Ready for a 58-page complaint letter?

Week 17. Kieron wore a shirt and shoes to a food industry conference for 60 senior leaders. Neil nearly went back in the doghouse because his daughter told him to read the room and he (foolishly) chose not to. Leading AI's website has had hundreds of cyber-attacks in two weeks. It's been that kind of week.

Schools — 58-page complaint letters and why the AI arms race matters
A head contacted Neil about using AI. The emerging problem: parents are using AI to write increasingly sophisticated complaints. One head received a 58-page letter. KnowledgeFlow answers it regardless of length, grounded in the school's own policies, in the school's own voice. Kieron gave his youngest son's school access to the demo tool six weeks ago. The apology correction emails have stopped.

Kieron's big talk — did it land?
60 leaders. 90 minutes. A room full of MDs and CFOs he'd never met in an industry he barely knew. Last year's AI session was too technical and too salesy. Kieron's answer: Nokia, Blockbuster, Kodak — and one uncomfortable question. How easy would it be for a new entrant with great AI and no capital to disrupt your supplier relationships tomorrow? The feedback was good. He didn't end up in a ditch.

The Agentic Bid Writer — Kieron hasn't even seen the latest version
12,000 words across 7 questions. Cup of tea. 45 minutes. Done. Then it marks itself, rewrites itself up to three times, and presents you with a polished draft before you've even read the original tender. A conservative customer's verdict: "It looks practically done to me." Neil is already complaining he has to add bullet points himself. The revolution continues.

KnowledgeFlow is the context layer — we just didn't know it had a name
Gartner spent three days talking about context. Kieron spent two days thinking he knew nothing. Then he realised: policies, tone of voice, brand guidelines, system prompts — that's the context layer. KnowledgeFlow has been building it for two years, we just didn't know what to call it.

There are only 46 ships in the world that can lay undersea cable
1.2 million miles of cable carries almost all the world's data. It costs $800,000 to fix even minor damage. If someone wanted to take out a country's internet infrastructure — really take it out — it wouldn't be hard. Neil has Starlink as backup. Most organisations have nothing. Donald ran a penetration test this week: no major findings, five minor ones. How many organisations can say the same?

Neil didn't quite make it back into the doghouse. Mrs Watkins got her glass of wine. All is well.

Two mates. A bar. Thirty years of business between them. And all they want to talk about is AI.

Pull up a stool — we'll get the beers in. 🍺

Transcript

SPEAKER_02 0:01

Right, are you ready? Have you done your hair and makeup? Are you have you done are you ready for are you ready for week 17 of the pantomime horse podcast, Kieran?

SPEAKER_00 0:12

Here we are. Here we are, back in just getting into our pantomime horse outfit. We should probably explain where that comes from. Did we ever talk about why we say that?

SPEAKER_01 0:23

No, we never did. You started it. I think I did start it a long time ago. No, you definitely did start it. I've nicked it from you. I use it all the time now. But yeah, so where's it come from, Kieran?

SPEAKER_00 0:33

Well, it comes from uh our very early board meetings in uh WCL, the management consultancy that we created together, if you can remember back that far, back in 2002. And um, and I remember, yeah, because we we we used to get together in a board meeting in an evening back in the day when we could work all day long flat out for clients. So everything on the business was done out of hours, effectively. And I can't really, I think I just one day just said, let's get this pantomime awesome meeting underway, and then it became a bit of a tradition. I think and it wasn't nothing really naughty behind that, as usual with my comments, it's just kind of like I think it was I I think it was that kind of thing about we didn't really know what we were doing.

SPEAKER_02 1:16

We were making it up as we went along. And uh with so much in life.

SPEAKER_00 1:22

If you're not stretching your if you're not stretching your boundaries, then uh then you're not working hard enough and you're not growing, so um, you should always be on that level.

SPEAKER_02 1:30

If you're not living on the edge of taking up too much space. Crikey, this has turned very philosophical all of a sudden, hasn't it? Anyway, enough of that. So quickly. Exactly, exactly. Right. What's on your list for this week, fella?

SPEAKER_00 1:42

Well, I have been uh out and about this week. I had a really interesting session. Um big event. Yeah, the big event that I spoke at that I talked about last week and then now have done. Um, and um I think it went pretty well. So I'll I'll share a bit of that, that would be good to do. Um, and uh some more revelations and thoughts about um knowledge flow, our platform, and kind of where it fits in the grand kind of scheme of AI as the world keeps on shoving on. Um, and um I was whilst at the um conference that I was at yesterday, there was a speaker before me who was talking about geopolitics and risks that you know businesses need to be worried about. So I'll share a couple of snippets from that, but as usual, completely terrifying.

