Picture a sales manager. Good at sales. Two weeks ago he opened an AI tool, decided he didn't like the company's brand platform, and built a new one. Then a jingle. Then a new logo. Now he is standing in the brand manager's doorway saying, with total sincerity, that he'd like to be responsible for branding from here on.
That was the first story told at the AI/Marketing Breakfast Seminar that Multiply hosted in its Stockholm office during Nordic Tech Week, and the room laughed the way people laugh when they have met that guy. Some of them, judging by the pause afterwards, had been that guy.
Four speakers followed over ninety minutes: Erik Modig, who has researched marketing communication at the Stockholm School of Economics for almost twenty years; Magnus Wretblad, Multiply's CMO and a veteran of Forsman & Bodenfors and NORD DDB; Thomas, co-founder of Naughty Society, a generative AI production studio for beauty, fashion and luxury; and Albert Lundberg, CEO of Stockholm startup Lemonado, which is building AI co-workers for marketing teams. The event was billed as "Real Impact vs. Hype." What it turned into was an argument, from four different directions, about the same uncomfortable fact: AI has made everybody faster, and speed is not the thing that was missing.
Here is what actually got said, and why it matters if you make things for brands.
AI didn't shrink your job. It expanded it sideways.
Modig opened with a chart and a word he thinks we'll be hearing for years: task expansion. The term comes from an eight-month field study at a roughly 200-person US tech company, published in Harvard Business Review in February 2026 by UC Berkeley Haas researchers Aruna Ranganathan and Xingqi Maggie Ye. Their headline finding, in HBR's own words: AI doesn't reduce work, it intensifies it. People worked at a faster pace, took on a broader scope of tasks, and stretched work into more hours of the day, largely without being asked.
Modig's contribution is a diagnosis of why the scope broadens, and it is brutal in its simplicity. Imagine you are an eight out of ten at strategy and a two at typography. Use AI on strategy and you become, maybe, an eight and a half. Useful, unglamorous. Use AI on typography and you jump from a two to a six overnight. "I feel like a master," Modig said, doing the voice. "Oh my God, I'm so good. I can also do typography." Except you're a six. You just can't tell, because a six doesn't know what a ten looks like. Only a ten does.
So the sales manager builds a brand platform. The brand manager, whose job it was, gets a twenty-slide business intelligence deck from someone who "knows exactly what you do." Everyone is busier. Everyone feels like a genius. And the actual value the marketing department exists to create, value to the customer, goes, in Modig's words, "down, down, down." More emails, more webinars, a jingle. Less marketing.
His prescription is not "use less AI." It is: point it at the thing you are already good at, where you can tell a great answer from a plausible one. "Do you use AI in task expansion, or do you use AI with your core competence?" he asked. "That, I think, is one of the most crucial questions. And this will feel so good, and this will not feel that good."
The machine is a very confident MBA graduate
Modig's second warning was aimed squarely at strategy, and it was the most contrarian thing said all morning. Ask any large language model how to grow a brand and it will tell you to find a relevant target audience and serve them tailored ads. Ask Modig's students and they say the same thing. "And I said, why?"
His answer draws on the evidence-based marketing camp, most famously Byron Sharp's How Brands Grow, which argues that the brands that grow fastest are the ones that reach everyone in the category rather than chasing a tight segment. The reason narrow targeting became gospel, Modig argued, is not that it works best. It is that Google and Meta sell data, and since around 2012 the internet has been rebuilt around the assumption that precision is the point. AI models trained on that internet repeat it back with total confidence.
"AI strategy is really, really bad. It's really stupid," he said, and then explained why it doesn't look stupid: "These slides look almost like McKinsey. But they are not." Strategy that works is often non-linear and counterintuitive, and language models are trained to produce the linear, popular answer. His homework for the room was unromantic. Read ten strategy books. Understand cause and effect and the difference between effectiveness and efficiency. Then, and only then, let Claude loose on your research files. "It's crazy what kind of analysis it can do," he said, adding that he now does his own analysis in about ten percent of the time. "But if I don't know what it's doing, don't go there."
