Hedaux , Hogg & The Machines 1

In this debut episode of "Hedaux, Hogg & the Machines", host Simon is joined by guest co-host Ed Hogg, CEO of SolvedBy.ai. Together, they explore the fast-moving state of AI technology and its practical applications across retail, hospitality, and customer-facing environments.

Key takeaways:

🔹 Data Security: Why enterprise security settings matter when using LLMs and the hidden risks of unvetted open-source models. 

🔹 Head Office Efficiency: How area managers are using AI agents to automate weekly store action plans in minutes. 

🔹 Shop Floor AI: How retail giants like Walmart, Tesco, Lowe's, and M&S are rolling out AI tools to frontline teams. 

🔹 What’s Next: What to expect from upcoming models, enterprise forecasting tools, and the realistic timeline for store robotics.

📅 Save the Date! Join us live at the Productivity Forum on 10th September in Birmingham.

🎟️ Secure your spot: ReThink Forum 2026

Download a copy of our first ReThink Productivity Retail Report

 #theproductivityexperts
Register for the 2026 Productivity Forum
Download a copy of our first ReThink Productivity Retail Report
Connect to Simon on LinkedIn
Follow ReThink on LinkedIn 

Transcript

Copy

Simon 0:05

Welcome to the Productivity Podcast. This is the first in a new series called Heddo Hog and the Machines, and I have a co-host, Ed Hogg. How are you doing, Ed? Very good, thank you, Simon. Welcome back because I know we we've been on a couple before. I think February 2025 was the last time we we spoke about all things AI. So we're going to do this every four to six weeks. It's a fast moving environment. Every day something new seems to change or happen, but hopefully we can uh enlighten you. I will keep it simple. Ed will make it complex, and Ed, I will promise to draw you back out of complexity anytime I don't understand what's going on. I think that's probably needed. Good. Uh for those that may have not listened to the February one, do you want to give us a bit of background on yourself and solve by AI?

SPEAKER_01 0:53

Yeah, so uh I'm Ever Hogg, I'm the CEO of Self by AI. I um founded Self by AI, which is an AI forecast and then decide business um in Edinburgh in 2020. And we have a team of data scientists who focus on uh solving problems using artificial intelligence in retail and hospitality. And one of the things that we really focus on is using AI productivity. Um and that means that we are very much spending most of our days looking at the latest models, looking at um what our customers are doing, what our partners are doing with AI in order to um be more productive um and save money and increase revenue.

Simon 1:36

And before solve by what were you involved with? What were you up to?

SPEAKER_01 1:40

Uh yeah, so I uh used to work in high-level motorsport. I used to um work for Bentley um designing um race cars there. Uh and then um I worked in workforce management as well. So uh um working in the shop works um on uh retail workforce management. Um so yeah, so that's my background. Uh the motorsport being being a more interesting.

Simon 2:02

Exactly. Exactly. Not not to do WFM a a discredit, but yeah, I'd take motorspot over WFM any day. Um sorry to all those WFM geeks out there, but uh that's just the way it is. So 18 months ago we chatted. Um, wow, I mean, AI's moving, a lot happens in 18 minutes and 18 seconds, let alone 18 months, doesn't it? So give us a bit of a state of the nation of if I'm working in retail hospitality, any customer-facing environment, what what's the state of the nation now in AI? Because we hear about Claude, Chat GPT, Gemini, Groc, um, and all the others that I've missed out because there's way too many to mention. What where are we? What's happening?

SPEAKER_01 2:43

Yeah, I I guess let's start on on the the big three, which is uh open II anthropic, and then uh in most people um who work which retail uh Gemini is also up there. So in 2025, February, um, we had 4.0 Sonic 3.5 from Flaud and then uh Gemini 2.0 Flash. And those were roughly 128,000 token models, um which were which were outputting roughly 16,000 tokens.

Simon 3:16

If you explain a to explain a token, I'll stop you. Go on, explain a token. This is my first one of pulling you back.