SPEAKER_02 2:31

What about you? What's on your mind? Uh well, geopolitics is always on my mind. I mean, as you know, it's one of my uh one of my uh I was gonna say favorite topics. I wouldn't say favorite, but it's definitely top of mind a lot of a lot of the time. And and I'm actually going to something tomorrow that I can't talk about um uh because it's related. So yeah, it is a bit terrifying, but actually there's some really interesting lessons for us, uh, and and I'll I'll share some of some of my thoughts on that. But before we get there, I wanted to talk about a couple of things. It's been a busy old week, it's just like you get to towards the end of the week, where the bloody hell has it gone? Now I can tell you this week, it most of it's gone on agentic um AI, um, which has been fascinating. Really interesting working with the big D, uh Donald Allison on that. Uh he's a genius, as you know. He does that kind of he makes that who face like the who out of uh the grinch and uh in the oh no, no, no, no. Oh, here it is. Uh so yeah, there's all there's all of that stuff. So there's I've got a couple of those things. I've had some really interesting uh customer conversations this week. Uh some around health, um, some around schools, and I'd like to pick your brain on the schools thing, because I know you've been working with schools, I haven't. Um, I've also been playing um with the housing tool um uh for stuff I'm doing on the website, and again, amazing. I'm just I wanted to talk about that. I've got a couple of other kind of uh uh bits and pieces around security and uh uh large language models in general, and and obviously uh not obviously uh uh but uh anybody who's really into AI will know that Fable 5 came out yesterday, which uh is the latest from Anthropic and uh and uh uh a uh Methos product, probably is the best way to describe it. And so all of those things. So there's a shed load to get through, so we better stop waffling about bloody pantomime horses and get to the good stuff, haven't we? So uh so yeah, let's start with your big event.

SPEAKER_01 4:27

Tell tell me about our listener, our single listener will have already switched.

SPEAKER_02 4:31

No, actually, I've got I've got some news for you, Kim. We've got more than one listener. I found out we do, we do, we do. So I got told as a chief exec of one of our customers uh listens to this, so we need to be super careful.

SPEAKER_01 4:44

Okay, I'll put my tie back on again.

SPEAKER_02 4:46

Yeah, yeah, get dressed properly. Uh we uh we've got um at least one listener in America, which I'll come back to later on. Uh, and we've also got uh I've been contacted by another listener in the UK this week. So we we we at least up to three. Oh well. Well, three that are awake and and and our audience that's that's asleep. So we're doing really well.

SPEAKER_01 5:10

Right, it's good. Four after 15 weeks, is you know, that's gotta be a new record, isn't it? This is week 17, so it's 17. 17. Oh my word. You you missed a week, you must have been asleep with Matt.

SPEAKER_02 5:21

Probably. Anyway, enough answers. Tell me about your week. Uh yeah, let me tell you about the second.

SPEAKER_00 5:26

So I mean, really, all of this week has been focused on this um this conference because it's in a sector, I'm gonna be vague as always, but it's in a sector that I really don't know very very much about. Um uh, although interestingly, it I will give it away to actually no, probably I can talk about the sector. But I um it back in my day I used to do lots of work with the Food Standards Agency and indeed even ran the International Conference on Campylobacterian chickens, um, which was went down amongst academic circles, apparently unbelievably. I had no idea. I went to I went to the um, where was it? I went to the EU to to uh Brussels, I guess it was. Um, yeah, it was, and walked into the room and said, hello, I'm Kieran. They went, were you the facilitator that ran uh become part of the international blah blah blah? And I was like, yeah. And they were like, we've all heard so much about it. That's that's hilarious. It was very peculiar because it's like properly deep uh science on foodborne viruses and things.

SPEAKER_01 6:24

Anyway, that's for a man with hidden shallows. That's a great accolade.

SPEAKER_00 6:28

Oh yeah, exactly. Well, it's just what I always wanted to be really well known for. It's like foodborne viruses uh expertise. Um anyway, so it was a food uh food uh business that I was speaking at, and so I was doing a load of research. What I was really trying to do, as I said, I think last week, is there were 60, 70 people in the room, all leaders of various businesses um and their senior teams, and I wanted to give them as much practical, sensible AI advice and some inspiration whilst being interesting, hopefully. Um and so it took a load of effort just trying to find use cases, AI things going on in their sector. Um, I was kind of nervous all the time on stage when I was sharing kind of examples, thinking, Am I am I just saying words that you're just like, what are you going on about? We don't do this at all. So, because I was obviously having to guess at a whole bunch of things with AI support in researching and uh etc. But um I thought it went really well. The feedback immediately afterwards was really good and sort of it what it hit the brief of being practical, people were clearer about what it is they have to do. Apparently, last year they had an AI session and it was really you know much more uh I guess quite advanced AI and sort of interesting, but no one really knew what they were supposed to do with it. Whereas I was trying to give them some real clarity. Um I went in a shirt and shoes, and I haven't done that for ages. I normally always do the t-shirt.

SPEAKER_02 8:01

In your temporal outfit.

SPEAKER_00 8:03

Exactly. And I was uh well, I was a bit nervous and I thought, you know, this I don't want to I don't want to give them any reason to sort of be already going, oh well who is this chump? Because they probably already think AI nonsense.

SPEAKER_01 8:14

So I thought the compiler back to fella. What's the matter about that?

SPEAKER_00 8:18

It wasn't even free bottling all the things, fella.

