Latest / The STANDOUT Agency Podcast / EPISODE 4: Making AI Work For Your Agency
Transcript
- 0:02SPEAKER_01: Hello, welcome to The STANDOUT Agency Podcast.
- 0:05SPEAKER_01: This is episode four and it is titled Making AI Work for Your Agency.
- 0:11SPEAKER_01: My name is Callum Healy.
- 0:12SPEAKER_01: I'm an AI consultant for marketing agencies,
- 0:14SPEAKER_01: helping them effectively embed and adopt AI into their agency.
- 0:20SPEAKER_00: And I'm Gareth Healey.
- 0:21SPEAKER_00: I'm an ex-agency owner,
- 0:24SPEAKER_00: author of the agency book, STANDOUT OR DIE, and now an agency growth advisor.
- 0:30SPEAKER_00: And together we are Agents of Change, which is also the URL of our website.
- 0:35SPEAKER_01: Absolutely, and that's become quite a nice podcast opener.
- 0:39SPEAKER_00: It has, yeah.
- 0:39SPEAKER_00: We're going to try and stick
- 0:40SPEAKER_01: to that, aren't we?
- 0:42SPEAKER_00: Weareagentsofchange.com if you want to learn more about us.
- 0:44SPEAKER_00: So, Callum, what's the titles there?
- 0:47SPEAKER_00: Yep.
- 0:51SPEAKER_00: The first question for me is to go back a stage and think about how,
- 0:58SPEAKER_00: before you started working with agencies,
- 1:01SPEAKER_00: you started to look at how AI works and got really interested in AI and started
- 1:06SPEAKER_00: to look at how it works and how it thinks.
- 1:09SPEAKER_00: What did you learn at that point?
- 1:12SPEAKER_01: Context, ultimately, is something that is very relevant to my studies.
- 1:19SPEAKER_01: Throughout my education journey,
- 1:23SPEAKER_01: I have a psychology degree and studied that for five years,
- 1:28SPEAKER_01: including sixth form, and context is a massive part of how humans learn.
- 1:33SPEAKER_01: But I was immensely fascinated by artificial intelligence,
- 1:38SPEAKER_01: because I'd spent so long studying the human mind,
- 1:41SPEAKER_01: and that's kind of Where I fell in love with using AI as almost my thinking partner
- 1:47SPEAKER_01: and something that's just become part of my daily work as it has with yourself
- 1:51SPEAKER_01: and I'm sure many people watching as they've searched up the keyword of AI of course
- 1:57SPEAKER_01: and there are two main topics of AI and machine learning is one of them,
- 2:03SPEAKER_01: deep learning is another and it's ultimately all about the kind of evolution and
- 2:08SPEAKER_01: the study of how AI comes to think.
- 2:13SPEAKER_00: There are many uses of AI of course,
- 2:15SPEAKER_00: but if we just go to our base use and what many people will be most familiar with,
- 2:22SPEAKER_00: when you type a question into the chatbot, what is actually happening?
- 2:30SPEAKER_01: Many of you may know that computers fundamentally are processing information
- 2:36SPEAKER_01: and ultimately the whole way that they operate is through binary code.
- 2:42SPEAKER_01: Which is ones and zeros.
- 2:43SPEAKER_01: Whenever you look at anything on a computer, it's just a massive string of ones and zeros.
- 2:48SPEAKER_01: And like you would learn a new language,
- 2:51SPEAKER_01: you know, we speak English if we were to learn French or Chinese.
- 2:56SPEAKER_01: It's all about how we translate our language to theirs.
- 3:02SPEAKER_01: And there is a study called Natural Language Processing,
- 3:07SPEAKER_01: which is all about how we can bridge the gap between Human language
- 3:14SPEAKER_01: and computer language.
- 3:15SPEAKER_01: And so this starts with ultimately words becoming numbers.
- 3:20SPEAKER_01: And there are certain processes of this.
- 3:23SPEAKER_01: Vectorization was an earlier one.
- 3:26SPEAKER_01: We now have things like embeddings and it is a very long and complex process,
- 3:32SPEAKER_01: but ultimately it's all about converting words to numbers and semantic meaning is
- 3:40SPEAKER_01: a factor of that.
- 3:43SPEAKER_01: There are things called embeddings,
- 3:46SPEAKER_01: and put it very simply,
- 3:49SPEAKER_01: words with similar meanings will have similar vectors,
- 3:53SPEAKER_01: and vectors are strings of numbers that translate to words,
- 3:58SPEAKER_01: and so they sit together in some infinitely multi-dimensional space,
- 4:04SPEAKER_01: where you see 3D,
- 4:06SPEAKER_01: 4D, 5D, this is like 1000D, completely It's a big map you see there.
- 4:12SPEAKER_01: Yeah, and it has to have this Level of complexity and these amounts of dimensions
- 4:19SPEAKER_01: because there are so many words and so many meanings between different words.
- 4:23SPEAKER_01: Every word has a relation to every single different word and it's all
- 4:27SPEAKER_01: about how they interact with each other.
- 4:29SPEAKER_01: Very simply AI has siphoned as much data as it can possibly handle and it looks
- 4:35SPEAKER_01: at all of the ways that words are usually interacting with each other
- 4:40SPEAKER_01: on say web pages,
- 4:43SPEAKER_01: Massive data sets which they're trained on and they'll realise that the word,
- 4:48SPEAKER_01: say, king is very similar to the word queen.
- 4:51SPEAKER_01: Those two come up in context a lot.
- 4:55SPEAKER_01: There's an association.
- 4:56SPEAKER_01: There's an association.
- 4:58SPEAKER_01: So it might, king might have an association with fantasy or knight or crown, things like this.
- 5:04SPEAKER_01: And they learn over time by taking in all of these data sets,
- 5:10SPEAKER_01: how the words relate to each other.
- 5:12SPEAKER_01: And so, a good way of describing word to vec,
- 5:16SPEAKER_01: and there's a lot of jargon here,
- 5:17SPEAKER_01: but I'll try to describe it as simply as I can,
- 5:20SPEAKER_01: is that you can do kind of arithmetic with these words,
- 5:26SPEAKER_01: because they're numbers, they're vectors.