SPEAKER_01 3:22

It didn't take us long, did it? Uh so a token is uh basically a way of the AI breaking down a word. All LLMs uh or large language models that these are, um, break down um the text that you put in into uh um tokens and you can split out words. Um if you think of hello, that's probably the the right size for a token. It's uh got um that that's probably the uh relatively simple uh set of five letters that the AI can understand always means one consistent thing. Um if you think of a word like unbelievable, that tends to have uh three tokens uh in it, with uh your um uh describing some parts and then believable being broken up into two, there are two different parts of it, depending on which model you're using. And what's really important about tokens is that um that's where the AI gets its context from. Uh so it's looking at comparing what tokens you've put in um to then outputting the next token that it thinks, or the most likely token, um, that's how you get your response. And uh they have fixed sizes of how big the input can be and how uh big the output can be. And we used to, 18 months ago, to be on 128,000 tokens, uh, which is roughly the size of a Hobbit uh novel. Um, and we used to be able to output 16,000 tokens, which is roughly the word count in this podcast time was too. It's roughly a 60-minute podcast uh length written down. And now ChatGPT 5.6, um, the most recent models um from Claude, particularly Opus 5, they are uh and then and also Flash, which was uh Flash 3, which was released by uh Google very recently, they are million token inputs, 128,000 outputs, which means that um uh ChatGPT uh 5.6 is the best example, is able to take in all of the Harry Potter series and will be able to write an eighth book in terms of the number of tokens. Which is great if you are particularly into writing uh word works of fiction, um, but doesn't really relate to um to retail. So to come back to that, you've you've now got these models which are able to take in large amounts of context and thousands and thousands of uh Word documents, PowerPoints, spreadsheets, uh PDFs, pictures, and they are all multimodal, and they're able to take those in and use a uh wider breadth of skill um to make uh higher quality things for us, such as uh spreadsheets that have the calculations embedded within them that have been tested, that we've done the work on, such as uh design uh a PowerPoint based off of some example design that's in the PowerPoint already with this text and all of these documents, and that can be really useful uh in uh retail. Um and all these LMs are also a lot better at acting like humans in responding, and so uh they've all been taught off of the many, many chats that we've had with customers. Um, and uh since 4.0, you've had lots of examples of people converting audio files. They may be examples of customers speaking to managers or um complaints from uh customers phoning into the service desks um or podcasts or um other examples of human interacting with human, and now the responses that they can give are much more human-like, um, to the point where a lot of the interactions you have online with customer service are AI, even if it feels like we're speaking to a human.

Simon 7:18

They're becoming quicker, more intelligent, can handle multi-complex pieces of data from multiple different formats in different points. Is that fair?

SPEAKER_01 7:30

Yeah, yeah. So um, if you look at the most recent Chat GBT models, um you've got uh soul that is uh deep reasoning and can go off and can spend an hour taking into account everything and and reason, but you can also use Terra and Luna to get back answers that would have taken 20 minutes um a year and a half ago in 10, 20 seconds, um having taken into account all of the context you want it to take into account. So yeah, it's making a massive difference. Uh Gemini uh Flash is is designed to do that. Um and uh yeah, 3.6 is it's so so quick um doing that. You you you aren't able to output as many tokens with with Gemini 3.6, but you're still talking about being able to output half the hobbit, which is enough to generate uh roughly 240 pages of um uh a Word document or a PDF, depending on the kind of context and quality of that document. So it is more than enough for most people to be able to use and get a written response that that enables them to be to be more productive.

Simon 8:41

And I I suppose the the warning signal that goes up when you you talk through that around consuming PowerPoints, Excels, all those other bits of of company information is where's it going? Where uh who's it sharing it with? How do you make sure it's um it's secure? I know kind of our internal rethink one we've got the not training any data with outside of the business, but there must be genuine concerns that everybody's teaching the machines to be more intelligent.

SPEAKER_01 9:12

Yeah, I I guess there's two sides of that. So there's the first side is that some businesses don't have the setting of don't use my data to train. Um, and therefore there are examples of um entire trading algorithms from uh large hedge funds in the US ending up in the training data for these models. And so that is 20, 30 billion pounds worth of uh alpha to those businesses just just being stolen, which is um uh uh and being used to train, um, which uh when we think about our own uh work, it it could be harmful. There could be strategies, there could be things that we don't want to get out, which could the models could be training on and reinforcing and repeating back to other people when they ask for it. So all of our hard work could be given to somebody else. Um but all three of the kind of ones we've talked about, so ChatGPT, um Claude, Gemini, if you're using Copilot within uh Microsoft, they all have uh now basically in the work in the enterprise version as standard not to uh train on that data until I get that data from somewhere else. There are a number of enterprises that are using their own open source models. And so open source means where um an organization has developed it and then they uh allow you to host that or work on it yourself. So they make it public what they what the weights of their model um are. And the weight is uh the little bit inside each AI um uh node, which tells you how to connect with other nodes in terms of multiplying and reducing numbers. And what it means is that um you can go and run a fraction of the cost. I think um somebody was telling me that uh the difference between using uh Opus 5 uh and an open source model is like a hundredth of the cost. And so you're um you're saving you're saving an awful lot of money there. But um nine of the ten best in the world were developed uh in China, and uh it is not public what training data that they use, and therefore it is entirely possible that those models could have backdoors encoded into them that would allow them to take your data when you're running it. Um and uh while they've been security tested to an extent, we will never know very specifically as to what level they've they've been. So you really do need to be careful about what you're putting into these models, especially if you're using some of the ones which are um uh open source or haven't been security tested by by a large number of organizations.

