Generative AI for Marketing: You've Tried It All — So Why Isn't It Making Money Yet?

A while back, a friend of mine who runs a marketing team sat across from me at dinner, looking miserable.

Since last year, he told me, his company had tried just about every flavor of generative AI: auto-written ad copy, batch image generation, one-click ad creatives, even customer-service scripts that could be dropped into a template. But when it came time to report up, the boss's first question was always "Where's the ROI?" And there he would sit, clutching a stack of PDFs with not a single number that could make anyone nod.

I told him he wasn't the exception. Most marketing teams at most companies are stuck in the exact same place.

A pilot means: a pile of tools tried, a pile of decks written, and still no glimpse of any money.

Scaling means: taking the two or three things that are already proven and plugging them into your daily workflow so they keep producing on their own.

In marketing, the worth of generative AI isn't "how many things you tried." It's whether you can scale it into profit.

Let's Start With Three Sets of Numbers

The first set: the pessimistic picture.

Bain & Company ran two rounds of research in 2024 — one covering 200 large U.S. companies, one covering 184. Each time it asked the same question: has generative AI in marketing exceeded your expectations?

Only 27% answered yes.

That means the other seven in ten either feel it's "nothing special" or are still lost in the maze of "pilots waiting to be signed off."

The second set: the optimistic picture.

Retail shows it most clearly. Retailers that use AI to tune their targeted ad spend get ad returns 10% to 25% higher than their peers. Spend the same 10,000 budget — a competitor might recover 8,000, while you bring back 9,000 to 10,000. Whatever extra comes back is pure profit.

The third set: a few concrete examples.

Etsy, the image-based e-commerce marketplace, turned AI into a "gift mode." Fill in two preference tags about the person you are buying for, and AI picks the best-matching persona from more than 200 profiles, then hands you a gift list. The "what on earth should I buy" question that used to eat an entire evening is now brushed away by AI in one pass.

Booking.com, the travel platform, folded AI into trip planning. Just say "a beach trip with the kids, six days," and it lines up the hotel, the flights, and the restaurants all at once. What used to be half an hour of homework now takes about the time to drink one cup of coffee.

P&G, the consumer-goods giant, goes one step further. Instead of the answers people type into questionnaires, it trusts the data: its AI reads real usage figures from smart hardware — like that Oral-B iO electric toothbrush, reporting how long you brushed yesterday and how many times, all of it. From there it works backwards to decide which product line to support, which is far more accurate than sending out dozens of surveys.

Finance, healthcare, telecom — the industries that regulators watch closely — can get the same advantage, as long as the line of privacy and compliance is drawn out clearly.

So is anything wrong with AI itself? No. The only problem is that the people who can take a project from "pilot" to "production" are far too few.

Where Does It Get Stuck?

In the game of scaling, the hardest part is not the algorithm — it is the organization.

CFOs and CEOs today sing the exact same tune: no extra budget, more speed, one more notch on the business target. But on the marketer's side the data landscape is growing more complex, data scattered across dozens of tools, while the call for personalization gets louder. Squeezed between the two, a person in charge can barely gather enough strength to drive one pilot all the way through.

So the key is one sentence: Don't scatter your energy. Two or three battles fought well are worth more than a hundred high-sounding promises.

Which directions should get the focus, then? Combining what Bain found, the effort falls into four pools:

• Smooth the workflow. Concept drafts, image generation, translation, compliance review, asset archiving — chain them all into one automated pipeline.
• Produce with personalization. Copy, visuals, ad creatives — move from "one by one, in a rush" to "in batches, aimed at each person."
• Insight with the digital twin. Give a customer a digital twin, test the versions in a virtual environment first, and only put real money in when the winner is clear.
• Measure it, then tune. The AI reads the ad data every day, pulls the reports scattered everywhere into a single view, pumps more into what wins, and pulls back what falls off.

These four are not four separate balls; they are the line of the one assembly: first make the process smooth, then scale the content, then use the data to skip the dead ends, and finally feed it all back into the chain.

Some Companies Made It Work — What Did They Do Right?

Bain reviewed the companies out front and distilled five moves. Here they are in my own words.

1. Set a Goal That Carries a Number

The easiest dead end is "a pile of promising ideas" — everyone feels confident, yet nobody can answer, in one sentence, "which KPI we will move, from what number to what number, next quarter."

