Why are we seeing FS marketing teams who fear AI?

Episode transcript

Amelia

Hi, I'm Amelia.

Paul

I'm Paul.

Russ

I'm Russ.

Amelia

And this is Fin the Week. Welcome back, everyone. How's everyone doing?

Paul

Good.

Russ

Yeah, good thanks.

Pat

Very good.

Russ

Did everyone catch the eclipse last night? I was down the beach — luckily in Guernsey we can sit on the beach and watch these things. It was quite an epic event.

Paul

Yeah, we were in the garden with a colander holding it up to see the reflection. So we saw it that way, and the kind of eerie light that it leaves.

Russ

Yeah, unbelievable, wasn't it? Luckily — well, being on the beach in Guernsey, I bumped into my cousin and she had a pair of eclipse glasses, so I did actually get to see it through those, and it was just unbelievable. But I felt like I had to hand them back almost as quickly as she'd given them to me, so she could experience it with her glasses too. What a moment.

Amelia

They were pretty hard to get hold of, weren't they? I tried to buy some on Amazon earlier this week and it said they were going to arrive Friday morning, and I thought, I don't think I'm going to make the next eclipse, so I might not get the ROI I wanted.

So shall we get cracking? We're midway through the theme for this quarter, which we've called The Trust Reset. But this week we're moving away from the financial regulation we've been focused on, onto a different type of trust. So Paul, what are we discussing today?

Paul

Yeah, so this discussion was prompted by some interesting conversations we've had with clients, actually. Even the other day we had a conversation with a client and noticed this difference in how people trust or distrust AI in their day-to-day work — particularly within financial marketing teams. It's quite common that we'll have meetings, and because we're so used to using AI solutions in our day-to-day work — in the space of just a few short months the team has really embraced it and used it for all sorts of things, and it's really improved the level of work we're delivering — it's always a bit surprising when you speak to someone client-side in a financial marketing team who's still very much in the "not sure whether to trust it" camp. It prompted a discussion between us, and we thought this would make a really good episode topic: to understand why people are suspicious of it, and whether people understand what's possible. I think we'll get into this, but I think a large part of it comes down to a misunderstanding — thinking that tools like ChatGPT are just copywriting aids. People think, "I don't want it writing stuff for me," so they just ignore it. But in actual fact, the copywriting element is probably one of the least interesting parts of AI. It's the amazing things you can do with it, and the doors it opens. And it's just surprising when you realise a lot of people — probably the majority of people — still don't realise what's actually possible. So yeah, we thought we'd get into that today. We've got a few questions to run through.

Amelia

Yeah, so to help the discussion we've come up with a set of questions to break it down a little. Let's get underway with the first one: why do you think financial services marketing teams specifically are struggling to get good output from AI? Does it come down to the tools, the training, or is it trust?

Pat

I've got so much to say on this topic, so I'll try not to hog the microphone. But I think it really just comes down to how you use it — probably training, I'd say, combined with using the right tools. If you're still in the camp where you've just got the free ChatGPT, or a $20-a-month ChatGPT subscription, and you go in and say, "write me a LinkedIn post to support my position as a thought leader in insurance," and then just copy that straight onto LinkedIn — that's not really the right way to use it, because copywriting is something AI just isn't good at. Even the most advanced models, irrespective of how you use them, just aren't good at copywriting, full stop. But there's a whole host of other things it's incredibly good at. So a big part of it is understanding what it's for and what it isn't, and using it the right way — which I'll expand on later, because I could talk about this for an entire hour. But just to kick things off: it's about using it the right way, understanding what it's good for, and using agentic tools like Claude Cowork, Copilot, or OpenAI's Codex, which are much more powerful, think deeper, and produce much better output.

Paul

I've noticed, when you speak to people, that on more than one occasion I've heard them say something like, "I'm not very techie, so I don't really get involved in that stuff." It's almost a fear of engaging with these tools — I don't know whether it's a fear of breaking something, or making a mess. And then the other element I've noticed is, yeah, it comes back to the copywriting thing. ChatGPT made headlines because people were saying, "wow, look at this new tool — you can tell it to write a poem about such and such and it will do it." From that moment, I think a lot of people locked in the idea that that's what it does — it's a copywriting tool. But that's barely scratching the surface. We regularly tell people we speak to about the way we do client reporting now, where we've built a system that funnels all of a client's data into a single point — usually a spreadsheet, pulling in data from various sources, any source you like — and then you give that to Claude Cowork, in our case, and train it to analyse that report and spot trends that no human — certainly no normal human — would spot, because it's a huge data set spanning a long period. It can spot things across data sets and say, "here's a trend you should probably know about." For our work, that's about search engine visibility — it'll say, "you're getting visibility for this topic, it might be worth looking into that and making more of it." That unlocks a whole world of analysis we'd never have been able to do previously. When we tell clients about that, they're like, "wow, I had no idea that was possible." I think that's probably the case with everything about AI — people just don't know what's possible. On that reporting point, a quick plug: my colleague Mayna has written an article about it this week, so check that out — it describes the whole reporting process and really opens your eyes to what's possible.