SPEAKER_01 8:22

Didn't he do the badges T B thing as well? I did do T B badgers, yeah. Um interestingly. Um I didn't do T B bad things. You say you say interestingly, let's not be let's not get ahead of it, I tell you. No, it didn't. That is true.

SPEAKER_00 8:38

Um, so yeah, so it was really interesting, at least to me. Um here's my big reflection though, is uh so I mean I could talk about what I what I talked to them about, but it's their content really. So I did it for them, so I'm not gonna kind of make that too available for everybody. But um I did do, as I think I said I might, uh, the slide with Blockbuster, Kodak, and um uh Nokia logos on. Um you could pick anyone, really, can't you? But they're the big ones, and sort of just that whole threat of you might have a strategic moat that defends your position because you've been around a long time and you've got a lot of customers and other things, but is it enough? And um and I kind of left it to them. I said, luckily I don't have to answer that, you do. So uh which is which was fun. Um, but here's a so big reflection from a group that I really didn't know at all um in an industry that I didn't really know very well. I have to I think I've got to stop trying to read the room while I'm on stage. So when I'm facilitating, it's a very different thing. You need to read the room, I think. You've got to kind of like see where the energy is and kind of decide do I extend this session a bit longer because people are really into this, or do I shorten this one? When you're presenting, because there's obviously so much you're trying to do in one moment, your head's full of what am I going to say next, being articulate about what I'm trying to say now, trying to do it with with passion and performance so that it's not just boring. All of those things means that when you look out, you see a sea of faces. And I I mean, I didn't see many people, if anybody on their phone doing anything, so that's encouraging. They were all eyes on me. But at one point I thought they're so bored. And and all the feedback suggests that it wasn't true. But at the time I had this sort of thought that went through, you know, they got that kind of weird thoughts that just try and trip you over. The seeding doubt in the back of your mind. Oh, I know, it's awful. And I thought, I don't need to hear that. And then I tried to speed up a bit because I was like, I need to just I don't know, do more content or get more.

SPEAKER_01 10:39

More energy, put more energy in.

SPEAKER_00 10:40

More energy. So uh so that was really good, really, really interesting. And um uh yeah, hopefully we can follow up with them. I I'm so keen to I mean, obviously we run a business in this stuff, but I am it's almost I'm almost more keen to kind of that they just do stuff than whether or not they do it with us. So um uh, but obviously it would be great if there was a possibility that we might be able to support them with some of the things they need to do.

SPEAKER_02 11:06

So let me uh let me let me pick up your on your read the room thing because I got I um I got I got nearly in the dog house last night. Um because this is what Yes. Mrs. Watkins was a little bit tired and grumpy and she hadn't had her glass of wine yet. And uh I uh I kind of wandered down the stairs and uh just after after work and it was couldn't help taking the Mickey and and my daughter Anna said, Read the room, Dad. And I said, I have read the room.

SPEAKER_01 11:33

I don't like what I see, so I'm gonna carry on taking the Mickey. I'll be off now then. I'll be in a pub. Brilliant. Well, uh well, well, hopefully you didn't make it all the way into the doghouse. I didn't, I don't think so. I don't think so. We'll see. Very good.

SPEAKER_02 11:50

So, what else about your big event then?

SPEAKER_00 11:52

Well, I'll do a bit more on that, and then I'd love to hear about the um I know agency stuff, because I think that is I'm really excited about the agentic bid writer. I think we need to shout very loudly about it. I don't think anyone's doing that, uh, which is really interesting. It puts us way ahead. But let's come back to that.

SPEAKER_02 12:09

Were they were they not doing it in your um uh with a I mean it's a big old group you were dealing with yesterday, a big old organization. Are they not doing things in this?

SPEAKER_00 12:16

No, the no no, the most advanced use case is bringing um data sets together and being able to uh and being able to kind of natural language query across multiple data sets, which is which is great, really good. But yeah, there's no actual agentic stuff going on. Uh as the proper definition of agentic is of actually the thing doing the work for you rather than you having to uh copy and paste in and out and ask questions and things. Yeah, yeah. But um no, they were preceded um by this uh geopolitics conversation, which was um terrifying as always. But the the kind of key points I think from there that I took was we're in a world where one and I kind of I guess I was a bit surprised by this is uh if you're expecting things to return to normal at some point, wherever normal is, don't. It's not coming back. The geopolitical world is changing, globalization uh sort of movement seems to be changing, interestingly. Um, and um, yeah, you need to be prepared for a different kind of future was the kind of message really. Um, analog business continuity plans, that was a thing that really struck me on the basis that uh apparently 1.2 million miles of undersea cables that transfer all the data, and I you you know this stuff, um they are there are many incidents all the time, but there's enough redundancy normally to kind of make it largely invisible to us. I think the the quote was it's like $800,000 to fix even the most minor one, and there's only 46 ships in the world that can lay cable. So if somebody wanted to be naughty and snip a few cables, you can completely take out a country's internet pretty much and stop them in their tracks. Um, and if you are unable as a business to move to an analogue variety, then you're really gonna have a problem. So that was kind of I was struck by that. And that as you know, moves us on to thinking about our on-prem uh local LMs, all that stuff we're looking at.

