Roger Hurni has run the same Phoenix agency for nearly twenty-seven years. He arrived with two mistakes, one about people and one about software.
Will and I first met Roger Hurni through WorldCom about a year ago, and I only worked out while researching this episode that he has been running Off Madison Ave for nearly twenty-seven years. In agency terms, that is close to a lifetime because, trust me it is challenging running an agency full stop. Sixteen years of Prohibition has already turned the pair of us grey, so I wanted to know how anyone survives that long without either selling up or losing their mind.
His answer was not a strategy. It was two admissions, which is precisely why he was worth having on. One of them cost him staff and new business. The other cost him a piece of software he never needed to build in the first place.
Lesson one, work out whether it is cheaper to rent than to build
Roger was in artificial intelligence long before the current noise. He started a machine-learning company that later moved to ontology-based learning, so when the wave arrived, building an enterprise AI inside his own agency felt like the obvious next step. That instinct turned out to be wrong.
“We made the mistake of trying to build a kind of an enterprise AI for our company when all this was happening the last couple of years. And I’m not saying companies shouldn’t do that. This is not the lesson. What I’m saying is that you need to evaluate if it’s cheaper to rent than buy. That’s not our core competency.”
After the build failed, they found something off the shelf, customised it to the way they work, and paid a fraction of what building and maintaining their own would have cost. He also pointed out that most people are wildly wrong about the size of the market they are trying to enter.
“People think there’s a handful of these things, and there’s not. There’s tens and tens and tens of thousands of AIs out there. Everyone thinks they’ve got a better mousetrap.”
His framing of the wider gold rush is the one every investor reaches for eventually.
“In a gold rush, it better be the guy who sells picks and shovels. So let everyone have the gold rush trying to build stuff. I’m going to be the guy with the picks and shovels using what’s already there.”
He credits that line to Mark Twain, as almost everyone does. Twain died in 1910 and was first given credit for the saying in 1982, which is a fairly relaxed approach to sourcing. The point survives the attribution though, and Roger applies it to Apple, which has spent this whole cycle declining to build a frontier model while everyone else sets fire to trillions. He reckons that patience is a strategy rather than a failure of nerve.
Lesson two, one toxic hire can empty your building
The second mistake is one almost every agency owner has made, and almost none will name in public – I had to hold my hand up to this too.
“We as a company made a massive mistake hiring somebody who was just toxic to the culture. And it ended up being a case where we weren’t getting business because of this person. We had some good people punch out, some of them without notice.”
What struck me was how long the correction took. It was not simply a case of spotting the problem and acting. He had to get the whole leadership group to agree that it was a problem at all, which is the part nobody warns you about. By the time everyone was aligned, good people had already walked out of the door without working their notice.
His conclusion was to stop importing senior people who do not understand how the place works.
“I would rather groom people who’ve been with me for years, understand who we are and why we work the way we do, and find ways to get them into leadership positions to grow the company. So I will not make that mistake again.”
He was careful to say this is not a universal rule, quoting his father’s line that what is good for one marriage is poison for another. He was much less hedged about the trade-off itself.
“I would rather have someone who’s a bit of a lesser performer, but fits in culturally, than a high performing person who’s toxic.”
He also made the point that this has nothing to do with seniority. The damage can come from the graduate on their first week just as easily as from a new managing director.
What is an AI philosophy, and why does it beat an AI policy?
An AI policy tells your staff which tools they may use and what they must declare. An AI philosophy tells them how to think while they are using them.
Roger wrote one for his own team because a policy document was not answering the question he actually cared about.
“So the philosophy was you never look for AI for an answer. I want you to look for AI to expand your options. We use it to find gaps. We do not let it replace judgment.”
He is equally unimpressed by clients at the other extreme, the ones who ban it outright. His view is that you can refuse to use AI for output and still be daft to refuse it for input, because comb-through-the-research work is exactly what it is good at. A blanket ban, he reckons, gets you outmanoeuvred by competitors inside a very short window.