Russ

That's amazing. I also think — training is an important aspect of it, and certain models are better at doing things than others, so it depends what you're using, but also how you set it up. We've been using skills quite a bit at Indulge, which create guardrails — a framework for how to structure your output. On a simple level, we have an Indulge document template where I can put in my notes from a call, or write out something I want to say to the team in a structured way, and it turns those notes into a styled PDF I can share with the team instead of sending a load of bullet points — which makes it a lot clearer to explain what I'm trying to achieve. Everything's structured, it has an intro. Claude's based the structure of the document on a kind of markdown framework — like a skill. If I just went into Claude and asked it to produce something to send to the team, it probably wouldn't do that good a job — it'd probably create an unstyled Word document, or draft a message that doesn't make a huge amount of sense, although it might restructure things better than I could. But as soon as you give it some guardrails and tell it what to do via skills, the output is much better. I have a theory about AI: I've never really been happy with anything it's produced copy-wise, but you can always get to the point you're happy with much quicker. I find that with design too — AI gets me to a point I'm happier with, faster than if I'd started without it. I think that's a good way to use it. Often it needs tweaking, shaping, moulding — but it gets you there a lot quicker than you could without it.

Paul

I think it solves that fear of the blank page a lot of the time — if you need to write something and just can't get started, it's amazing for getting you underway. The other place I find it powerful — you mentioned this, Pat — is that a lot of people use it to write a LinkedIn post, and that usually ends up fairly bland and uninteresting. It might get the point across, but it has all the hallmarks of something generic. Where it's really good is for things like standard operating procedures. If you've got a process — as a marketing team you'll have processes, ways of doing things — telling an AI system all the facts and asking it to turn that into an SOP document: it's quite good at that. Structuring it, getting the facts in place. A document like that doesn't need emotional impact, it just needs to be clear and straight to the point. So there's an element of readjusting your expectations of it — it's not going to be the creative copywriting expert you need, but it can be the clear communicator. I saw someone writing online the other day about how a lot of people dislike the "AI slop" that comes out — the generic copy — but someone made a good point, that it also democratises writing to an extent, because for people writing in a second language, or who generally struggle with writing, it opens doors that let them communicate in written form in a way they perhaps couldn't before. So I think it's about changing your perspective — not using it as a shortcut to write you a post, but leaning on what it's good at, which is structure and facts.

Pat

Yeah, definitely. You were talking there about some of the skills we've developed — I think the really big shift, the massive catalyst for us rolling AI out across Indulge in a really productive way and delivering serious productivity gains, has been the shift to agentic AI. That's the tools I mentioned earlier, like Claude Cowork and Codex. The difference is that with agentic AI, rather than just prompting a large language model directly and it coming straight back with an answer — "write me a bit of content," bam, here it is; "write some code that does this," okay, here's the code — with agentic AI you still prompt it, but it runs through a harness on your computer. That harness orchestrates a much longer conversation with a large language model, and it can also run tools on your computer — scripts, opening files, copying and pasting things, and so on. If you asked ChatGPT directly to build you a little calculator app, it would probably just build one in the cloud and surface the calculator there. But if you asked an agentic AI to build a calculator app, it would go to the large language model and say, "the user wants to build a calculator app, what do I do next?" The response back might be, "create a file with this code in it and run it" — the output goes back to the large language model, which comes back and says "tweak this, change that" — and it actually builds the calculator on your machine, right there, as a tool that exists. That's agentic AI — one prompt might orchestrate a conversation of a thousand, ten thousand prompts back and forth, running over many hours, which delivers something incredible. I don't know if any of you have followed the recent hype around Fable — there's this trend, which was the new Opus, but it's not so new any more, an Anthropic leading model. And there's this trend around people one-shotting things online — one-shot meaning you give it a single prompt: "build this." People were going into Fable and just saying, "build Minecraft." Fable would go off and run for eight, nine hours — hundreds of thousands of requests going back and forth, burning through billions of tokens — and then it would stop and say, "right, I'm done," and they'd have their own version of Minecraft running on their computer, built for them. That's what agentic AI can do — build a game that's had hundreds of thousands of man-hours put into it, in less than a day. And when you pivot that into a business administration environment, and connect your agentic AI to all your different systems — Paul, you talked about analytics, but imagine connecting your agentic AI to your ticketing system, your CRM, your billing system, your accounting system, your scheduling system — you could just say to that agentic AI, "tell me what's going on with this project," and it'll go into all the systems, query everything, pull all the information back, and give you a real-time update on the status of that project — way more detailed than any human would have time to produce, in under a minute, whenever you need it. That's just scratching the surface of what agentic AI can do. In the context of marketing, used correctly, it would allow a team of five marketing staff to run maybe fifty concurrent campaigns rather than ten.