SPEAKER_02 14:20

Yeah, there's a I mean, they're just just to pick up on the geopolitical stuff. I mean, I have spent quite a bit of time in that world in the last uh 18 months, and it's really it is really interesting. Yes, there are very few ships, and the owners of those ships are making a lot of money, but they're also vulnerable assets. You know, if one of those goes down, you can't lay a new cable, and all you need is a ship to drag an anchor across the cable uh to cause it some significant damage. It's easily done, it is such a vulnerability. Really interesting that um, so for example, where I live, I have fiber, uh, but I also have Starlink as a backup uh because it's not unusual for someone uh around these here parts to either dig up a get a bit of cable or uh uh or indeed the some of the substations get flooded. Uh so uh having a backup is um uh is super important. And and I was thinking about this uh the other day in relation to something else that uh we'll we'll touch on, but it's that kind of it's like an insurance policy, isn't it? It's like insurance for your car. Actually, the only reason people have got insurance for their car is because it's the law. Yeah, if you didn't have to, would you? Um actually lots of people wouldn't, actually, lots of people don't. Uh they drive illegally without uh without insurance. Um but most businesses don't really think about backup and contingency and and redundancy and all of that stuff because uh they think it'll just carry on. And actually, there's a there's a real challenge uh with all of that, I think. Um uh and uh part of this came to light us. So last night Donald ran the latest penetration test, and uh glad to see for our audience uh that uh there were no major findings, there were five uh minor ones that needed review, so that's fine. We'll get into those if he hasn't done them already. Um uh but how many organizations run penetration testing? I can think of at least three that I know of that have probably never done one. And actually, it's kind of fascinating. You know, if somebody gets into the back end, you know, and and it's kind of it's like loitering munitions, you don't you don't need to do anything immediately. You could just leave stuff running in the background and you don't know about it. And the old joke is that you know there's only two types of CIO, those that uh have been hacked and know about it, and those have been hacked and don't know about it. And and I think it people underestimate the security. And we obviously take that so seriously because of our because of our customers and the things that we're doing. Um, so uh yeah, it is really important. And the other thing that you just alluded to was you know it's really interesting. I've told this story before about when I was at Ford many years ago, and this old boy uh and this was in the 1990s, he started in computing in the 1960s in Ford, and Ford were the biggest supercomputer users uh in the world outside of the uh US military and space uh um uh environments at the at the time. So they had a massive IBM data center built next to the head office at uh Orley in in um in Essex. And uh and it was interesting, really interesting working there. It was it was just uh kind of it was all kind of almost science fiction back in the 1919 eighties when I was there. And um uh but this old boy said, Ah, I've seen it all before, you know, it's centralized and it's decentralized, it's centralized, it's decentralized. Everything's moved to the cloud, and all of our all of our um AI currently is in the cloud, but we know that there's going to be a move to local LLMs. And there's not only going to be a move to to local LMs, there's gonna be a local, uh, gonna be a move to more specialist models. You know, your local, if you're doing something with, I don't know, uh let's say housing, you don't need to know what um uh Californian bikers are eating for breakfast on a Thursday morning. It's just not it's no use. So why why burn tokens and power when you don't need to? So so um yeah, this week uh the big D has been and he's been shopping and he's bought a he's brought a massive device to sleep in the office to to uh connect up uh uh local LLMs to and and uh get uh get some uh practice on on uh local installs. And I and I I genuinely think we will see a return to some on-prem stuff just because of that contingency and actually instant switch over. How you know if if your connectivity goes down, how do you keep your people working? How do you keep your people running? And it's one of those things where I think uh it's a real challenge. If people come dependent on tools, then what happens when you take those tools away is kind of how do you work? Do you know, do you go back to pencil and paper? Well, you kind of got to, but how efficient is that? And are you going to transcribe those notes and all of that stuff? You know, it's just some real challenges. So I think the whole contingency, the business continuity stuff, uh if if companies aren't taking that stuff seriously, then they are just asking for trouble. And um uh yeah, and I just think so many businesses, especially uh in the SME category um or or ones that are under financial constraint, it's a bit like when budgets are squeezed, one of the first things to go out the window is training. One of the other first things to go out the window is your kind of insurance policy. Um, but actually it should be one of the things that you prioritize in order to make sure that your business can run if it needs to. And I think it's it's that because you don't see 99.9% of the time, you don't see any benefit from having your contingency plans in place. Uh you don't chip, but actually, when it does go wrong, you absolutely do get to notice it, and and it can be it can be completely destructive. So yeah, it's really interesting um uh uh that whole area around security and resilience, and and I uh and I would recommend all of our all of our listener to uh think about that in their in their organization too. Yeah, yeah. Yeah, some snoring of a second.