- 5:28SPEAKER_01: I have an equation which I learned,
- 5:30SPEAKER_01: which is simply,
- 5:31SPEAKER_01: if you take the context of king,
- 5:32SPEAKER_01: and king has a vector,
- 5:34SPEAKER_01: and you take away the Vector for man,
- 5:39SPEAKER_01: these are all numbers but in our language it's king and man,
- 5:42SPEAKER_01: take away the vector for man,
- 5:44SPEAKER_01: you add the vector for woman,
- 5:47SPEAKER_01: so you take away the masculine element of this monarch,
- 5:51SPEAKER_01: you then add a feminine sort of context of the monarch,
- 5:56SPEAKER_01: that equation would equal queen.
- 5:58SPEAKER_01: So it connects to queen on that map.
- 6:00SPEAKER_01: So if you see these on a big embedding space you'll see how it's all about space and
- 6:05SPEAKER_01: so King and Queen will be close.
- 6:07SPEAKER_01: Man and woman will be close.
- 6:10SPEAKER_01: Differences.
- 6:11SPEAKER_01: It's all very interesting and complex stuff.
- 6:14SPEAKER_01: Is that why we hear a lot about neural networks?
- 6:18SPEAKER_01: Neural networks, yes.
- 6:20SPEAKER_01: Ultimately, I was surprised by this.
- 6:25SPEAKER_01: We modelled the AI brain after the human brain to some degree because that's
- 6:29SPEAKER_01: the only level of intelligence we knew.
- 6:31SPEAKER_01: All of the animals have similar brains to us as well.
- 6:36SPEAKER_01: We have neurons and AI has artificial neurons.
- 6:39SPEAKER_01: There are things called artificial neural networks,
- 6:42SPEAKER_01: which is the very start of kind of these AI algorithms,
- 6:47SPEAKER_01: this thing, recurrent neural networks.
- 6:49SPEAKER_01: Now they're transformers.
- 6:51SPEAKER_01: So if you see GPT, the actual acronym of GPT is Generative Pre-trained Transformer.
- 6:59SPEAKER_01: This was the revelation,
- 7:00SPEAKER_01: I believe in 2017 by Google researchers and it made it so LLMs,
- 7:07SPEAKER_01: Large Language Models, as we interact with today mainly, it's allowed them to exist.
- 7:13SPEAKER_01: And these have hundreds of billions of what's called parameters,
- 7:19SPEAKER_01: these variables,
- 7:20SPEAKER_01: and you can almost imagine these as little dials, as weights and as biases.
- 7:27SPEAKER_01: And we gave AI the algorithms to learn.
- 7:32SPEAKER_01: Very simply,
- 7:34SPEAKER_01: It fails over and over and over again until it finds the correct tweaking
- 7:39SPEAKER_01: of all these billions of dials to get from the input to the output to get
- 7:44SPEAKER_01: the correct word.
- 7:45SPEAKER_01: And so we can't exactly map how AI gets from the input to the output in the correct word.
- 7:54SPEAKER_01: And fundamentally, AI is based on probabilities.
- 7:57SPEAKER_01: So when you see these machines,
- 7:59SPEAKER_01: They are going from one word to another, over and over and over again.
- 8:05SPEAKER_01: And the highest probability in the calculations,
- 8:07SPEAKER_01: these are trained by calculus,
- 8:09SPEAKER_01: all these mathematical equations,
- 8:12SPEAKER_00: they happen again, then it predicts the next word, then it predicts the next word.
- 8:18SPEAKER_00: So when anybody's using,
- 8:19SPEAKER_00: we're talking about agencies here of course,
- 8:21SPEAKER_00: but when anybody's using AI in the chatbot form,
- 8:25SPEAKER_00: It's not really understanding
- 8:26SPEAKER_01: what we're saying, isn't it?
- 8:27SPEAKER_01: It's predicting the next,
- 8:29SPEAKER_00: well, not just the next word, it's predicting the next set of words, isn't it?
- 8:33SPEAKER_00: Which is a fact,
- 8:36SPEAKER_00: but again,
- 8:37SPEAKER_00: another fact that I just find incredible,
- 8:38SPEAKER_00: because it doesn't feel like that when you use it,
- 8:40SPEAKER_00: because it doesn't feel as if it's mechanically predicting the next words.
- 8:45SPEAKER_00: It feels like it's actually,
- 8:47SPEAKER_00: you're actually having a conversation, going backwards and forwards.
- 8:50SPEAKER_00: And sometimes, you know, I almost feel human.
- 8:53SPEAKER_01: Yeah, exactly and so instead of AI knowing,
- 8:57SPEAKER_01: say if you ask it to generate you an essay,
- 9:01SPEAKER_01: it's not going to understand all of the essay and all the words and
- 9:05SPEAKER_01: then just give it to you all at once,
- 9:07SPEAKER_01: it's actually just very quickly going,
- 9:10SPEAKER_01: Here's the first word,
- 9:11SPEAKER_01: here's the next word, here's the next word, here's the next word.
- 9:13SPEAKER_01: That's based on the training.
- 9:15SPEAKER_01: If you give it a role,
- 9:16SPEAKER_01: writing Shakespearean,
- 9:17SPEAKER_01: it's going to use the Shakespearean data set which it was trained on.
- 9:21SPEAKER_01: And all of this is well and good,
- 9:23SPEAKER_01: and I'm sure some of you will be interested and some of you will be thinking,
- 9:27SPEAKER_01: how can it help my agency?
- 9:28SPEAKER_01: And the reason it's important to understand this is because...
- 9:33SPEAKER_01: It doesn't fundamentally understand the words that we're giving it and the outputs
- 9:39SPEAKER_01: that it gives us but it's just very good at predicting
- 9:45SPEAKER_01: and it's got these mathematical formulas which allow it to give coherent thought
- 9:51SPEAKER_00: but it isn't trained on your
- 9:54SPEAKER_01: agency and so a human with inference will understand
- 10:01SPEAKER_01: the task much better only if they had a little bit of context about your agency.
- 10:05SPEAKER_01: This isn't how machines think.
- 10:07SPEAKER_01: They need to be trained on your
- 10:10SPEAKER_00: data as well.
- 10:11SPEAKER_00: And it still gets,
- 10:12SPEAKER_00: even though I mean I didn't say how clever it is and how amazing,
- 10:14SPEAKER_00: which of course is true, but it still gets stuff wrong as well, doesn't it?