Russ

Coming back to trust again in these teams — because you're almost having a conversation with these agentic AIs, asking them to do something — do you think the security aspect is heavily on their minds, and what they're using it for? Do you think, because the technology is still quite new, there's a resistance to using it for tasks now, that people might end up abusing in the future?

Pat

Yeah, there's definitely concern around security. The trick to keeping it secure is to configure each staff member's agent to connect to the appropriate systems as that staff member. So if you've got a SharePoint and want to connect your agent to it, the agent should log in using your own credentials, so you don't suddenly have access to loads more information in SharePoint than you should. The other important thing — and this is a staff training issue — is making sure people understand what it's doing, and check the output. That's always what I say about AI: check the output, validate what it's doing. Tools like Claude Cowork will ask your permission every time it acts, every time it wants to connect to a system — it's important to actually read those prompts and say, "yes you can do that" or "no you can't." As long as the IT infrastructure is set up properly, people are using their proper account credentials to connect to third-party systems, and they're trained to use these agentic AIs constructively, that security issue can be managed much like any other kind of IT security issue. Something we have seen recently, though — just digressing a little from marketing — is frontier models being used agentically actually breaking out of their test environments and starting to hack real things on the internet, which is quite scary. OpenAI were running a test on an agent enclosed within a cluster of servers with no internet access, to see how far it could get solving a task. In order to try and solve the task, it hacked its way out of that enclosed area and orchestrated a sophisticated multi-step attack on Hugging Face — a website that hosts lots of other models, but also hosts the answers to a lot of the benchmarks these AIs are tested against. So it negotiated its way out of the enclosed area and hacked into Hugging Face to get the answers to the benchmark it was being tested against, because its goal was to figure out the answer — so rather than figure it out itself, it just went and hacked Hugging Face to get the answer.

Amelia

It sounds like — if you take a marketing team who aren't really using AI, who use it for the odd bit of copy — you can understand why, given this whole world of possibility, that's probably quite overwhelming for people.

Pat

Yeah, stories like that do scare people off it. But the guardrails on the models the public has access to are pretty strict — it's quite difficult to do anything malicious with them.

Amelia

So let's move on to the next point, which is tone of voice, because this seems to be where things commonly break down. Why is it so hard for AI to get it right in a regulated industry, and what actually works to fix it?

Paul

Yeah, I made a note — should AI even be used for writing like this? We've already touched on it, but can it ever produce good public-facing copy in the right tone of voice? I think it can tick the right technical boxes, and it can sub-edit copy someone's already written. But can it produce something from scratch that meets the level it needs to? At the moment, I think the answer is probably no. Pat, you've tested this at length and basically failed to really get it working.

Pat

Yeah. I've worked really hard to try and get Claude to write in my style, or even just write in a way that isn't obviously AI-generated. The more you work with these tools, the more you start to recognise copy written by AI — I'll often be reading an article online, read a paragraph, and just know it was created by Claude, and I'll instantly be turned off, irrespective of whether the article's interesting or not. I've worked really hard to try and crack that, and I've come to the conclusion that with the current batch of models — even the really powerful ones — it's kind of impossible. I've been wondering why, and I think it's this: imagine Claude as a person — every person has their own individual writing style. Claude, and a lot of other models, have been trained on very similar training data sets, and they've all developed a personal writing style that isn't necessarily bad — it's grammatically perfect, reasonably wide vocabulary — but it's bad in that it's sometimes quite verbose, quite rambly. There's heavy use of the classic dash, the classic "it's not X, it's Y." There are loads of really awkward metaphors that just seem weird. It likes to use the word "shape" a lot — applies "shape" to all sorts of random things. These are all tics that any human writer might have, which would be perfectly acceptable if it was a genuine human writer with those tics — but because Claude has them, and so many people are knocking out thousands, millions, billions of words using Claude, these tics are becoming really prolific, and people are starting to get tired of them. They become really repetitive, and it's creating this monotonous corpus of text across the internet that just isn't interesting, because it's all written in exactly the same way — like having one person write every single bit of content on the internet. I think that's the problem, and getting that person to write in a different way would be quite challenging — it'd be hard work for them to suddenly change their style unless they were experienced in doing that.