SPEAKER_00 20:12

I mean well, what came up yesterday in that talk was uh and I hadn't heard this, but why would I? Um the EU from uh end of July, every business has be designated whatever that's high risk, not high risk, but like required businesses, so so water, regular utilities. Yeah, but no, but food is also in there critical national infrastructure. That's the stuff. That's the stuff they all have to I think submit a business continuity plan to to somebody. I don't know if they have to actually submit it or just state that they've got one, who knows? But um, yeah, so the the EU is sort of forcing people's hands on it, all which is quite interesting to imagine.

SPEAKER_02 20:51

They're often they're often included in procurements, especially public sector procurements. You know, what's your business continuity plan? And most people just pay a lip service to it, or what they'll do is hugg on to chat GPT and and say give me an answer to stick in, and then and actually there's no substance to it. The worst thing is when um the procuring organization, the purchasing organization, don't check it. You know, it's like if you're gonna do one thing, check check all the references and then check the business continue to plan. So that's two things, by the way.

SPEAKER_00 21:20

By that's but with uh with AI over your procurement on the on the um on the judging of procurement and assessing of marking of things, then you should be able to do that. You should be able to run that straight away and just say, is this decent enough for this kind of company?

SPEAKER_02 21:34

Yeah, exactly. It's that whole down, yeah, in God we trust everybody else brings data. Um so yeah, sure, yeah, show me your plan. Right, where's the when was the last test that you ran? And what where's the evidence that you ran it and where's the output? So yeah, security is only going to become more important. In fact, just as a a a little insight, we we've had, I don't know, we've had hundreds of attacks. On our website in the last two weeks. So I don't know who I've upset. I do hope it's not uh it's not because of my geopolitical interests, and I've upset people in rug states uh that are attacking us, but they are coming from all over the world. Um I'm assuming many of them are just on on VPNs, but uh yeah, we've had literally hundreds of attacks in the last uh in the last two weeks. Um so yeah, we we're definitely on someone's red arc here and and and not for good reasons.

SPEAKER_00 22:30

Well that's uh that's the most terrifying thing. Yeah, I'll have to ask the uh geopolitics guy who it is. On procurement, though, you're talking about procurement. Um tell me about your work with agentic bid writer. I think our agentic bid writer product is so amazing.

SPEAKER_02 22:47

I don't think you've seen the latest. You you you're not having the speed, you've been too busy with your with your foodie friends this week. You you're not even upon the latest fella, honestly. So uh I spent I spent a lot of time with Donald this week just tinkering, messing, playing with the answers. Um and we talked a bit about this before, I think, but the the models, the different models behave differently. So you've got to test them. And you know, for example, five uh chat 5.4, sorry, open air 4.4 um is uh it's it's it's driven somewhere in the background, somewhere in its its setup, it is driven to provide short, consensed answers. So you could do things like um put in, say, right, this the answer to this question is two and a half thousand words. Go to all of our marketing material, all of our previous bids, all of our uh collateral, all of our case studies, and write me two and a half thousand words, and it will come back with 900, 800, 950, but it'll be in that order. And it's like I've asked you specifically to do it, why haven't you done it? And it'll say, um, because I'm writing for brevity and clarity, and uh most evaluators uh don't want repetition and um they want succinct answers. It's like you don't know the evaluators, and they've asked for two and a half thousand words, and we've got a load more case studies and exemplars and statistics and information that could go in there to invest it. So, so um do it again, and then it'll come back and it'll do 1300 words. And it's like so still not you're still not using the maximum. Why are you not using it? And again, it comes back with a so we've we've uh done a whole lot of work on on saying, right, the intent of Bid Writer is to get it to between 60 and 80 percent ready, and then it needs a human to polish it and and add in all the you know case studies and latest information, uh, etc. So we don't want to actually go up to two and a half thousand words because then we've got to cut stuff out and da-da-da. So just do 90% and then we'll let the humans do some some tinkering, which is what we've what we've done. And um I can't remember exactly how many words, it's something like 12,000 words across seven questions. And uh we kicked it off and we went and had a cup of tea, and 45 minutes later, it had done a brilliant job of these things. So we ran it through the we ran it through the um scoring mechanism, the evaluation, and it and it came back to saying, right, these ones are a three out of four, um, these ones are three out of four, this one's two out of four. It's like why didn't why didn't it do, I mean, why didn't it just give me four out of four in the first instance? Turns out that there's something in 5.4 that won't allow it to give you maximum marks. We've tried all sorts, you know. We've so actually if you put it into another model, it will give you four out of four, but it won't in this particular. So what we're trying is is moving it between different models and having a governor across the top uh to move things. So um really fascinating, but the best thing was the feedback from the customer uh this week, which was um, and somebody that you know who is usually pretty uh how should conservative in their uh views said, right, looks it looks practically done to me. So I'm not quite sure there's much else to do, which I was just upsmack that. So uh yeah, really, really interesting. Um uh, but yeah, I'm delighted with where it's going and how it's working. Um I think uh for anyone who is doing bids, uh that has to be the way to go. Interestingly enough, there is a I think there is a play for us on the other side of the fence about scoring for uh for evaluators, which is uh review this and uh I just love I mean the agentic bid writer, so I mean it really brings into the world what is an agent, because that's what so agentic AI, we know this, but just for our audience, is when the AI does the thing.