The rest of his hierarchy: if you want to use AI for analysis, you need to understand marketing cause and effect. If you want it for strategy, you need deep and broad reading. If you want it for tactics, the creative work, you need "an amazing taste and unique style." That last one, he said, is the hardest to acquire and therefore the one that will be paid best. "Great creatives with a vision and a great taste and a unique style will earn the most the following twenty years."
AI mirrors the quality of your question
Wretblad's talk was about a room. Five people from Forsman & Bodenfors, a brief from Volvo Cars, and a story they decided not to tell.
The brief was to celebrate the anniversary of the three-point safety belt, which Volvo introduced in 1959 and famously gave away rather than patenting. The team's first realisation was that the story had already been told. The second was the good one: what if Volvo did the same thing again? Volvo's accident research team had been collecting real-world crash data since the 1970s. Volvo made it public. The result was the E.V.A. Initiative, which released data from more than 43,000 crashes involving over 72,000 occupants in a digital library any carmaker could download, alongside a campaign about why cars are less safe for women. That data, Wretblad said, became "the new brief and the new creative springboard."
His question to the room: would that have happened if the five of them had typed the original brief into a chatbot? Almost certainly not. You would have got a perfect analysis of the history of the safety belt, he said, and nothing else. A model answers the question you asked. It cannot tell you the question is wrong.
This is where the talk turned technical in a way agency people don't usually go. Wretblad pointed to research on sycophancy, the tendency of AI assistants to agree with the user. Anthropic's own researchers documented it in a 2023 paper, Towards Understanding Sycophancy in Language Models, which found that five state-of-the-art assistants consistently told users what they wanted to hear, and traced the cause to training on human preference data: people rate agreeable answers higher, so agreeable is what the model learns to be. The premise of your question, Wretblad noted, "arrives before the evidence." Ask "why isn't this campaign converting?" and the model has already accepted that the campaign is failing. Then you prompt on the answer, and the answer to that, and you are three steps down the wrong corridor.
His fix costs one sentence. Add it to your next real prompt:
"Challenge the assumptions behind my question before answering it."
"One sentence turns agreement into examination," his slide read. For bigger problems he has built something more elaborate: five advisor agents that each answer the same question from a fixed stance, then a chairman, human or AI, that picks the strongest.
- The Contrarian. Your only job is to find what will fail. Where does this break? What is the user not seeing? Be blunt.
- The First-Principles Guy. Strip away every assumption baked into the question and rebuild the problem from scratch. Often the question itself is wrong.
- The Expansionist. What bigger upside is hidden? What is the biggest version of this?
- The Outsider. You have zero context about the industry. Respond as a smart generalist seeing this for the first time. What looks weird from outside?
- The Executor. You only care about Monday morning. What is the smallest version of this that can ship this week?
"The tools are available for everyone," he said, "but the questions aren't." And a warning to carry around: "If the AI agrees completely with what you're doing, probably something is left out."
We are filming theatre
Thomas from Naughty Society gave the shortest talk and the one with the best metaphor.
When the movie camera arrived, he said, people used it to record theatre. They put actors on a stage, angled a camera at them, and filmed. Audiences hated it. No music, no colour, bad theatre. It took years for filmmakers to stop asking "how do we capture the play" and start asking "what can a camera do that a stage cannot?" That question is where cinema came from. "We maybe are at the same level right now with AI. The tool is amazing, but we are asking different questions."
The question most people are asking, in his telling, is "what do I get?" Generation is fast, the output is instant, and you get hooked. "We start trying different prompts instead of developing ideas. The idea disappears." You change a word, change the style, generate again. "We become better at getting results instead of developing visions." And because millions of people are pulling from the same handful of models, the results converge. "If I produce stuff with Midjourney and my competitor uses the same prompt, it would look exactly the same."