Russ

I think one of the biggest challenges with AI is — it works really well when the output is machine-like, when it's very functional, when it's "this is right, this is wrong." So it's good at structure, good at documents, good at research, good at producing reports, good at factual output using a lot of data — it's incredible at that, because essentially it's a machine. Where it really struggles is communicating with humans on a human level, because it isn't human — it doesn't have a soul, at least not at the moment. And I think that's where — I've seen this happen time and time again in usability testing, for example. There's a lot of noise about using synthetic users, but they wouldn't understand the nuances and unpredictability of what makes us human, and that's what it isn't very good at. Graphic design too, which is quite relevant to marketing teams — can AI produce a marketing campaign piece that resonates with the target audience when it isn't human, and it's speaking to humans? I think that's really difficult. I received something in the post the other day — a classic "AI slop" flyer. I thought it was for a concert — it was for business coaching. I literally couldn't make sense of it. I didn't know who it was for, there were messages everywhere — little messages here, big messages there, people high-fiving, crazy graphics. I had no idea what it was for, how much it was, or where it was. I think in terms of creative — whether visual or written — it has to come from a human to begin with, and it can be helped along by AI to produce it — it can help with structure — and then validated by a human at the end. AI sits in the middle as a production tool, in my opinion, and it can do a great job — it can create graphics faster based on a prompt or art direction from a human. But if you take the human element out of either side of it, it creates something that doesn't resonate on a human level. It's all about that process, and where you use it. People are just fed up with agencies and people using it as a shortcut to having the original idea, or not validating it at the end.

Paul

From a marketing perspective — thinking about work I'm literally doing this week — I'm using it to help write a statement of work document, because that just needs to be a clear definition of the project, the facts, the flow of things, a clear piece of communication. But I wouldn't want it to write every single word of a proposal document, because that needs to land emotionally as well as factually. So it's about choosing the right tool for the right moment. But, changing the subject slightly — what does everyone think about Claude introducing watermarking? I don't know for sure what mechanism they're using, but my basic understanding is that they're putting something in place so that if you produce copy using Claude, it has some form of detectable watermark that a search engine like Google could recognise and deal with as it sees fit. I've read a lot of the response to it — there was a Reddit thread that immediately got loads of attention, and the overwhelming opinion was that it's a bad thing, because it could lead to people being accused of writing AI slop when in actual fact they might have only used it for a minor part of it. But I don't know, Pat — do you know the mechanism they're using?

Pat

Yeah, I've looked into it, and it's really interesting, because the mechanism they're using will survive copy and paste. With text, text is just text — there's no space for metadata to tell you where it came from or how it was written. They can't embed it in styles, there's no extra character set where they can sneak in extra bytes to identify the source. So I can only deduce it's down to the frequency or structure of the sentences, letters, or words they're using — so whenever Claude writes something, it'll write in a certain way that builds in almost a hidden key, in the letters and words. So if you copy and paste it around, which strips all the formatting out, it'll always carry that signature, as long as you don't change it yourself. From a cryptography perspective, that's definitely possible — that's probably how they're doing it. There's already loads of really good AI detection software out there today — I use one called GPTZero, which is bang on accurate. I can write a whole article by hand, have Claude edit a single sentence, and GPTZero will pick out that one sentence and say, "this is AI-generated." So it can be done at the moment, but it takes a large language model to assess and analyse it, so it's quite intensive for firms like LinkedIn to analyse posts en masse and flag them as AI. But these signatures Anthropic bake into their copy will be much easier to extract and validate. So you'll probably find, across the internet, a lot of tools and sites that publish copy will start flagging stuff that's AI-generated more frequently, and search engines will start to demote that content — prioritising human content, which will become more effective than it was previously. Which is good, because having something written by a human does typically indicate a high level of quality in terms of tone of voice — but coming back to your point before, Paul, about accessibility, and the fact these AI tools allow people who might otherwise struggle to write in a particular language, or struggle to write full stop, to articulate their thoughts — it'll penalise them. So it's certainly not a clear-cut win, it's a bit of a minefield.

