Making Content with AI: The Giants Stopped Typing in a Chat Box Long Ago

A while ago, a friend who works in marketing came to me to vent.

His company had everyone writing copy with AI. The output, as he put it, "tasted like dishwater": every line looks correct; every line puts you to sleep. His boss kept pressing him — why is everyone else's AI so magical while yours is so flat?

I asked him: how do you actually use AI?

He said, how else would you? Open ChatGPT, type in "write me a promo line for a yogurt," and then start pulling cards. Pull ten times, pick the one that looks decent.

That's exactly where the problem is.

You're treating AI like an intern who just types faster. But nothing about your company lives inside it. It doesn't know who your users are, and it doesn't know what voice your brand should speak in.

So I went and dug through how several consumer-goods giants do it. The strongest impression I came away with: everyone is using the same technology, but how they use it is already a generation apart.

Where's the gap?

Most people are "using" AI. The giants are "raising" it.

Out of a batch of cases from around 2023, I picked three companies to tell you about.

1. Unilever: Raising AI Under Its Own Roof

Unilever first.

How long has this company been in the brand business? In 1884, the Lever brothers gave their laundry soap a proper name and proper packaging — among the earliest merchants to take "branding" seriously. A hundred-plus years later, this company that started out selling laundry soap and mayonnaise is doing nothing that feels old.

It didn't have its employees all typing into public chat windows. Instead, it used OpenAI's models to build its own internal interface, with several GPT-3-based applications inside dedicated to writing marketing copy. Brand managers run experiments in there every single day.

One example. Around Thanksgiving, they ran an AI menu recommender: tell it what's in your fridge, and it puts recipes together for you. Search "what can I do with the leftover turkey" and out would come a pile of turkey sandwiches made with Hellmann's mayonnaise, every one of them looking delicious. Search "what about the leftover pumpkin," and it had no good ideas left.

Fun, right?

But the truly valuable stuff comes later.

They have a tool called Alex, whose job is reading email. When consumers write in, it first reads the sentiment and the actual issue in each message, then drafts a reply for the service agent inside Salesforce. The result: agents' email-handling time was cut by 90%.

Not reduced to 90% — cut by 90%. What used to be four hours a day on replies is now twenty-four minutes.

Brutal, right?

There's another one called Homer that writes product listings for Amazon, its tone automatically aligned to the brand's voice.

The supply chain is using it too. Palm-oil sourcing used to rely on reports; now the truck traffic around plantations, market intelligence gathered by crowdsourcing, even aerial photos taken on cloudy days, all get fed to AI to analyze whether the sourcing is actually sustainable.

Some will say: that's a giant, I can't afford to learn from it.

The opposite is true. Translate this playbook into moves a small company can get its hands on, and it comes down to: write your prompts specifically enough; feed the customer knowledge base you've accumulated into your own AI; then copy how they watch sales data — which SKUs should be stopped, and which are sleeping gems worth a push.

The tools can be small. The method is the same.

2. PepsiCo: Draw the Track First, Then Floor It

Second story, PepsiCo.

What surprised me most about PepsiCo wasn't how aggressively it uses AI, but that it drew two no-go zones first: hiring — no AI allowed; one-on-one precision targeting of consumers — also not allowed. And it brought in Stanford to build an ethical-use framework together.

Set the rules first, then go all out.

Now look at where it does let loose.

Internally there's a tool called Ada, named after the nineteenth-century mathematician Ada Lovelace — the one later generations regard as "the first programmer." This Ada's job is testing creative: throw a creative in, and it first tells you roughly how the audience will react. Turnaround gets faster, and the return on ad spend becomes something you can actually measure.

On the marketing side it gets flashier. The personalized greetings fans receive can come straight from soccer star Messi, every line different from the last. Healthy-snack development moves faster by analyzing posts on social platforms.

Inside the company, it's just as busy: AI bots recommend internal roles and more challenging new tasks to employees; demand forecasting and inventory planning are handed to AI; farmers are given AI tools to help raise yields; machine learning is used to dig into the details of operations so the company can hit its greenhouse-gas reduction targets. Even the ESG report: AI assembles the agriculture, value-chain, and product pieces into one sustainability report.

When you're holding a new tool in your hands, daring to say up front, "these two places we won't touch," is harder than daring to use it.

And it's precisely because the track exists that the racing team dares to floor the accelerator. Most companies get this order exactly backwards.

3. Salesforce: Turning Prompts into a Product

Third story — let's talk about the pain point itself.

Why do most marketers never get anything special out of chatbots?

Because of that blank input box. You type in "write an email," and the most it can give you is an opening line anyone could write. If you want anything better than mediocre, you have to feed in your business background and your customer situation, one big paragraph after another.

The trouble doesn't stop there. The customer profiles you feed in may stay in the chat history and then walk straight over to your competitors. You can turn off the "train the model on my data" option, but the whole workflow is still awkward, and confidential data is something you wouldn't dare touch at all.

Salesforce's approach was to simply turn this whole thing into a product. Its Einstein suite of tools is the equivalent of a ChatGPT that doesn't start blank: the common outreach scenarios come pre-built, and generative AI then rewrites for the specific customer. And this isn't something that only serves big clients like Unilever.

They ran a paid pilot, where email administrators could build workflows like this: generate audience segments and personas from sales data, draft email subject lines and body copy, then loop back to evaluate campaign performance. When it officially rolls out, industry-specific and company-specific prompts will work like plug-and-play plugins. Even PR — the trade that has been slowest to accept AI — could find itself pushed forward.

To put it plainly: prompt engineering shouldn't be a required course for every marketer. It should be built into infrastructure.

Back to That Cup of Dishwater

Now let's go back to my friend.

His problem isn't AI. His problem is that he's riding a race car like a bicycle.

Break the giants' approach apart and it's three moves: feed your own data in, draw the no-go zones, and distill the prompts that work into tools. Not one move is mysterious — each one just takes a bit more unglamorous grunt work than "opening a chat box and pulling cards."

AI won't eliminate the people who make content. The people who know how to feed AI their own data are eliminating the people who only know how to pull cards.

Next time you write copy that tastes like dishwater, don't rush to blame the AI.

First ask it: what have I actually told it?

Here's hoping that soon, you simmer that cup of dishwater into a pot of good soup.