- 10:17SPEAKER_00: Exactly.
- 10:18SPEAKER_01: Hallucinates, just gets facts wrong.
- 10:23SPEAKER_01: Ultimately it's a genius because it's very fast,
- 10:25SPEAKER_01: it's ingested all these facts and it's very good at retaining that information.
- 10:30SPEAKER_01: These algorithms are very fascinating but it's a generalist and they made it that way.
- 10:35SPEAKER_01: All of these incredible corporations and Gemini
- 10:40SPEAKER_00: with Google and OpenAI,
- 10:42SPEAKER_00: ChatGPT,
- 10:43SPEAKER_00: Claude as we're favourable of right now,
- 10:47SPEAKER_01: they wanted to make a product that anyone could use but A core philosophy
- 10:54SPEAKER_01: that we have is that it's good to specialise, to niche, to focus.
- 11:00SPEAKER_01: And so we want to use AI in a way that makes it a specialist on your agency.
- 11:10SPEAKER_01: And this
- 11:13SPEAKER_01: corresponds greatly to the architecture of how these machines think.
- 11:18SPEAKER_01: Because once you understand They are prediction machines, they need data to use these formulas.
- 11:25SPEAKER_01: It compels you to give it as much context of yours as possible at every opportunity.
- 11:32SPEAKER_00: So you talked about context there and obviously that's important
- 11:37SPEAKER_00: but it just strikes me what's the difference between If you are running an agency,
- 11:43SPEAKER_00: which is the business that we are in and focused on,
- 11:47SPEAKER_00: what is the difference between giving your AI your agency context,
- 11:52SPEAKER_00: or maybe even your client context, and just using it out of the box?
- 11:57SPEAKER_00: Because as we have been talking, these things are immensely clever.
- 11:59SPEAKER_00: Is this context really needed?
- 12:04SPEAKER_01: It's a testament to how intelligent these machines are.
- 12:10SPEAKER_01: It's artificial intelligence.
- 12:12SPEAKER_01: It almost doesn't do it justice to calling it artificial because it's so clever.
- 12:18SPEAKER_01: It's so good at ingesting so much data,
- 12:20SPEAKER_01: I'm talking trillions of different data points,
- 12:23SPEAKER_01: that it has a great knowledge on most things and ultimately it's got knowledge
- 12:29SPEAKER_01: on what us humans have put on the internet and what we have given it.
- 12:35SPEAKER_01: It's not creating anything new,
- 12:36SPEAKER_01: it's just learning what's already out there but can read and ingest things quicker
- 12:41SPEAKER_01: than anyone.
- 12:43SPEAKER_01: The problem with that is it can answer questions fluently and confidently,
- 12:47SPEAKER_01: even if it's the wrong answer or a suboptimal answer.
- 12:52SPEAKER_01: The way these machines work in general is the regression to the mean.
- 12:57SPEAKER_01: We're talking about probabilities earlier,
- 12:59SPEAKER_01: it's all about predicting what's the most likely next answer, it's the average.
- 13:04SPEAKER_01: Things that if someone's a personal trainer,
- 13:08SPEAKER_01: Their advice would be much better than AI's advice,
- 13:12SPEAKER_01: but if they use AI, it would drag them down to the average.
- 13:15SPEAKER_01: Someone who has no idea about fitness can greatly bolster their knowledge by AI,
- 13:21SPEAKER_01: but it's meant to be
- 13:23SPEAKER_00: medium, mediocre, if you will.
- 13:25SPEAKER_00: We're starting to realise that in some recent meetings I've had on
- 13:31SPEAKER_00: the creative generation with AI.
- 13:35SPEAKER_00: I've been speaking to a number of agencies recently that Not themselves,
- 13:40SPEAKER_00: but the market is realising that whilst generating creative concepts,
- 13:45SPEAKER_00: visuals, copy,
- 13:47SPEAKER_00: it's brilliant art,
- 13:49SPEAKER_00: but it does create average ultimately because it's bringing in everything
- 13:54SPEAKER_00: and creating an average.
- 13:55SPEAKER_00: So there's certain agencies I'm talking to that are actually not stopping using AI,
- 14:00SPEAKER_00: but only using it to a certain point and using it with More human input.
- 14:06SPEAKER_01: It's very easy to spot a chat GPT generated brochure, I've noticed.
- 14:11SPEAKER_00: It's pretty much
- 14:12SPEAKER_01: everywhere.
- 14:13SPEAKER_01: And they look great,
- 14:14SPEAKER_01: you know,
- 14:14SPEAKER_01: it would have taken a graphic designer a lot of skill to do that in the past.
- 14:18SPEAKER_01: But they do that because it's the most likely next answer.
- 14:22SPEAKER_01: This is why they're so similar, because that's the average that they're trained on.
- 14:27SPEAKER_01: A little tip,
- 14:28SPEAKER_01: there is a variable called temperature,
- 14:31SPEAKER_01: which basically makes the probabilities more random.
- 14:35SPEAKER_01: Kind of makes all the calculus, the formulas that these are trained on.
- 14:39SPEAKER_01: Instead of being the most likely next answer,
- 14:41SPEAKER_01: the probabilities can get differentiated in the sense.
- 14:46SPEAKER_01: So if you say turn the temperature up to the max and you're generating a visual,
- 14:51SPEAKER_01: it'll give you something vastly different because it's not going on average anymore.
- 14:55SPEAKER_01: That's a little tip.
- 14:56SPEAKER_01: So in this case,
- 14:58SPEAKER_01: to go back to your question,
- 15:00SPEAKER_01: If I was to generate a proposal or an agency was to generate a proposal for
- 15:05SPEAKER_01: a new potential client, it would invent the whole thing.
- 15:09SPEAKER_01: It would invent the price,
- 15:11SPEAKER_01: the deliverables,
- 15:12SPEAKER_01: the time frame,
- 15:16SPEAKER_01: the certain sort of quality markers that an agency might have, the guarantees.
- 15:24SPEAKER_00: By pulling all those proposals in from all the resources you've got.
