Marketers Want to Use Generative AI? Think Through These Four Questions First

A little while ago, I had coffee with the CMO of an FMCG brand.

She asked me one question: "Should our company get on board with generative AI tools?"

I asked her: what do you want to use it for?

She thought for a moment. "Write Xiaohongshu (China's lifestyle social platform) posts? Plan campaigns? Train customer service?"

I smiled.

This is the most common confusion hanging over marketing in 2026. Generative AI is genuinely on fire — but the vast majority of companies have no idea why they're buying it.

It wasn't until I read a study in a marketing academic journal, sat down with 10 executives for interviews, and pulled together a very clear analytical framework that I figured out how to answer her.

Today I'm going to walk you through that framework.

First, get clear: which "AI" are you talking about?

There's a common misunderstanding in marketing circles — calling every kind of AI "generative AI."

In fact, AI comes in two flavors, with very different goals, mechanisms, and use cases.

One is called Analytical AI. Its job is "prediction." Feed it a hundred thousand customer purchase records, and it can tell you who's most likely to repeat-purchase next month, which price point converts best, which ad has the highest ROI. Music apps recommending playlists, banks deciding whether to approve your loan — that's all its work.

The other is called Generative AI (Gen AI). Its job is "creation." Give it one sentence, and it hands you back a piece of copy, an image, a video. A global beverage brand using AI to produce ad posters, copywriters using AI to draft the first version — that's its lane.

They sound like the same thing, but they run on completely different logic.

Analytical AI eats structured data (numbers, tables) and outputs prediction scores. It's accurate, stable, and explainable — but it requires your company to have clean data of its own.

Generative AI eats unstructured data (text, images) and outputs new content. It's flexible, plug-and-play, and doesn't need your own data — but it will make things up with a straight face. The industry calls this "hallucination."

Which kind of AI you want depends entirely on what problem you're trying to solve.

Here's the framework: two questions, four quadrants

That scholar proposed a very practical two-dimensional framework. I'll translate it into plain English.

Dimension one: is the data fed to the AI "general-purpose" or "custom"?

General-purpose — for example, the conversational AI products that are everywhere on the market. It has seen all the public content on the internet; it can chat about anything.

Custom — for example, a finance-domain large model (LLM) trained by an international financial data company. It has only studied finance content; ask it for a recipe and it's lost, but ask it for an earnings report analysis and it outperforms the general model.

Dimension two: does the AI's output need to be reviewed by a human?

No — AI-generated content goes directly to the user.

Yes — AI-generated content is first reviewed and revised by an employee before being sent out.

Cross the two dimensions, and you get four quadrants.

I call them the fast, slow, steady, expensive playbooks.

Four quadrants, four playbooks

Quadrant one: Fast.

General-purpose model + no review.

Typical scenarios: have conversational AI summarize customer reviews, organize meeting notes, list competitor key points.

This usage is the cheapest and the fastest — and the most likely to blow up in your face. Because nobody's watching, the moment AI fabricates something, it lands right in front of the customer.

Good for internal tasks where being wrong doesn't really matter.

Quadrant two: Steady.

General-purpose model + mandatory review.

Typical scenarios: use conversational AI to write Xiaohongshu posts, draft campaign plans, then have someone from marketing revise them before publishing.

This is how 80% of companies are playing today. AI lifts efficiency; humans catch the falls.

Expensive? More than quadrant one, because you have to pay people. But far cheaper than a custom model.

Quadrant three: Private.

Custom model + no review.

Typical scenario: a supermarket chain's app. You open it and ask "where are the canned goods," and it tells you the aisle number and shelf position. That kind of Q&A can't use a general model, because the general model doesn't know your store's specific layout. So you have to train it on the supermarket's own inventory data. But if it answers wrong, it's no big deal — the customer grumbles at most.

The key to this usage is having your own data.

Quadrant four: Expensive.

Custom model + mandatory review.

Typical scenario: use a finance-domain large model to draft securities disclosure filings, then have legal and the CFO scrutinize every single word.

This is the most expensive playbook — and the most solid. The AI absorbs your domain knowledge; humans sign off one more time. Any scenario where an error carries an extreme cost has to sit in this tier.

How to choose? It comes down to your "fault tolerance"

I told that CMO: which quadrant to pick really comes down to a single question —

"If the AI gets it wrong this time, how much damage can you absorb?"

Low cost (you get an internal weekly report wrong — fix it and move on) → pick quadrant one, optimize for fast and cheap.

Medium cost (an official social media post goes out wrong, damaging the brand) → pick quadrant two, run a human review.

High cost (a commitment to a customer is wrong and triggers compensation) → pick quadrant three or four, bring in a custom model.

See — this isn't a technology decision. It's a business decision.

Many companies want to train their own large model from day one, when their actual use case would be more than covered by a general-purpose conversational AI plus an intern reviewing drafts.

And plenty of companies, to save money, shove everything onto a general-purpose model — and one piece of AI-generated misinformation knocks 5% off a listed company's stock price.

Spend where your fault tolerance sits.

Then there are the minefields you can't avoid

After picking the model, I have to cover four minefields. These all actually happened in marketing circles between 2024 and 2026.

First, copyright.

A certain AI illustration tool was caught producing images that sometimes closely resembled a particular artist's work. If you use that kind of image in a brand campaign, the artist can sue you.

The courts are still fighting these cases — but as a brand owner, don't treat legal risk as a luck problem.

Second, bias.

Research has found that different companies' large models actually have different political leanings. One leans left, another leans right. If your target audience is a neutral-to-conservative middle class, and your AI copy keeps skewing extreme, conversion rates will quietly drop — and you won't be able to find out why.

Third, privacy.

Feeding customer data into a general-purpose model is equivalent to handing your most valuable asset to someone else. An executive at a Fortune 500 company told me their internal rule is that no customer PII (personally identifiable information) is allowed into a general-purpose large model. That's the floor.

Fourth, the black box.

Analytical AI can at least tell you "why I predicted this." Generative AI often can't explain why it wrote that sentence.

Regulators have started requiring high-risk AI systems to be explainable. The AI you bolt on today to save trouble may be next year's compliance minefield.

A final note

My CMO friend went quiet for a while after hearing all this.

She said: "So should our company hold off for now?"

I said: "Not 'hold off' — 'think it through, then get on board.'"

With generative AI, the real question was never "to use or not to use." It's "where to use it, how to use it, and who reviews it."

Think of it as a razor-sharp knife. A chef uses it to plate a Michelin-star dish; a beginner uses it to cut themselves.

Four quadrants, four minefields. Think those two questions through first, then decide whether to buy that knife.

Technology is never the answer — clear judgment is.

May your choices match your insight.