Simon 11:54

Yeah, makes sense. Yeah, it's it it feels it feels like at the moment people might not be quite sure where it's going. So they're either hesitant to put it in or they're putting it in and it's going somewhere to train. But yeah, the that whole open source thing, great, you get it at a reduced cost, but do you really do you really know what's wrapped around it, I suppose, is is the risk, isn't it?

SPEAKER_01 12:13

Uh yeah, especially if you're um uh like uh critical infrastructure. Um uh I I was speaking yesterday to a to a hospital and uh they were talking about the fact that they will never be able to use it because they just the very, very small chance that a model has been trained of if you are a hospital and this infrastructure diagram gets uploaded to it, it becomes a security risk. And so that kind of stuff is really scary to uh organizations. And so yeah, I there you're that in enterprises, in the businesses that uh tend to be focused on productivity and listen to this podcast. Um, I don't think we're gonna see open source models um in in those in those businesses for a very long time.

Simon 12:56

Okay, so so in retail head offices, what kind of things are you seeing happening now? Where's where are people pushing the boundaries?

SPEAKER_01 13:04

Yeah, so I think uh in terms of in um head offices, Copilot is probably the most interesting uh example. So in Copilot, uh you now have access to all of the top models. Um and so and the way that Copilot charges is that it does it based off of the complexity of the task that you're sending. And um most organizations uh are who are using Copilot have uh still got those enabled because it comes as default uh on. And so businesses um who are using that in retail are able to give area managers, managers those kind of um tools to be able to produce really high quality reports. So yesterday I I was in uh the Meadowhoe shopping centre in Sheffield and uh was talking with an area manager who was showing me that every Monday morning they have an agent which they have built in Copilot that uses uh ChatGPT 5.6 to go and get all of their spreadsheets uh in terms of EPOS sales from the last week, all of their customer feedback stories, all of their um actions, uh promotional marketing material for the week to send the Monday morning email automatically to their staff members, particularly their um their store managers, that goes, here are the five things that you need to focus on in your store, particularly, and here's the five things that we as a business are focusing on this week. And uh these are the kind of things that you need to be telling your staff. And that reads completely automatic and uh is uh great. And that um I love it, it's uh really good use of AI. That is exactly the kind of thing that um empowering head office staff and and area managers to uh to be doing in in retail is really positive. Um the problem is that it's not gonna last forever. So um people who are using those uh big models are spending about a thousand more credits, which is the um way that uh um Copilot works out the cost of uh uh the tokens you're sending in uh than you are if you're doing some of the basic tasks. So if you, for instance, were having a conversation with Copilot and going, can you tell me um what uh the weather is going to be like in uh Winchester on Saturday, people go through and um do that very quickly for a very small amount of credits. But um because when we use the bigger models, we are throwing in more complex stuff with it already charges you a hundred times more in CodePiler and is a hundred times the number of tokens you've used, and people are using a lot more tokens per query. And so most enterprise businesses will get to a point in the very, very near future where the CTO will be stopping them from having access um to it. And to be honest, the the small and middle models are able to do those kind of functions, you'd still be able to do that meadow hall area manager example um uh with some of the smaller models, but it is just worth being aware that in Copilot, some of the models um won't be available to you in your organization.

Simon 16:28

So that's that's head offices, and again, lots of press around people then opening up AI to colleagues on the shop floor. In retail, I think we've we've got a few examples there that you want to work us through.