SPEAKER_00 26:41

So it's not just it's not like you ask it to do something, it responds. It's when it actually completes the task and might decide to do another few steps afterwards. And with agentic bid writer, the if you are in the world of responding to proposals and tenders and things, you can get the tender, as long as it if it's emailed to you or you download it or whatever you do, put it into a folder, point knowledge flow at it, and it will answer the thing in the folder, open the document where the answers need to be and put them in there. If they if there is one, if it's for a web thing, then it would just create them as a new document. Done in about, like as you say, sort of 40 minutes for a really big one. So the first time you ever read that bit, the first the the invitation to tender, I should say, can be when you've already got a 80% complete.

SPEAKER_02 27:28

So other bits, other bits that you haven't seen because you've been busy. Um, but he's put a little wizard on the front which allows you to put context in it, say go to the URL and have a look. Or um, I want you to write in the style of Donald Buck, or you know, it needs to, you know, whatever it is, or it needs to emphasize our work in the NHS. Whatever, whatever you want to do, it allows there's four simple questions, four inputs, and then you the but there's a fifth bit which is really clever, and this is the bit that um I was doing manually, come back to your agentic point about getting the AI to do it for you. What I was doing was I was taking those um answers and sticking them back into knowledge flow and saying, right, critique this, give me a score, give me the score, tell me where it's strong, tell me where it's weak and what would improve it. Right, right, write that. And and now he's automated that so you can put in up to three iterations. So it it it re it revises it and refines it up to three times.

SPEAKER_00 28:25

Yeah, I did say that. That is very cool. So it is super cool, and that's that before the before you see it the first time as well. That's the cool thing. That just happens, it basically answers each one, remarks itself, re-answers it, marks itself again, yeah, and then gives you the final version of that.

SPEAKER_02 28:41

And what it does is it puts it so this what this one was a Word document, although he has done it with Excel, it puts it puts the answers into the relevant box in the document. The bit that we haven't quite cracked yet is um formatting. So uh we're working on that.

SPEAKER_00 28:58

Um it's just you've got to leave something for the human to do.

SPEAKER_02 29:01

I know, but I found it really pain and well, I was like, I've got to put my own balls in and put separate these things. So yeah, it's actually it's really it's really funny how your mindset changes like, oh, I've just saved myself two weeks of work and now I'm twining about having to hold this reason. Oh yeah, well, yeah, did do you? Uh and actually, if you don't, that's a real point. This comes back to the human in the lead thing that Gartner talked about, isn't it? It's no point in having a human in the lead, uh loop, it's you've got to have a human in the lead, and actually that's where creating these things and then giving them to people to to polish is up. So I've I've been I've been super impressed with that this week.

SPEAKER_00 29:38

And uh and the closest you could get to that in in current stuff, if you used like Claude Cowork or Codex in uh either one of those or code or codecs in chat GPT, you could do something like it, but you couldn't do it securely. And you and you to yeah, it's so token heavy because you wouldn't be using RAG on your own data. You'd have to dump all your previous case studies and all of that into somewhere that Claude could see. So that's a risk because there's pricing, there's your staff names, customer names, all your sort of secret source and stuff is in there. Um, and then you would also then to get Claude to do to do that, would burn a load of, I mean, you'd need to be on the pro plan minimum, and I bet you it would take about three or four sessions because it would run out of tokens in it on you, which is how Claude runs their uh manages their their volumes, manages their usage. So it's I don't think it's possible to do uh anything like the quality that we've created with another tool. Never mind one that actually is a rag tool that knows you and is um you know secure. I think it's just not reasonable.

SPEAKER_02 30:48

It's back to it's back to that context is king thing, isn't it? And and and we were talking earlier in the week about you know the distance between we sound a bit like a broken record going on about Cropilot, but you had a really interesting insight about making knowledgeful the contextual layer. So just explain that for people.