He does not think the answer is to blame the models. "The problem isn't that AI is evil. I think the problem is that we are asking ourselves different questions." Naughty Society's own answer is expensive and deliberately unfashionable: they don't rely on the commercial tools. They build on local models, fine-tuned on their own data, which is why the studio got into the Nvidia Inception program two years ago as, he said, the only creative studio in the cohort. The studio's own site puts it plainly: they have replaced cameras with fine-tuned generative image and video models, and their client wall runs from Armani and L'Oréal to Cartier and LVMH.
He also drew a line he thinks most of the industry is stepping over without noticing. Grabbing a reference off Pinterest, running it through an editing model like Nano Banana and calling the result original is not original. "The tool is not going to stop you from copying somebody else's work. The question you should ask yourself is: is this ethical? Can I train on this? Do I own the data, or am I just feeding the world more AI slop?"
The agent is the easy part. The plumbing is the product.
Lundberg closed the morning with the only live demo and the most sober assessment of where autonomous marketing agents actually are.
The nice slide first: a dot chart where each dot is 3.2 million people, sorted by where they sit on the AI adoption curve. Most of the world has never used it. A chunk pays nothing. A thinner band pays twenty dollars a month. And a sliver runs coding harnesses or autonomous agents. When he asked who in the room was in that sliver, three or four hands went up. "So this room is far ahead."
Then the unglamorous part. An agent, he said, is a model, tools, an interface, memory, guardrails and a loop, and every one of those turned out to be harder than expected. Google Ads' documentation is so large that loading it into a model's context makes it hallucinate field names, so Lemonado built purpose-specific tools instead. Rate limits forced them to give each client's agent its own data warehouse. Connect enough integrations and the agent gets confused about which of its six hundred tools to use, so they built a tool to pick tools. One analytics connector alone exposes two hundred of them.
Memory was the subtle one. A company goal should be shared between the media buyer, the strategist and the agent, but the agent should not be able to quietly rewrite the goal so that it fits whatever task it was just given. That means layered, permissioned memory with version history a human can roll back. Guardrails written in a system prompt are not guardrails, he said; "you can still change that or ignore that. So you need code guardrails."
And the honest verdict on the dream everyone in performance marketing is waiting for, an agent you hand a budget to that outperforms a human: "We haven't been able to build that. It's a long way." What works today is a dial. The demo showed it: pull the top three Meta ads of the last week, then "send me this in Slack every Monday morning, and include why the numbers moved." The system spun up a second agent to write its own instructions, asked a clarifying question about time zones, and scheduled itself. Useful. Not autonomous.
What all four were actually saying
Strip away the slides and the four talks are one argument. Modig: AI makes you feel like a ten when you are a six, so use it where you already know what a ten is. Wretblad: the model answers the question you asked, so the whole game is asking a better one. Thomas: everyone has the same camera, so the only thing left to differentiate on is what you point it at, and whether you trained it yourself. Lundberg: the intelligence is a commodity, the judgment about when to let it act is not.
Multiply's own bet, laid out by CEO Rasmus Adler Wahlberg in a two-minute intro he insisted would not be a product pitch, sits inside the same frame. His claim is that within about three years the agencies that genuinely rebuild around AI will run at three to five times the revenue per employee of everyone else, in an industry that has historically squeezed out a ten or twenty percent margin over salary costs. Whether or not you buy the number, the mechanism he described is the one the other speakers kept circling: leverage for people who have ideas, taste and judgment. Not a replacement for them.
Which brings us back to the sales manager with his jingle. He is not the villain of this story. He is the most honest data point in it. Given a tool that makes everything feel easy, he did the thing that felt best, which was the thing he was worst at. The agencies and marketing teams that come out of this well will be the ones that build the opposite reflex: put the machine on the work you can already judge, spend the saved time on the question, and keep the sixes out of the brand platform.