SPEAKER_01 16:42

Yeah, so um I guess uh very recently the the Argos sales, so they have um been sold to an organization that uh is coming in with a concept of we're gonna put AI in all of our stores. They they are retail turnaround guys who specialise in putting AI in the hands of staff. And if you look at some of the examples, um so uh in the US, we've got kind of Lowe's who are doing this expertise in the worker's hand, this kind of chat tool that people are able to access um and uh work through that kind of concept of what is the problem that uh you're you're after, chat with the bot, get an example of it. Um uh and uh they call it my low um companion, and uh it's now linked 1700 of their stores. Uh and then Walmart are putting every single staff member, all 2.1, 2.2 million of them, on a literacy um course for AI, and um so that every single staff member should be able to take advantage of AI at some point in their in their journey, and that and that makes a massive difference in terms of being able to take advantage of it. In the UK, Tesco uh have got like 280,000 colleagues training their agents. So um they basically gave their uh customer assistant shop uh shopping assistance to all of the all of the staff, including their uh it's got meal planning and it's got like dietary preferences and uh what what would you do with leftovers and and how would you go about building building a basket for me? Um and it's giving it two colleagues to test, and those colleagues are on the shop floor and testing them. And that is a really cool use of AI, um, and uh part of the Tesco's development progress on it. And then uh MS have also uh got um 11,000 copilot licenses um for all of their store managers and store support colleagues. Um and uh the proposed use is like the morning huddles, like sales insights stuff. Um one question that is of interest is is how often that's gonna get used. Um so uh buying those copilot licenses, giving every Tesco's employee th those tools, 2.1 million Walmart staff. Um how many shop floor staff are regularly gonna be on their phone interacting with the AI rather than either doing the role interaction with the customer, which uh um most most staff will still walk you around a store as to which aisle the item is when you go and ask them, rather than querying it on the on the app or or pointing you towards the app. Um, and then the other um thing is um the majority of tasks that are done by shop floor staff are physical, and that is more of a robotics thing, and uh that has been coming for a while. Um humanoids are um like they're being worked on uh in AI, and there's a lot of AI humanoid work going on, and particularly in terms of the idea of being able to replenish things or do uh the washing up or do the in a cafe or um do the uh ironing or dry cleaning, but uh your most of the demos for those are in the very early stage, and I think we've we've still got a while to go um and a lot of capital investment by Silicon Valley uh to make that real.

Simon 20:33

And I I think it was yesterday I was reading on LinkedIn uh Tesco's are trialing a robot that moves around the store, counts the gaps, and is more accurate in terms of replen. So there's bits that are teasing through, but they feel like they're really you know one two shop trials at the moment. Obviously, the cost of implementation across an estate is significant at that type of scale, but people seem to be more um accepting of dipping their toe in the water and having a go, is they're finding ways. I think all of this comes back to some of the wider stuff we talk about on the podcast of managing uh cost pressure, national living wage, and I all the other stuff.

SPEAKER_01 21:12

Yeah, I mean uh if you presume that it's um the uh staff member at head office is on somewhere around 40 to 45,000 pounds, and you could use a four to five thousand pound AI tool to replace 90% of their work. That is the the calculation uh everybody in Silicon Valley who's betting on ChatGPT and Claw being trillion-dollar businesses is is hoping comes true. Um we have yet to see that properly. There is a lot that it can do, uh, but um most of the the labour costs are still in the the physical world and um the the replenishment, the driving the lorries to uh two sites, the um yeah, the the distribution centers. Um and while robots are are making a difference there they're still quite a way away from from being being able to do that. So um, yeah, AI is slowly doing it, but we're talking um we're talking inches rather than miles here in terms of uh the speed or the distance in which which AI is going to saving, saving those those costs um while those pressures are moving at a significantly faster rate.

Simon 22:33

Yeah, and there have been organizations recently that have done some head office um job round cuts and and used uh we're moving more to an air approach on that. I yeah, those those calculations will be interesting. I think there's there's some stuff and some intelligence that you're always going to need the expertise, so we shouldn't we shouldn't be too scared at this moment in time. And Ed, just to finish off this first episode, we'll speak again in kind of four to six weeks' time. Is there anything you can see pending on the horizon that you think we'll be talking about on the next episode?

SPEAKER_01 23:03

Yeah, um so Tan Allen is uh at the White House again this week to talk about uh ChatGPT six. Um and uh that model is in theory so much better um than uh 5.6. Um and then that's that'll be really interesting. And then uh there is uh rumors that at least two or three of the US and European open source models uh will become available um in the next four to six weeks, and it'll be really interesting to see where those are because those will be developed under the eye of the US and EU uh governments, and therefore, in theory, will be capable of being used by enterprises because there is no fear that your data is being taken back by a uh an adversarial actor. Um so yeah, that those are the kind of two things that I'll be watching. Um, and then also in terms of the time series forecasting world, the world that I spend most of my time in, um there is uh there's a few more um uh LLM-based models and the next next steps of those models that are rumored to be coming out in the next couple of next couple of weeks. So it'll be really interesting to see um how much improvement they bring to the time series forecasting, uh particularly in in new stores, which is a problem that um uh it would be really nice for everybody to have. Um and and when we relocate to different places. So yeah, those are the kind of things that I'm watching in terms of the AI world over the next six weeks.

Simon 24:36

Perfect. Well, thanks. Enjoyed the first episode. I think it's safe to say the machines aren't taking over quite yet. We've not entered the Terminator World, but um I think they're they're here to help if you can harness the data safely in the in the correct way. So thanks again, Ed, and we will catch you on the next episode. Thanks, Alan.