If you cannot say your goal in one sentence, AI will gladly carry your "no clear result" into the next year.

The companies that know what they want leave no room for ambiguity. One global financial-services firm set the goal "shorten the time it takes to bring marketing campaigns to market by 50%," and rebuilt its AI content system, its tech stack, and its assistants on that single figure. A media company set the goal of AI producing "a different piece of content for each recipient" — and the click-through rate of its new campaigns jumped 5 to 7 times as a result.

The more concrete the goal, the more you get out of the AI; the fuzzier the goal, the more the AI just gives you one verdict: "Covered."

2. Take the Battle Nearest the Money First — Don't Spread Too Wide

Pilots most easily turn into "ten gardens" — the team is so drawn to the very idea of "AI innovation" that they forget to decide which plot of land could give the earliest harvest.

The right approach: pick the two or three closest to the money — for example, auto-writing direct-mail copy, auto-posting to social media, and auto-building landing pages. Let these produce numbers first, and go deeper only later.

There is a good example in a well-known consumer bank: for its app-banking and search campaigns, it built one AI creative assistant. The assistant took 75% of the creative production time away and, as a side estimate, new-account openings could be lifted by another 20% to 25%. It did not try to do too many things at once: from day one it locked onto a single metric it could move, and ran that whole distance to the end.

3. Let the People Who Use It Have the Vote — Don't Send the Tool from Above

A tool that is complained about and left on the shelf — nine times out of ten the same story: the IT department "air-dropped" it in, disregarding the rhythm of the team and not reflecting how people actually work.

The correct direction is the inverse: first have the marketing people draw up their everyday workflow, and only then do the data or tech team turn that into AI. Or follow the practice of one financial firm: nurture a group of internal "AI super users" — the first batch really learns the tool, then carries the colleagues along.

A tool that is thrown in until nobody uses it is a waste; the same tool, co-built with the people who need it, is where real adoption begins.

4. Put the Learning Into the Weekly Calendar — Not Just Issue Licenses

Give everyone an account for ChatGPT Enterprise: the first week new, and within one month back to where everything was. If the tool is not folded into the rhythm, it remains an ornament.

The teams that win treat it as a wheel that does not stop: weekly, a fixed slot on the meeting agenda — the team reads out the prompts they tried that week, compares the results against the previous week, and iterates the week after. AI is not a firework at the end-of-year meeting; it is the same door a team walks through every single week.

Giving license is the cheapest step; but what actually flips people's practice is a feedback that arrives weekly and repeats over and over.

5. Don't Rely on In-House Work Alone — Treat the Ecosystem as a Lever

The marketing-tech ecosystem has always been fragmented, and vendors keep shifting positions from quarter to quarter. Rather than constructing every single piece in-house, set aside a small slice of budget each quarter to trial a shortlist of vendors: Adobe is strong in image generation and digital experience, Jasper in copy at scale, Synthesia in digital humans, Typeface in brand content. Betting everything on one vendor is too risky; it is better to test two or three at small scale, see which one fits your business, and then scale it up.

And a reminder: agencies and creative partners are going through the same AI makeover right now, and your competitors are likely already using AI quietly. Treating the ecosystem as leverage is like adding lubricant so that scaling runs faster.

In Closing: Think It Through — Scaling Is Where the Real Results Show

Unlike the early days when it first emerged, generative AI in marketing has already moved from "should we try?" to "have we scaled up?"

The leading teams are already preparing for the next round. Search habits are shifting: the customer may no longer click through a list of links one by one, but instead hand the question to an AI and get a direct answer. Would the old rules of channel placement still hold there? And a step further: what if your customers themselves are also an "AI" — how do you write marketing copy for that? That is a new battlefield. It will not stop at adjusting the placement logic; it changes how the whole brand presents itself.

Now, back to my friend.

I did not ask him to drop AI. I told him one honest truth: in the pilot stage, no matter how much more time you throw in, the answer will not get any clearer. So shrink the list. From the "I've tried everything" pile, take out the two bets nearest to the money; back them with data, a process, and a budget. Then ask: can you hand me one number within three months?

When "I tried AI" becomes "I have an AI pipeline running," we can talk about the next generation — and have the confidence to do so.