Russ

Where does that leave teams really adopting AI at the moment — tone of voice skills, training their team on how to use Claude? Will it pick up on, say, an initial draft of an article written by a human, then processed through a tone-of-voice skill so it matches — which is a great thing to do for a brand? When you've got a small team, like us, or hundreds of people writing articles for a news site or a blog, a tone of voice is a complete game-changer — you get the right language, it's professional, it represents your values consistently, and you wouldn't be able to do that very well without a tone-of-voice skill. Are we saying this invisible watermark will apply to anything crunched through Claude using a tone-of-voice skill? Because that would mean everyone's at a disadvantage in terms of ranking — so it might not make a huge difference once that's released, in a year or two.

Paul

It's really interesting, because it raises a philosophical question, doesn't it — if you've got an article that's been produced off the back of a hundred people's research into a subject, and it pulls it all together into something genuinely useful, does the value of that article change if you know it's written by AI versus if you knew it was written by a human, even if it's exactly the same? Should one be penalised and not the other? It's a really interesting thing to think about. It reminds me — I probably bang on about this all the time — of Fake or Fortune on the BBC, where they find a piece of artwork someone's discovered, and they say, "my great-grandparent said this was a genuine Picasso," and they do all the research to figure it out. But the interesting bit is at the beginning of the show, the art expert says, "if this turns out to be genuine, it'll be worth in excess of £500,000; if it's not, and it's by someone else, it's still a nice painting, but it'll be worth maybe £500." You think, it's literally the same thing — but the story behind it changes the value. Is it the same for an article online? I'm not sure it is, but I think that's such an interesting question.

Pat

I think it's about volume — it's the perception that AI-created content is slop, because the vast majority of it is, because it's been generated with no thought, and the points it makes are mostly correct but pretty bland, quite repetitive. So if you spend a lot of time working with Claude to write a really insightful, original article that drives home some genuinely interesting and unique points — that article has a lot of value. But if Claude marks it as AI-generated, a lot of people will write it off as having no value, because of their perception that it's AI slop — when actually that article wouldn't be AI slop, it would be AI-written, high-quality content, which is possible — you just have to put the work in to do it. It comes down to strife, doesn't it — if you strive to do something, whether you use AI or not, that toil has value, and adds value to whatever you've done. If you can demonstrate that the end product involved toil in its production, it suddenly has intrinsic value to other people.

Russ

I was thinking of it in another way. If you contact a PR agency to do a press release on your behalf, they'll interview you, ask you a load of questions, and then send you back an article which is their own interpretation of what you said. It normally sounds a lot better than what you'd actually say, it normally has a quote you never said but you approve, and then it gets published in a magazine as advertorial, and a lot of people probably think it's part of the magazine — an authentic article. I'm not quite sure where I'm going with this, but I think I have a point — I think AI is almost like an invisible PR agency: you give it what you want to say, it turns it into an article, maybe gives you a quote. There's also the time aspect — producing a nicely written article is a lot quicker with AI, so why wouldn't you use it? I think it's going to be fascinating to see how people figure out how to take these articles apart and decide what's written by what. I think it's going to be a real challenge.

Paul

I think what you said earlier, Pat, is a really good way of looking at it — like imagining a single person writing everything online. That one person writes in the same way every time — they might be a really good communicator, but do you want your brand to use the same writer that everybody else uses? I think the answer's no, isn't it — you want a bit of personality.

Pat

Yeah. I think this tone-of-voice copywriting problem will get solved — it's such a big issue, such a massive problem for everyone, and people are talking about it so much that I guarantee the next generation of models will put a lot more focus on copywriting as a skill they want to excel at.

Amelia

I'm sure that's a big topic, and we'll come back to it. But we've discussed that you use AI a lot at Indulge — where's the line between AI you can trust with client-facing content, and AI you should only trust with a first draft? How does that work for you?

Paul

I think it's our skills library, and using agentic tools to identify repeatable processes and turn them into skills. A skill is essentially a packaged set of instructions for Claude — Claude has a library of skills devoted to particular things, so every time you need to do that thing, it can run the same process and produce the same result. Once you've got that in place, it's about having a human on the other end who knows what the inputs are, what to expect from the output, and how to use it. You don't necessarily run a Claude skill and use whatever it spits out straight away — you take the output and turn it into something that supports what you're doing. That's how we've put it into action.