- 15:29SPEAKER_00: You could even say make me a McKinsey
- 15:32SPEAKER_01: or Bain proposal and it'll take the logic
- 15:35SPEAKER_01: and maybe there's books online or there's articles and that's all well and good
- 15:39SPEAKER_01: but it isn't yours.
- 15:42SPEAKER_01: You know,
- 15:43SPEAKER_01: we're mainly dealing with agencies and helping agencies that have been in business
- 15:48SPEAKER_01: for a long time and they've already got established processes and established offers
- 15:53SPEAKER_01: that work so well.
- 15:54SPEAKER_01: We don't want AI inventing the wheel, nor should it.
- 15:58SPEAKER_01: I think it's all about making whatever already works so well
- 16:02SPEAKER_01: for your agency
- 16:04SPEAKER_01: even quicker and therefore you can spend a little bit more time
- 16:08SPEAKER_01: on making the quality good as well.
- 16:11SPEAKER_01: It invents the whole thing in a proposal in this case.
- 16:14SPEAKER_01: We are on Claude,
- 16:16SPEAKER_01: and I love co-work, as we'll explain a little bit about a bit later in the episode.
- 16:22SPEAKER_01: I have basically let it ingest, or ingested myself, our whole framework.
- 16:28SPEAKER_01: So the STANDOUT system, which we covered in episode one.
- 16:32SPEAKER_01: So it knows I do an AI audit at this kind of When I start to work with agencies,
- 16:38SPEAKER_01: it's...
- 16:39SPEAKER_01: And so it knows the price,
- 16:41SPEAKER_01: it knows that I do a 90-minute diagnostic of the pillars and then
- 16:45SPEAKER_01: a 90-minute report delivery,
- 16:47SPEAKER_01: I've got a separate presentation deck,
- 16:50SPEAKER_01: separate large comprehensive report, all of these details that I have made...
- 16:57SPEAKER_01: It understands and it's every time I interact with Claude and I talk about my offer,
- 17:04SPEAKER_01: it will pull up that file that's been saved by me,
- 17:07SPEAKER_01: or multiple files,
- 17:09SPEAKER_01: and it will understand exactly what my offer is because I'm the one who has given it
- 17:13SPEAKER_01: that extra data.
- 17:14SPEAKER_01: This is called RAG fundamentally, it's Retrieval Augmented Generation.
- 17:19SPEAKER_01: Which basically means AI has had all of this training with all of this data,
- 17:23SPEAKER_01: you're giving it more data and in this case it's your data,
- 17:27SPEAKER_01: it's the ones
- 17:29SPEAKER_00: that's important.
- 17:31SPEAKER_00: Your proposals, agencies do proposals of various shapes and sizes, so I guess
- 17:38SPEAKER_00: Are we saying then,
- 17:39SPEAKER_00: for if agencies listening out,
- 17:40SPEAKER_00: that we may upload your bank of the last 12 months proposals to AI and
- 17:48SPEAKER_00: to whatever AI you're using,
- 17:52SPEAKER_00: upload the files,
- 17:53SPEAKER_00: upload the credentials deck, upload the proposals, it knows more about you.
- 17:57SPEAKER_00: Is that a job done?
- 18:00SPEAKER_01: Yeah, I'd say there's multiple steps,
- 18:03SPEAKER_01: of course,
- 18:03SPEAKER_01: and there's ways of doing it,
- 18:05SPEAKER_01: but in general,
- 18:08SPEAKER_01: People have caught on very quickly to the fact that they should bring
- 18:11SPEAKER_01: in external files into chat histories.
- 18:14SPEAKER_01: But the problem with most agencies that I see,
- 18:17SPEAKER_01: and this is particularly applicable with large teams,
- 18:21SPEAKER_01: is that the context dies in the chat.
- 18:24SPEAKER_01: So, especially if everyone's on personal plans,
- 18:28SPEAKER_01: they might have a great amount of context,
- 18:31SPEAKER_01: you know,
- 18:31SPEAKER_01: they might have ingested their certain offer of a kind that they're doing a proposal
- 18:36SPEAKER_01: for and then their AI gives better outputs because it has that extra information.
- 18:41SPEAKER_01: That can't really be shared across the whole team.
- 18:44SPEAKER_01: And it forgets, doesn't it?
- 18:45SPEAKER_01: And it forgets that.
- 18:46SPEAKER_01: Yeah.
- 18:47SPEAKER_01: So the idea is to have an agency plan,
- 18:49SPEAKER_01: of course,
- 18:50SPEAKER_01: which I'll always recommend and that has implications with governance,
- 18:54SPEAKER_01: as we'll discuss a little bit later.
- 18:59SPEAKER_01: Explain to people what that means.
- 19:00SPEAKER_01: So an agency plan is a central file?
- 19:03SPEAKER_01: Oh, yeah.
- 19:03SPEAKER_01: So sorry,
- 19:04SPEAKER_01: the agency plan is purely it's a The company bought an owned subscription to
- 19:12SPEAKER_01: the tool.
- 19:15SPEAKER_01: We'll discuss with contracts, but in general, everyone wants to be under the same hood.
- 19:21SPEAKER_00: Of course, we don't want people using personal plans.
- 19:24SPEAKER_00: That's shadow AI,
- 19:26SPEAKER_00: especially free
- 19:26SPEAKER_01: plans because you might put sensitive data,
- 19:31SPEAKER_01: model trainings, turned on, ingest that context.
- 19:35SPEAKER_01: There's been some horror stories with that.
- 19:38SPEAKER_00: Step number one, invest,
- 19:41SPEAKER_00: and it is an investment,
- 19:42SPEAKER_00: because particularly if you're a larger business,
- 19:44SPEAKER_00: this is a lot of money,
- 19:46SPEAKER_00: but you've got to invest in the business plans and not have people using their own
- 19:54SPEAKER_00: Plans and not having people using their own favorite AI chat that they're using at home.
- 20:02SPEAKER_01: Completely.
- 20:02SPEAKER_01: For all sorts of reasons.
- 20:04SPEAKER_00: To get this right, but obviously for data protection and all that sort of stuff.
- 20:08SPEAKER_01: We have gone to Claude.
- 20:09SPEAKER_01: I mentioned Cowork is an amazing environment.
- 20:12SPEAKER_01: It's almost like the difference in how I explain the different types of Claude.
- 20:19SPEAKER_01: There's three.