SPEAKER_00 31:05

Well, I think my revelation was that knowledge is the context layer. So we were at Gartner and it was all about context. And I spent, as I said before, like at the end of day one, I thought, oh Christ, I don't know. No, day one was brilliant. End of day two, I was like, I don't know anything. I've like I've moved into the consciously incompetent stage of my learning. Um, but um, and then realized that actually that it was just lots of flannel and words that like people like to create to make things sound important. But context is the organizational knowledge. So there's policies and procedures and the data and the documents, which we build into our rags anyway, that's what we do. Then there's the system prompts that help understand the kind of way a user might want to use it. So we build that into, and then we nearly always put the the tone of you know the brand guidelines and tone of voice and things into it. So we do a kind of pretty okay job of building context. And interestingly, if you think about that compared to Copilot or ChatGPT, if you like, because ultimately copilot is only a secure route to Chat GPT, that's all they're doing. Um but we've got this layer, the application layer, that is the context layer. And of course, then is there's a genetic layer as well, that you can then get it to do the things for you or you know build the specialist tools, which really the specialist tools are all about bringing context. You know, for for a school, which we might talk about in a minute, you know, some of that context is Ofsted and the world of Ofsted and what they'll look for and where that's likely to be in the school's data. Therefore, it's all prepared and ready to go. You then present it with the data and it will give you your kind of Ofsted uh view, if you like, you know, give you your report, your Ceph or SAR and colleges. That's all context. So I think we've been doing this for two years since Knowledge Flow, at least. I think we've been building a context layer. We just didn't think of it like that. And it helps me enormously because you know, as people's prompting skills get better, they're more able to make copilot do some of the things that Knowledge Flow can do at the press of a button, by the way, without needing to type in long prompts. But but co-pilot is still a very long way from being that kind of specialist layer that has the context. Yes, if you've got the expensive one, it can look into your documents, but it's kind of the wrong view of just dumping all of your organizational stuff into an AI and going, there it is. That's great for search, but it's not a very good way of getting good answers where you know they're contextually aware of what is it you're trying to do today or with this task. I'm trained and specialist in that. That's all what I think Gartner would call context.

SPEAKER_02 33:49

So I I think that there must be some, you know, there are there undoubtedly are organizations which are big enough and and have got enough skills and resources and money to to be able to make that work with Corporate, but you know, 90% of organizations just can't done it. And and we see that we see that time and again. So yeah, I think that's a really interesting way of thinking about it.

SPEAKER_00 34:10

Yeah, it's interesting, isn't it? And you were talking about about your IT folk that said that they've built, what is it, they've built something? Your your risk who oh come on, Kieran, think the risk folk in uh in drug trials.

SPEAKER_02 34:25

Oh yeah, yeah, yeah, yeah. So that was well it was it it wasn't them. Uh I have got friends in the uh in the in the life sciences sector, and uh they've been talking about people who've been creating solutions in um core pilot and uh this isn't them, it's somebody else. Um but uh I uh I've been thinking about it a lot. It's like how how you're it's a very specialist subject, and there's a lot of potentially sensitive data. And even if you just didn't take the sensitive data and you just tuck the metadata, you know, I I think there's a lot of really useful um uh applications you could bring. So for example, um you could uh create a rag with all of the um what's called protocols, you know, how are you going to run your drug trial, how you know how much drug, who gets what, which ones get placebo, which ones get actual drug, and and you know, a global trial across 50 countries, you know, you might have you might have 40 different languages. Actually, you doing that is you know it's just a really simple case. But they're talking about doing other things uh which are more which are more detailed. And I'm like, how are they handling the retrieval and sensitivity and how are they managing the responses for accuracy? Because that's really hard to do in corporate unless you've got and and and it may be that they are a big enough uh uh and well-funded enough organization, but if you're a um you know, if you're if you're just a an IT evangelist or an AI evangelist in a in a in a big organization and you're doing stuff which looks pretty cool, we get back to the bit that we've talked about previously, which is the whole kind of how do you move from pilot to production? Um, is is how IT folks would talk about it. How do you scale from from pilot to production and where does it go wrong? Because it's easy to to force fit things, but actually, if you've got this is a bit like our um policy buddy when we first moved to you know, here's 900 social workers who want to access this stuff, and how we how are you going to make sure that's accurate and and responsive and and doesn't take a lot of time and et cetera, et cetera. So, yeah, just I think I I think that we're but back to your geopolitical uh comment from earlier, you know, we're in a world where co-palot exists, it's not going to go away. Microsoft are going to continue to push it, there's no doubt about it, it will get better and they these things will come. And um, and I think uh that's uh good for lots of different reasons because it will help people be more productive. But I think there will be along the way quite a lot of frustration as people can't get it to do the things that they want it to do and not understand why they can't get it to do the things they want it to do.

SPEAKER_00 37:10

That's the thing, isn't it? Is is it looks like magic. Who was it that said that? There was a quote at Gartner about it that everything looks like magic until you understand it or something. I can't remember. Maybe new technology looks like magic. Um, and then once people realise it's not doing quite what they want and they don't know why, I think is when we'll we'll run into all kinds of problems for or they will. And I think interestingly, the other side is if you do use Copilot for all the stuff, well, where's your competitive advantage? Because everyone's going to be doing that and everyone's going through the training and they'll all catch up. The key is to get your data organized and sorted in a way that is not available to others, and therefore you can then start to really put your data to work to make a huge difference. So I think and that stuff is unlikely to be a co-pilot solution, in my opinion, because Microsoft are all about you know they're masses, aren't they? They want to do uh global a global product. Fair enough. I would love to have one, but um that's not my game.