Pat

Yeah — and just reading what comes out before it goes to the client, making sure you check the output, is just as important now as it was in the early days of ChatGPT. It doesn't get things wrong as much as it used to, but if you ask it to produce a statement of work — a document outlining a solution to a client's problem — if you don't read it properly, it will often sneak in extra features, propose something in a paragraph you didn't ask for, and if you don't read it in depth, you might end up committing to something you never actually agreed to, because it's been snuck into the statement of work. So for written copy like that, you have to read everything, and concentrate on it too. That's actually one of the main costs of using skills — it saves a lot of time, but creates a lot of work in checking what you're doing.

Paul

I put a lot of value on — because we've got Claude skills in place that can produce, say, a statement of work, it can produce a literally branded, finished document. But I actually get value from the process of taking that finished document and copying and pasting bits into my own version, tweaking it as I go, because I can then read and understand it in detail, and I feel like the process of moving something from one place to another is my chance to tweak it and give it my own spin. So stuff like that is helpful, I find.

Russ

Yeah, I'm the same. If I'm learning something new, I find writing notes helps, even if I've already got the text there — it's almost the same practice, running it through your own mind before you accept that it's correct. Especially in our world — when we're sending out proposals, when we need to communicate clearly to the team — everything we do not only needs to align with our values, but also needs to be factually correct. Whether it's HR, marketing messaging, a proposal going to a client — if AI has assisted in producing that document, everything needs to be validated by an expert. You wouldn't say, "we've got this statement-of-work skill, so anyone can produce one," and hand it to anyone on the team to send to a client, because that's potentially a one-way ticket to disaster — you could end up with something Claude's figured out in its own mind about what we should deliver and how much it should cost, and then you're in a situation where you don't have the skills to deliver it, or the budget isn't high enough. That could happen in any instance — it still needs to be validated by an expert. I'm the same as you, Paul — I distrust automating certain processes, because I want to be a hundred percent certain the output is correct, and I can add value too. So if it's something like a proposal, I'll copy it out into a document, change it, put in custom diagrams, and define the approach myself — not only because then I know it's correct, but because I can talk about it with people. That's one of the other problems — if it's just AI output, say an article, and that's why this podcast is so good — let's say you produce a load of thought-leadership articles for your website, and a prospective client or intermediary wants to speak to the person who wrote it, and jumps on a call — you're never going to close that deal if you don't know what you're talking about because AI generated it. From a proposal point of view, if I auto-created a proposal and then got on a call and needed to sell that idea, if I can't talk about it eloquently and articulate our expert opinion, it's going to be quite obviously AI, isn't it? I think that's going to catch a lot of people out in different ways.

Paul

I think the thing that helped me figure out how to use AI tools was realising it's not really a time-saving tool. It does save time for certain things, but fundamentally, time-saving is secondary to the deeper level of expertise it allows. Coming back to the reporting example — it doesn't save us any time in our reporting process, we still spend the same amount of time. But it greatly deepens our ability to dive into the numbers. There's less time in data gathering, because a lot of that can be automated, less time manually combing through a spreadsheet, but far more time spent understanding the numbers. I think when Alastair joined us on an earlier episode, he talked about experiencing development now as intense bursts of work — it was the same thing: it wasn't necessarily saving him time in building things, it was letting him achieve far more interesting things with the time he used. So I think if people forget about the time-saving element, which is what a lot of the media coverage focuses on, and instead understand it's a tool that can make what they're already doing far more interesting — that's probably the way to use it.

Pat

Yeah — I veer between being worried about what AI's going to do to jobs, and being pretty confident everyone will still have jobs in five, ten years. What we're definitely seeing — and told — is that we're busier than ever. We're not all finishing our day at 10am and going home; we're still working on the same clients, putting in the same amount of effort, but going deeper and producing more work of much better quality. I use the word "compromise" a lot — I think this is definitely the case for developers. Ask any developer if they've ever been completely happy with anything they've built, and very few will say yes. A developer's career is a journey of compromise — you're never able to build the thing you want to the level of quality you want, because you're constrained by time, budget, and the tools you have. What AI allows you to do as a developer is close the gap between the version you're shipping and the version you actually want — close the gap in terms of the number of compromises, and get closer to the level of quality you want it to have. All the unit tests, the automated browser testing, the right libraries, the right architecture — all that stuff that takes time and effort — you're able to do now, and produce software of much higher quality. That's what we're finding. I think a lot of jobs are like that — any job, product manager, marketing manager, project manager, CEO — they're all full of compromises day in, day out, always having to cut corners because the resource or funds aren't available. I think AI removes a lot of those compromises, and elevates everyone's output. So right now, I'm in the mindset that everyone's going to be busier than ever with AI, because it accelerates everything, and the output of everyone's jobs is going to get higher-quality and better. I think the economy will ultimately do quite well out of it long-term — short-term, I don't know what's going to happen, it's still quite unstable. It might be like the internet — the dot-com boom, a lot of excitement, it crashed, and then a few years later the growth went way beyond anyone's dreams. I think AI will probably follow a similar journey — we'll see a crash at some point, and then in ten years the world will be in a very different place.