- 20:20SPEAKER_01: There's chat, Cowork and Code.
- 20:23SPEAKER_01: Chat is your search engine.
- 20:25SPEAKER_01: Cowork is your employee, because this is my preferred environment to ingest all this content.
- 20:30SPEAKER_01: So you're saying, how does an agency go on to do that?
- 20:34SPEAKER_01: And just for the record, code is more like the developer in this case, but Cowork, the employee.
- 20:40SPEAKER_00: How do you go about doing that?
- 20:42SPEAKER_00: Forget the business plan first.
- 20:45SPEAKER_00: But then a central file, a central information repository.
- 20:52SPEAKER_01: Exactly,
- 20:53SPEAKER_01: and in psychology,
- 20:56SPEAKER_01: in my degree, I did research methods, and that's all about data analysis.
- 21:00SPEAKER_01: Fundamentally, there's two different types of data.
- 21:03SPEAKER_01: Quantitative,
- 21:05SPEAKER_01: which is all about the Files that exist that you can quantify them,
- 21:11SPEAKER_01: it's frameworks,
- 21:12SPEAKER_01: it's files,
- 21:14SPEAKER_01: anything written down,
- 21:15SPEAKER_01: numbers,
- 21:16SPEAKER_01: definitely that's the main kind of makeup of quantitative data
- 21:21SPEAKER_01: because you quantify it, of course.
- 21:23SPEAKER_01: And then qualitative,
- 21:24SPEAKER_01: so there's nuance that you get through conversations and perhaps
- 21:28SPEAKER_00: it's the knowledge
- 21:28SPEAKER_00: and IP trapped in people's heads.
- 21:30SPEAKER_00: So just examples of that for an agency might be the quantitative might be your
- 21:37SPEAKER_00: If you have rate cards,
- 21:39SPEAKER_00: your pricing mechanism,
- 21:42SPEAKER_00: your capacity,
- 21:45SPEAKER_00: a number of people,
- 21:49SPEAKER_00: and results,
- 21:50SPEAKER_00: case studies of work you have done in the past where you can prove ROI,
- 21:56SPEAKER_00: quantitative stuff,
- 21:57SPEAKER_00: and qualitative might be,
- 21:59SPEAKER_00: well many examples,
- 22:02SPEAKER_00: The obvious one is the language,
- 22:05SPEAKER_00: the tone of voice that the agency uses,
- 22:08SPEAKER_00: words and phrases and language that you,
- 22:12SPEAKER_00: as your agency brand,
- 22:14SPEAKER_00: want to use and potentially elists Weareagentsof, Words,
- 22:18SPEAKER_00: Phrases,
- 22:21SPEAKER_00: Tone of Voice that you want to avoid depending on your agency,
- 22:26SPEAKER_00: depending on your own personal view, the market that you're selling to.
- 22:31SPEAKER_00: You've got to really define that as you would do
- 22:33SPEAKER_00: in any marketing communication Brief and brand communication brief,
- 22:38SPEAKER_00: but this is about cementing this
- 22:40SPEAKER_01: into the AI so it knows.
- 22:42SPEAKER_01: Yeah,
- 22:43SPEAKER_01: and brand voice and tone of voice is a good example because in every single bit
- 22:48SPEAKER_01: of context and area of your agency that you want to ingest,
- 22:52SPEAKER_01: it's going to be both quantitative and qualitative.
- 22:55SPEAKER_01: So brand voice is one that's mainly qualitative because it's picked up
- 22:59SPEAKER_01: by people's natural way of speaking and interacting with clients.
- 23:04SPEAKER_01: I worked with an agency recently where as part of the quantitative data we got
- 23:09SPEAKER_01: the BBC English grammar guidelines and ingested that and
- 23:14SPEAKER_01: that was alongside more quantitative data in the form
- 23:18SPEAKER_01: of this agency's past LinkedIn posts,
- 23:22SPEAKER_01: website copy,
- 23:24SPEAKER_01: perhaps it's slide decks that they've had internally in the past that they've made
- 23:28SPEAKER_01: which is like the kind of source of truth which is already out there.
- 23:33SPEAKER_01: We did that and we also did,
- 23:35SPEAKER_01: we had an interview where we actually just spoke or it's from transcripts
- 23:40SPEAKER_01: from podcasts and things like this where it isn't exactly planned but it captures
- 23:45SPEAKER_01: that kind of nuance that the files and the quantitative data wouldn't get.
- 23:50SPEAKER_01: So it's about bringing all of that together and the way that is the best
- 23:55SPEAKER_01: to make these files is to start in one chat,
- 23:59SPEAKER_01: so I'd do this in co-work,
- 24:00SPEAKER_01: Bring in all of those files,
- 24:03SPEAKER_01: put them in the chat history,
- 24:05SPEAKER_01: ones that would usually die if you were to move on,
- 24:08SPEAKER_01: but the idea is we're going to put all this context into a chat and then take it out
- 24:13SPEAKER_01: and get it to remember that in the form of a file.
- 24:17SPEAKER_01: Capture these podcast transcripts and meeting transcripts where we've done tone
- 24:21SPEAKER_01: of voice questions,
- 24:22SPEAKER_01: capture the BBC guidelines, website, LinkedIn, that's all in there.
- 24:26SPEAKER_01: Then we go, okay AI, I'd like you to make me an MD file.
- 24:31SPEAKER_01: This is the preferred one.
- 24:32SPEAKER_01: They can read, you know, Various types of documents, but MD files are the simplest.
- 24:37SPEAKER_01: Markdown files.
- 24:37SPEAKER_01: Markdown files, and it can then format all of that context and lay it out nicely.
- 24:44SPEAKER_01: Once that file is saved, you can then keep that.
- 24:47SPEAKER_01: It's making all of these little different ones.
- 24:49SPEAKER_00: That's its source of truth, isn't it?
- 24:50SPEAKER_00: Yeah.
- 24:51SPEAKER_00: One of the things that I think we should,
- 24:53SPEAKER_00: just before we move on, these are the foundations, aren't they?
- 24:57SPEAKER_00: Things like brand voice and pricing.