Then I told him about a phone call I got while Will and I were in Paris, at the same WorldCom event Roger was at. One of our account directors had been sat down by a client’s marketing director and told that they had started using Claude, it was really good, and why did they not just produce all the output themselves and let us push it out to our media contacts. I put the same scenario to a panel at an AI and PR conference in London a few weeks ago. Roger did not need the panel.
“I would probably say to that person, we’re probably not the right agency for you. There’s always going to be bad actors. There’s always going to be people who are looking for shortcuts.”
The research backs up why that matters more than it sounds. Roger cited a paper he had read from Wharton, and having checked it, he was right on the substance. Steven Shaw and Gideon Nave published their work on artificial cognition in January 2026, extending Daniel Kahneman’s system one and system two thinking into a third system that sits outside the human brain entirely. Their term for the failure mode is cognitive surrender. Across 1,372 participants, people accepted correct AI answers 93 per cent of the time and wrong ones 80 per cent of the time.
“That’s my fear, that because it’s a new tool, people are just accepting the answers that come out of it, even if they’re wrong. And they’re reporting higher degrees of confidence, and that’s a danger zone.”
Lesson three, ask people to write the page AI cannot possibly write
This is the bit I have already stolen. Roger sets new employees three tasks. First they write something themselves. Then they have AI write something for them. The third task is where it gets interesting.
“I want you to write a page on something AI cannot possibly write. There’s no way in the world AI could write this for you.”
People push back and tell him AI can write anything, which is when he produces the example that ends the argument.
“AI cannot write about when you were eight years old making chocolate chip cookies with your grandmother in the kitchen and the feelings and emotions that you had. It doesn’t know that story. It’s not been trained on that.”
Underneath the exercise sits a term he uses that I have not been able to shake, which is expensive sameness. If two people join with the same experience, the same background and the same tools, something else has to separate them.
“It’s the person who comes back with that messier document, the handwritten notes in the margins, the person who’s like, hey, I didn’t like these couple of answers, so I dug deeper and I went back to the research and I found this that it didn’t flag.”
“The future doesn’t belong to people who use AI the fastest. It belongs to people who use it the deepest.”
We ended up on LinkedIn, which now runs a button letting you report a post as AI slop, introduced at the end of July. My problem with it is the obvious one. The platform spent years handing premium users a tool to write their posts for them, and is now asking those same users to police the results. I said on the episode that I give the button two months. Roger was gentler than me, and framed it as Microsoft realising that AI does not want to be trained on AI.
He closed with a line he loved but could only half remember, so here it is properly attributed. The author Joanna Maciejewska wrote in March 2024 that she wants AI to do her laundry and dishes so that she can do art and writing, rather than the other way round. Roger is hoping that becomes the future. So am I, although my dishwasher remains stubbornly unintelligent.
If your team is producing work that reads like everybody else’s, the fix sits upstream of the tools, and it is the thing our public relations management training spends most of its time on.
Listen to the full episode with Roger Hurni
Watch the full episode with Roger Hurni on YouTube, or browse every previous episode at embracingmarketingmistakes.co.uk. His book Outthink, Outperform covers the behavioural marketing side properly, and the second one, on competing with AI rather than against it, is being written now.
Embracing Marketing Mistakes is the weekly podcast I host with my Prohibition PR co-director Will Ockenden, where senior marketers, founders and agency leaders talk honestly about the decisions that went wrong and what they learned from them.
So here is the question Roger left me with. If you handed your best person and your laziest person the identical AI tool tomorrow morning, would you still be able to tell them apart by Friday?
Chris Norton is the founder of Prohibition and an award winning communications consultant with more than twenty years’ experience. He was a lecturer at Leeds Beckett University and has had a varied PR career having worked both in-house and in a number of large consultancies. He is an Integrated PR and social media blogger and writes on a wide variety of blogs across a huge amount of topics from digital marketing, social media marketing right through to technology and crisis management.