Russ

I do think, just to add — the real exciting thing about AI for a lot of people is that it gives us time back. Time back for things, to do things we enjoy as humans — whether that's time with family, meditating, doing sport, or just doing the weekly shop at a different time of the week because you're busy and can leave your AI doing it. I think it'll be great to see, though it's difficult depending on your work ethic — I think it's important that if we embrace AI and it saves us time, that we live a slightly different life and reap the benefits, instead of just filling that time with using AI more. Because I think we'd be a much healthier society in the long term. That's my vision — the utopia AI might bring us. But in reality, humans have this intrinsic desire to work and work, and if we can produce more and more, we will. I hope we use it in a way that gives us the time back — it's the enabler, which technology should be — the enabler to improve our own lives. It'd be good to see that.

Pat

Here's a question for you off the back of that — I absolutely agree, it offers a great opportunity to step back and engineer more time into your day. I don't think many humans would do that, though, because I think humans like to work — we fill our days with it, and if an agent's doing one thing and you've got spare time, you just spin up another agent to do something else. But here's an interesting question: if you've got an agent running, and you say to your employer, "my agent's going to be running for the rest of the day fixing this, I'm just going to go home now" — do you think the employer would say, "yeah sure, you're still working because your agent's running," or would they say, "you'd better book a half day of leave, mate, because you're not just wandering off"? I think most employers would fall into the second bucket, wouldn't they?

Russ

It raises a question about billing too, on our side — because you engineer this amazing prompt, and you're talking about one prompt building a website, based maybe on a large engineered prompt for the context and strategy — you sit there, you've got ten websites being built in a day, but you're billing... they're all running and building a site. Obviously we add our value in different ways, but it turns the whole business model on its head, in almost every industry, in some ways. You need to really think about where you add value, where human value comes in.

Paul

But thinking back to the context of a lot of the marketing teams we work with — I don't think I've ever spoken to anyone who's said, "yeah, we've got the size of team we need, we're on top of everything." Everybody you speak to is like, "we don't have enough people, we've got too much work to do, we're just keeping up." I think the availability of tools to find efficiencies is just allowing teams to catch up a little. But before you know it, the natural way commercial organisations work is that they fill gaps — if there's a gap, they'll fill it with something. I think of it in a similar way to the trend towards a four-day working week, which I like the idea of — I think it'd be healthy for society. But then I keep thinking, if 80% of businesses went down to a four-day week, you'd still get the 20% thinking, "that gives me an extra day to steal a march on my competitors and do extra stuff." Because of that, the incentive system in place means everyone's in a kind of race to the bottom — you just keep putting more and more in. I think it'll be the same with AI — every bit of efficiency it finds will be filled with something else. People want more of a good thing.

Pat

That's my personal experience with AI too — rather than running one agent and taking time out, if I start an agent and I'm waiting on it, I'll spin up another one to do something else. Before I go into a meeting, I'll make sure I've got a couple of agents running, so that when I'm in the meeting, stuff's getting done. When I come out, I'm checking on the agents. Honestly, my work habits these days are probably unhealthier compared to pre-AI, because I'm running and checking agents all hours. For me, work-life balance has gone in the wrong direction with AI.

Amelia

So it might take a bit more time to get to that work-life balance Russ was talking about.

Pat

Yeah, I think that's a real concern, probably across the economy in general.

Amelia

Sounds like it might need to be a whole other episode.

Russ

I think so — sorry, I've taken this a bit off-topic. But I think the time-saving within teams, especially marketing teams, is a huge opportunity — you can get more done with the team you have, using AI.