- 25:01SPEAKER_00: Rate cards,
- 25:02SPEAKER_00: et cetera,
- 25:02SPEAKER_00: foundations of the agency,
- 25:04SPEAKER_00: but one really important one that we mustn't skip over,
- 25:08SPEAKER_00: I don't think, is process, isn't it?
- 25:10SPEAKER_00: Your processes in your agency should be in this thing as well.
- 25:13SPEAKER_00: Completely, yeah.
- 25:14SPEAKER_00: Or if it's not in one file,
- 25:16SPEAKER_00: it can be in a number of these marked down files as reference points.
- 25:21SPEAKER_01: Yeah, and that's almost the next step, this fundamental context of an agency.
- 25:28SPEAKER_01: Simply, if you were to hire or have a new hire,
- 25:33SPEAKER_01: what would you teach them in the first couple of weeks?
- 25:35SPEAKER_01: It's all these different things.
- 25:38SPEAKER_01: I've done this co-work,
- 25:39SPEAKER_01: Claude onboarding with a few agencies and it's all about our brown voice,
- 25:43SPEAKER_01: even AI policy,
- 25:45SPEAKER_01: so that every time you use an environment it reads that policy and it's aware of
- 25:50SPEAKER_01: the do's and don'ts and it can inform team members.
- 25:53SPEAKER_01: It's company wikis really,
- 25:55SPEAKER_01: but it's in AI,
- 25:56SPEAKER_01: there's frameworks they use,
- 25:58SPEAKER_01: it's values and mission and past credentials, all this kind of stuff.
- 26:03SPEAKER_01: But as you say,
- 26:03SPEAKER_01: processes,
- 26:05SPEAKER_01: it's packaged context,
- 26:07SPEAKER_01: but almost in a more meticulous format in the sense of,
- 26:15SPEAKER_01: I've come to learn,
- 26:16SPEAKER_01: being relatively new to the agency world,
- 26:18SPEAKER_01: that things like SOPs aren't the most common thing to see.
- 26:23SPEAKER_01: Standard Operating Procedures.
- 26:26SPEAKER_01: The fact that I haven't even explained the acronym there, it says a lot.
- 26:32SPEAKER_00: I did explain it because I think you're absolutely right.
- 26:34SPEAKER_00: It was a shock to you,
- 26:35SPEAKER_00: but it's not such a shock to me,
- 26:37SPEAKER_00: although it's something that I do reflect on often and think that agencies are
- 26:43SPEAKER_00: behind other industries, massively behind in this case.
- 26:48SPEAKER_00: Processes are rarely written down in smaller independent agencies.
- 26:53SPEAKER_00: Not in every case,
- 26:55SPEAKER_00: of course,
- 26:55SPEAKER_00: but how often do we come across agencies that we talk to and say,
- 27:00SPEAKER_00: you need to put your SOPs,
- 27:02SPEAKER_00: processes, standard operating procedures, whatever you want to call them, into
- 27:07SPEAKER_00: A file and a document that AI can read and refer to and form its knowledge base and
- 27:16SPEAKER_00: before anybody needs to go any further we get the response,
- 27:18SPEAKER_00: actually hang on a minute,
- 27:19SPEAKER_00: we know what we do,
- 27:20SPEAKER_00: of course we do it every single day but it's not written down anywhere or
- 27:24SPEAKER_00: if it's written down it's in a poor format that's not been referenced
- 27:29SPEAKER_00: or Amended recently,
- 27:32SPEAKER_00: lots of stuff that people are doing are in either everybody's or certain people in
- 27:36SPEAKER_00: the agency's heads.
- 27:38SPEAKER_00: So that's step one, isn't it?
- 27:40SPEAKER_01: And this is the quickest, easiest win with AI, I'd say, bar none.
- 27:49SPEAKER_01: Almost separate personas.
- 27:53SPEAKER_01: Agents is the word for it, really.
- 27:56SPEAKER_01: ChatGPT has GPTs, Gemini has Gems, Claude has Skills and Project Folders.
- 28:03SPEAKER_01: Skills are a little bit more advanced.
- 28:06SPEAKER_01: I take a liking to them.
- 28:08SPEAKER_01: We said a lot before about how AI is trained on these massive data sets.
- 28:13SPEAKER_01: When you give AI a role and you give it a workflow and you tell it its tone of voice
- 28:19SPEAKER_01: or what kind of role it should play,
- 28:23SPEAKER_01: what it should be the expert in, It's almost getting rid of all the noise.
- 28:27SPEAKER_01: Instead of looking at its whole data set and giving its best inference
- 28:30SPEAKER_01: and making all these predictions,
- 28:32SPEAKER_01: it's going, okay,
- 28:34SPEAKER_01: if you're an agency that specialises in web design,
- 28:38SPEAKER_01: it's going to go, right,
- 28:39SPEAKER_01: I'm going to focus on all of the web design data that I've been trained
- 28:42SPEAKER_01: on right now.
- 28:43SPEAKER_01: Let's get rid of everything else.
- 28:44SPEAKER_01: I'm the web design expert.
- 28:45SPEAKER_01: I'm not going to look at law.
- 28:47SPEAKER_01: I'm not going to look at being a Don't worry too much
- 28:52SPEAKER_00: about your first example,
- 28:53SPEAKER_00: don't worry too much about Kings and Queens and connecting those,
- 28:56SPEAKER_00: that's probably not that relevant but focus on this,
- 28:59SPEAKER_00: focus on how we work,
- 29:00SPEAKER_00: pricing,
- 29:01SPEAKER_00: who we work with, what our tone of voice is, what we do as an agency of course.
- 29:06SPEAKER_01: And the good thing about these package systems is
- 29:09SPEAKER_01: that Once AI knows what your workflow is,
- 29:14SPEAKER_01: it's the source of truth and it can teach your team.
- 29:18SPEAKER_01: So if leadership create this process or it's
- 29:21SPEAKER_01: A very experienced team member who's done it for many years and knows it better
- 29:25SPEAKER_01: than anyone,
- 29:26SPEAKER_01: if they make sure that all of that knowledge in their heads is in the AI system,
- 29:30SPEAKER_01: if a new team member gets stuck,
- 29:32SPEAKER_01: perhaps it's a junior that's just been onboarded and they're struggling,
- 29:35SPEAKER_01: as long as the agent has been configured by the person who's very knowledgeable,
- 29:41SPEAKER_01: the junior can then ask it questions on what the workflow is
- 29:43SPEAKER_01: and how they should do it.
