls

All those case studies I just mentioned are dazzling. But honestly, for marketers, the things really worth thinking hard about are the following traps.

Trap One: Hallucination.

Generative AI will make things up with total confidence. Ask it to write a product description, and it might invent features that don't exist. This isn't a bug — it's the fundamental nature of these models. They're probabilistic, not deterministic.

One study found that 68% of participants using AI-assisted writing submitted the AI's first draft without changing a single word.

Good grief.

That's like hiring an intern and signing off on everything they hand you without even glancing at it. Anyone who's spent time in the corporate world knows how dangerous that is.

Trap Two: Efficiency Is Not the Same as Effectiveness.

Someone ran a comparative test, pitting AI-generated ads against ones created by human creative teams.

The AI ads had three times the click-through rate of the human ones. Sounds like AI won, right?

But the human ads generated 9.5 times more qualified leads than the AI ads.

So who actually won?

High click-through rates might simply mean the AI wrote more clickbait-y headlines. But when it comes to content that actually converts prospects into customers, you still need the insight and creative depth that humans bring. That's the difference between efficiency (doing things fast) and effectiveness (doing things right).

BCG and Zeiss collaborated on a medical AI assistant that they claimed produced responses over 90% of which were "ready to go directly to patients." But what did they actually do? They still placed an optometrist in the middle — every response was reviewed before being forwarded.

Chew on that. Their words said 90%, but their actions said otherwise — they quietly slipped in a layer of human review.

Trap Three: Copyright.

DALL-E 3 and Midjourney can generate stunning images. But people have discovered that they sometimes reproduce copyrighted works from their training data almost verbatim — without telling you where the image came from.

If you use such an image in a commercial ad, a copyright lawsuit could come knocking. Whether they sue the AI company or you as the user, it's going to be a serious headache.

Trap Four: Bias.

A study tested the political leanings of 14 large language models. The findings? OpenAI's ChatGPT and GPT-4 leaned toward left-wing liberalism, while Meta's Llama leaned toward right-wing authoritarianism.

AI is not neutral. Its bias stems from training data and from the model's own weight design.

If your marketing content is generated by an AI that "comes with built-in bias," you could end up alienating a segment of your customers without even realizing it.

Trap Five: The Black Box.

Analytical AI models are relatively interpretable — you can roughly understand why a particular prediction was made. Generative AI? You ask it a question, it gives you a response, and you may have absolutely no idea why it said what it said.

In certain industries, this is fatal. If your AI recommendation system suggests an inappropriate product to a customer, and the regulator asks "why did you recommend this?" — you can't exactly say "I don't know, the AI just felt like it."

So, How Should You Actually Use It?

I've laid out all these pitfalls not to scare you away from AI. Quite the opposite — you have no choice but to use it.

Of the 300+ Chief Data Officers surveyed, more than three-quarters believe Generative AI will fundamentally transform their business environment. But at the organizational level, only about 20% are actually deploying it. The rest are still stuck at the stage of individual employees tinkering with ChatGPT.

In other words, most companies are still hovering at the doorstep. If you start thinking seriously about how to use it now, you're already ahead.

How to get ahead? I'd boil it down to three principles:

For low-risk tasks, let AI run free. Internal memos, data summaries, creative first drafts — the cost of getting these wrong is low, and AI is fast and good. Don't overthink it.

For high-risk tasks, put a human between AI and the customer. Review ads before they go live. Revise client proposals. Vet social media content before publishing. AI gets the ball rolling; humans finish the job.

For tasks that need proprietary data, get your data house in order first. Don't rush into building a custom large model. If your data is still messy, scattered across departments, and outdated, feeding it to an AI just means garbage in, garbage out.

The AI era has arrived. But no matter how powerful the tools are, the people wielding them are what truly matters.

Just like Walmart's negotiation bot. It did save Walmart money. But the real decision-maker was the person who decided "should we even use a bot to negotiate with suppliers?"

Tools are leverage. You are the fulcrum.