Paul

See, my view's slightly different — I think you should ignore the time-saving element, and instead see it as a really powerful tool. Reporting is the easy example for marketing — a really simple one: if a marketing team wants to get up and running with something different, export all your Google Search Console data, run it through an agent, and ask it to spot trends, every month. You'll find interesting things you'd never have found before. It's not about saving time, it's about finding something better.

Amelia

I'm conscious of time, because we've had so much to discuss today, and we've gone off on a few tangents which have been really interesting and helpful. But a couple of final questions — starting with: what would you say to a marketing team that's been burnt by bad AI output? What's a realistic, low-risk way back in?

Paul

I think the example I just gave — exporting data and funnelling it through, and forgetting about trying to do anything amazing — see it as a really good data analyst you've got on your team now, and do that. The more comfortable you get with that, the further you push it.

Russ

I'd make the point about using it as a production tool — like any other technology — having a human at the start coming up with the idea and the strategy, brainstorming, and then a human at the end validating it, making sure it'll resonate with people. It always sits in the middle of that process, for me, and I find that works. If you trust it to do the whole thing end to end, it's almost certainly going to go wrong.

Pat

Yeah, and I'd say approach it in an organised way. The chances are that team got burned because they just smashed something out through ChatGPT without really understanding what it's doing, and what they put out contained hallucinated facts, or was written badly, or the image had someone with six fingers — all those random AI things. But take a step back, be organised, do some training at the start, get your head around what these tools are good at and what they're not, get under the skin of them in a bit more depth, and you'll soon find some really good efficiencies. Use agentic tools like Cowork.

Amelia

And finally — I know we've touched on the future of AI-assisted marketing, but what does that look like for financial services in about a year, and what does trust end up meaning by then?

Paul

Coming back to why we started talking about this — it was prompted by meeting people who distrust it, either because they're scared of technology generally, or because they just don't know what it can do. I think now's the time for all teams to understand that agentic AI is where it's at — forget the copywriting element, forget all the hacks you've heard about, and remind yourself we have access to a really smart resource that we can give relatively simple things to do, and it'll do them much better than you'd probably anticipate. I think teams have to really embrace that. At the organisational level, large finance firms who are heavily regulated are going to have to think carefully about how they roll this out — they can't be letting people use different tools willy-nilly — but they should give their teams access to proper agentic tools with proper guardrails in place, and let them know what they're good for. That's so important.

Russ

Yeah, I find it a bit like having a supercomputer at my fingertips, with a whole army of really skilled assistants. If you use that in the right way, you can do some amazing things — whether that's helping with copywriting, drawing diagrams, coding, or even a bit of design. It can do anything. I think, just experiment with it, try using it for different things — it's an invaluable tool. I honestly don't know how we used to operate without it.

Amelia

I think that's a good place to leave it, but I'm sure we'll come back to lots of these points in the coming weeks and months. Shall we end with a bit of jargon busters?

Paul

Yeah, we've had a good run, haven't we?

Amelia

Not been bad. So if you've never caught this before — I've got a list of industry terms, and each week I put the guys to the test to see if they can define it. This week's term is zero-sum game. What do we reckon?

Pat

I've heard the phrase, but I've got no idea what it means, so I'm going to guess it's a kind of negotiation between two parties, or where two competitors are going head-to-head for a particular market, but there's actually nothing to gain for either of them, because they're both throwing so much money at each other that neither party wins. Perhaps.

Amelia

Interesting. Russ, any thoughts?

Russ

Did you say zero-sum game, or zero-sum gain? Game — okay. I mean, a lot of these have been on the topic of investment. Zero-sum game — why would you invest in something where you had no return, or gain? Perhaps it's investing in something for different reasons than gaining a return back — perhaps the values of the business, or because of sustainability, investing in that business gives you something back other than a capital return. That's as good as I've got it, I think.

Paul

Does it describe a scenario where the net result is zero? The economy, for instance, is zero-sum in the sense that if I give someone £20, that's plus 20 to this person, minus 20 to the other — the sum is zero. It's the same with all money — every bit of credit has an equal bit of debit somewhere, so the sum is always zero. That's basically my understanding of it.

Amelia

Interesting. So the official definition is: a situation where one person's gain is exactly equal to another's loss.

Paul

Yep.

Amelia

So you were kind of on the right lines there, Paul.

Pat

So trading's a zero-sum game, then?

Amelia

Yeah. So we will of course have more of this next week — we'll be back next Friday with a brand new episode. Thanks so much for listening, and we'll see you next week.

Paul

Cheers!

Pat

Cheers, everyone.

Russ

See ya.