- 29:46SPEAKER_00: Which fills a bit of a gap that we've always had in agencies
- 29:48SPEAKER_00: that these processes we said earlier generally aren't written down.
- 29:52SPEAKER_00: New people come into the business,
- 29:55SPEAKER_00: do our best to onboard them of course,
- 29:57SPEAKER_00: partner them up with somebody knowledgeable in the business that understands
- 30:00SPEAKER_00: the agency and has worked in the agency for a while.
- 30:04SPEAKER_00: But there's not that,
- 30:04SPEAKER_00: unless that person's there with their buddy or whatever in the early days
- 30:07SPEAKER_00: of their onboarding, there's generally no real reference point.
- 30:11SPEAKER_00: So AI can actually not just solve problems, but it can help to educate people.
- 30:17SPEAKER_00: Yeah, and the more you accumulate this knowledge,
- 30:20SPEAKER_00: the more
- 30:21SPEAKER_01: future-proofed you are to AI innovation in the future.
- 30:25SPEAKER_01: Because I think undoubtedly,
- 30:27SPEAKER_01: the more people rely on AI tools in their business,
- 30:31SPEAKER_01: which is where it's going,
- 30:33SPEAKER_01: The higher prices these companies can then inflict on their users.
- 30:39SPEAKER_01: And so if it comes to a point
- 30:41SPEAKER_01: where your trusted AI provider starts charging you 10 times more,
- 30:46SPEAKER_01: the context is the value.
- 30:48SPEAKER_01: And as long as this is packaged and it's accumulating over time,
- 30:52SPEAKER_01: every time you use AI chats,
- 30:53SPEAKER_01: I have a system that On every session,
- 30:57SPEAKER_01: it takes all the context in the chat and it files it like a librarian
- 31:01SPEAKER_01: across things like offers and decisions and perhaps it's new revelations in
- 31:06SPEAKER_01: the business or a change to my offer, things like this.
- 31:10SPEAKER_01: The more AI is accumulating this knowledge,
- 31:12SPEAKER_01: the more prepared you'll be for the future because you can plug all this context
- 31:16SPEAKER_01: into a different tool.
- 31:18SPEAKER_01: Your agency's context is the value.
- 31:21SPEAKER_01: It's all about documenting what's in your heads.
- 31:25SPEAKER_00: So, Callum, this episode is about making AI work for your agency.
- 31:34SPEAKER_00: We've talked a lot about how AI works,
- 31:36SPEAKER_00: how it's immensely clever,
- 31:37SPEAKER_00: but also how it is a machine and it's doing things,
- 31:42SPEAKER_00: predicting words, not actually conversing.
- 31:47SPEAKER_00: We often get,
- 31:48SPEAKER_00: and we have often received requests in the last six to 12 months,
- 31:54SPEAKER_00: From agencies and about the thing that they most want,
- 31:59SPEAKER_00: we talked about prioritising what they want in AI earlier,
- 32:03SPEAKER_00: and we often get the request for, I want a company brain, that's my goal.
- 32:09SPEAKER_00: Deliver us a company brain, guys.
- 32:11SPEAKER_00: I don't blame them.
- 32:12SPEAKER_00: No, neither do I.
- 32:13SPEAKER_00: That's more than possible, an AI company brain and more than desirable.
- 32:21SPEAKER_00: When we've had those conversations,
- 32:22SPEAKER_00: we often feel like we're putting the brakes on people's ambition.
- 32:26SPEAKER_00: So we always say to people,
- 32:28SPEAKER_00: you need to bring it back to a human brain,
- 32:32SPEAKER_00: your knowledge and your experience and your memories Written down in the form
- 32:39SPEAKER_00: of data before you even start to think about how you connect that to
- 32:45SPEAKER_00: and how you start to deploy that as an AI company brain.
- 32:49SPEAKER_00: There is a stage before that that we have already touched upon,
- 32:52SPEAKER_00: but just to finish off,
- 32:53SPEAKER_00: summarise for us what that stage is and how you bridge that gap if you are
- 32:59SPEAKER_00: an agency.
- 33:00SPEAKER_01: Gareth, you are completely right.
- 33:03SPEAKER_01: The analogy of the human brain is one I love because I have that background
- 33:07SPEAKER_01: in psychology and completely AI itself without any kind of context
- 33:14SPEAKER_01: of your agency is pretty suboptimal.
- 33:19SPEAKER_01: Really do the work that you want it to, mediocre outputs.
- 33:23SPEAKER_01: It's like if you were to,
- 33:25SPEAKER_01: God forbid this happened,
- 33:28SPEAKER_01: but if one of your team members' memories was wiped suddenly,
- 33:32SPEAKER_01: they wouldn't have any idea of what to do.
- 33:34SPEAKER_01: They might have their general intelligence, but they don't know anything about the agency.
- 33:41SPEAKER_01: Starting,
- 33:41SPEAKER_01: a lot of agencies have the ambition,
- 33:43SPEAKER_01: especially agencies that are doing very well and have the money to spend and want
- 33:46SPEAKER_01: to get ahead of their competition to create some grandiose kind
- 33:50SPEAKER_01: of complex company brain.
- 33:53SPEAKER_01: And going back to what we were saying at the start of this episode,
- 33:57SPEAKER_01: This brain,
- 33:57SPEAKER_01: the difference between just it reading files from RAG and creating its own brain,
- 34:03SPEAKER_01: you can use AI in all the algorithms where it learns the meaning between words,
- 34:09SPEAKER_01: but it's the meaning between words in the context of your agency.
- 34:13SPEAKER_01: So these brains,
- 34:13SPEAKER_01: which could set you back sort of 50 grand
- 34:16SPEAKER_01: and there's things like this technical knowledge graphs
- 34:20SPEAKER_01: and there's automated ingestion and you plug it into any AI tool,
- 34:24SPEAKER_01: they would perhaps know You could configure it to go, okay,
- 34:28SPEAKER_01: this client relates to this team member,
- 34:31SPEAKER_01: this team member does these processes,
- 34:33SPEAKER_01: has this tone of voice, these are their strengths and weaknesses.
- 34:36SPEAKER_01: It knows more about the relationship between your own data.
- 34:41SPEAKER_01: But the problem is,
- 34:42SPEAKER_01: if you don't have that data tracked and documented, you can't possibly make that.
- 34:48SPEAKER_01: It's like you said, it's a brain without any memories.
- 34:50SPEAKER_01: And so the start of the process and why I recommend co-work Is to get the core data
- 35:00SPEAKER_01: over time, like what would you teach to a new employee in the first couple of weeks?
- 35:05SPEAKER_01: Things like simply a document explaining who is agency X and it's all
- 35:12SPEAKER_01: about these are the credentials we have,
- 35:15SPEAKER_01: these are the people we help,
- 35:16SPEAKER_01: this is when we were founded, these are the awards that we've won.
- 35:20SPEAKER_01: This is the team size.
- 35:21SPEAKER_00: I think that's a good way of looking at it.
- 35:25SPEAKER_00: You've got to start with the basics, but imagine you're appointing a new managing director.
- 35:33SPEAKER_00: You're retaining the agency ownership,
- 35:36SPEAKER_00: but you're employing your first MD and you're bringing them in.
- 35:41SPEAKER_00: You're giving them the basics,
- 35:42SPEAKER_00: but then over time you're going to give that person,
- 35:44SPEAKER_00: because they're senior and going to run the business for you,
- 35:46SPEAKER_00: you're giving them more and more context, data, information.
- 35:50SPEAKER_00: So it's almost looking
- 35:54SPEAKER_01: at it like that, isn't it?
- 35:55SPEAKER_01: Completely, yeah.
- 35:56SPEAKER_01: There's examples that I've worked with with agencies in the past of what information
- 36:00SPEAKER_01: to digitise.
- 36:01SPEAKER_01: It's who is AgencyX,
- 36:04SPEAKER_01: it's brand voice,
- 36:06SPEAKER_01: it's frameworks that they use most often,
- 36:10SPEAKER_01: it's ingesting cred decks and ingesting those awards.
- 36:17SPEAKER_01: Who they work with.
- 36:19SPEAKER_01: So there's all these types of things.
- 36:21SPEAKER_01: And like I said,
- 36:22SPEAKER_01: simply it's for each one of those kind of context files,
- 36:26SPEAKER_01: you get all of the quantitative data,
- 36:29SPEAKER_01: all of the files,
- 36:30SPEAKER_01: the spreadsheets, the frameworks, the things that are already written down.
- 36:35SPEAKER_01: Maybe it's LinkedIn posts that you've done, notes that you have about a certain topic.
- 36:40SPEAKER_01: And then you can actually get AI once it's ingested all those files to go,
- 36:45SPEAKER_01: You can get it to interview you,
- 36:47SPEAKER_01: ask me a hundred questions to understand who the agency serves or what
- 36:56SPEAKER_01: the agency is fundamentally and then it will dig deeper.
- 36:59SPEAKER_01: You can sit with a leadership team, record the transcript, answer all these questions together.
- 37:03SPEAKER_01: There's Whisperflow, which we now love, where you could do this.
- 37:09SPEAKER_01: You can dictate your thoughts rather than typing them all.
- 37:20SPEAKER_01: So, ultimately,
- 37:21SPEAKER_01: we get all of this data compiled,
- 37:24SPEAKER_01: and we save it as markdown files,
- 37:26SPEAKER_01: that's the preferred, it's almost the format that AI can read best.
- 37:34SPEAKER_00: And quickest.
- 37:35SPEAKER_01: And quickest, and that's, so it's platform agnostic.
- 37:40SPEAKER_01: We're massive fans of Claude,
- 37:42SPEAKER_01: With teams that I've worked in that use Claude as well,
- 37:44SPEAKER_01: we can package that information in plugins,
- 37:46SPEAKER_01: everyone downloads the plugin, they get all the skills and markdown files.
- 37:50SPEAKER_00: That's a great point though actually,
- 37:52SPEAKER_00: if you do choose to transfer or use to LLMs at once, you can use the same files.
- 38:00SPEAKER_00: And that's the future proofing element
- 38:02SPEAKER_00: and so you can then
- 38:03SPEAKER_01: package these,
- 38:05SPEAKER_01: use it in your AI provider of choice.
- 38:12SPEAKER_01: So we just say, package it, put it in any provider that you want.
- 38:17SPEAKER_01: Alan, I'll wrap it up now, I'll just
- 38:19SPEAKER_00: say that...
- 38:21SPEAKER_00: So once you've got that, that's almost your start of your brain journey.
- 38:27SPEAKER_00: Yeah.
- 38:28SPEAKER_00: Once you've got that, then it can start to think more...
- 38:38SPEAKER_00: And
- 38:45SPEAKER_01: that is the strategy.
- 38:47SPEAKER_01: If you start with that step,
- 38:49SPEAKER_01: you can't go wrong because it's almost the start of the journey
- 38:52SPEAKER_01: to more technical systems.
- 38:55SPEAKER_01: And I do believe that in 15,
- 38:58SPEAKER_01: 20 years time,
- 38:58SPEAKER_01: who knows how rapidly these tools will evolve and how the industry will as a whole,
- 39:03SPEAKER_01: but I imagine every agency will have their brain anyway.
- 39:07SPEAKER_01: Companies that install them for you and all this,
- 39:10SPEAKER_01: but to make their jobs easier in the future and to get to that level,
- 39:15SPEAKER_01: start documenting and digitizing everything that you can about your agency so
- 39:22SPEAKER_01: that AI can become your employee, not just an off-the-shelf tool.
- 39:28SPEAKER_00: Great, great summary.
- 39:30SPEAKER_00: Enjoyed that, Cal.
- 39:31SPEAKER_00: Thank you very much, Gareth.
- 39:32SPEAKER_00: Always a pleasure.
- 39:33SPEAKER_00: Hope you guys enjoyed listening and We'll see you next week.
- 39:37SPEAKER_00: Yeah.
- 39:37SPEAKER_00: If you don't want to miss the next episode,
- 39:39SPEAKER_00: hit subscribe on your favorite channel, podcast application.
- 39:45SPEAKER_00: And see you next time.
- 39:46SPEAKER_00: Great.
- 39:46SPEAKER_00: See you next week.