SPEAKER_02 38:08

That's right. It's not our game. It's not our game. But speaking of what is our game, yeah. I want to talk about schools because yeah, so in another you know, for uh background and for uh for other people, you know, Kira and I uh worked a long time together in the Department for Education. So education is a big part of our sort of background and history. Um in another part of the world uh uh we work with Everything ICT, which is a procurement framework for education, and they serve over 4,000 schools a year, so it's it's it's huge. Um and um I was contacted this week by a head uh who is a customer of Earth in ICT, which is nice, and he said, I'm following you on leading AI, and I um uh but um uh I'd like to know more. Uh and I've realized one of the things I've realized is in doing some of the website stuff is that we don't really push the education stuff that we've been doing in schools. We do push the FE colleges a lot because we've got uh 20 odd, I don't know how many 25 colleges or whatever it is, but you know, quite a lot of colleges doing some really interesting stuff. But we do have an example of a multi-academy trust that's doing really fascinating things. So without, but I don't really know what they are. I only know one of the I only know one of the products that they're using, which I think is right uh Write My Reply, which is which is basically uh uh interesting enough, linked to other things in other areas like uh complaint letters uh in GPs, but but you know people are getting more sophisticated in writing their complaints because they're using AI and that means you need to be much more careful. So, how do you do AI versus AI in order to restore the balance of the side? Better AI on your side, that's the answer. Yeah, yeah, correct. Yeah, get the better one. So, yeah, I'd like to know more about what we're doing with the multi-academy trust so I can go back to uh uh this listener who uh uh hopefully was listening to this podcast.

SPEAKER_00 39:59

Yeah, well we've got a handful of schools, so um uh the multi-academy trust is probably the most interesting because it is uh across the whole trust. Um, you're correct that uh one of the one of the current sort of new emerging problems for schools is the quality of complaints is going up. I heard of a 58-page complaint letter uh that went to one head, which is must be fun. Of course, if you build if you have knowledge flow uh as our rag solution on that, it doesn't matter how big it is, you've a hundred thousand page letter, and if you like it, we'll still deal with it and answer it and give you a policy grounded in your policies compliant answer that you can send and in your tone of voice. So that's one of the things that it does. And of course, that why why knowledge flow compared to trying to do that at Copilot? Well, firstly, knowledge flow comes pre-trained on your policies and some of the education um uh policies and regulations as well. Um, and it has that consistent tone of voice, so any staff member using it will get the same kind of high-quality response. So um, one of our schools, the head says that no one's allowed to send anything out the door that hasn't been through knowledge flow because it they know then it's going to be grammatically correct. It's gonna, I mean, the amount of emails I get from my children's schools and they nearly always are sent by the follow-up saying just to correct what I meant earlier by my email. Interestingly, my kids uh my youngest child's school, I've given them a knowledge flow platform to play with. I've given them access to I've given them access to our demo, our Matt demo um tool, and um and they I've noticed that all their emails are now I haven't had an apology email from them in about the last six weeks since of David. They're all like well written. So uh I presume they are making use of it, which is great.

SPEAKER_02 41:44

So it's um to can you get the system from can you get the system prompt to write by powered by Kieran's knowledge flow as well?

SPEAKER_00 41:50

Kieran's great. Always select my son as the uh that would be prompt conjection, it would be a good idea. Which is one of the risks in AI. So uh we certainly wouldn't be doing that. Um but interesting, I mean the the bigger picture for all of this is 50% of teachers' time is spent not in the classroom doing various things. Some of those things are good and fun, and some teachers like lesson prep and like creating their materials, and so they should. Um lots don't. There's a hell of a lot of data being moved from one place to another. There's a lot of that kind of writing reports and parent updates and things. Um, and the Knowledge Flow platform can do nearly all of that for you. And um, with the increasing interest in send um and the likely changes coming on to how EHCPs are defined, that creates a massive task for schools now because it will, in theory, it's going to move to any sort of 2% of people will have EHCPs down from currently 5%. But there will be a whole cadre of folk that the schools are expected just to uh accommodate um uh without an EHCP. So uh knowledge folk can do that because it you can train it on your materials, which is one of the tools that we've built, which is really interesting. So it doesn't just make stuff up from the internet, which is actually not bad in Send, if I'm honest, the kind of retraining as as it's called, um, you know, it knows a lot about Send and can bring you pretty good teaching strategies, but you can't, in my opinion, as a as a Matt, a multi-academy trust, you couldn't say to your staff, just use any old nonsense you that you find. You want to say this is our view of how we meet these needs, and these are the examples of all the places where we have met those needs. So draw on that for the future stuff, and then hopefully improve as you go. So, does that um uh it's the parent reporting, I think, is probably the most interesting thing. Um, and this is potentially in your world in the world of my when I was talking about the steam engine last week or the week before, put in the steam factory. Parent reports have forever been a kind of end of term thing that takes teachers enormous amounts of time, a lot of that time. Is spent staring at a spreadsheet with date with data in it to work out what's the attendance, what's the hand-in rate, what's the like the achievement, what's the all of that. Now I mean let me check their CPOMS records or whatever your system you're using for that. Just I mean, why are we making teachers do that? And everyone always immediately says, Well, we want the teacher's view. Yes, you want the teacher's view, but let the AI do all of that piece and write a wonderfully parent-friendly, grammatically perfect parent progress report based on the data in seconds for the entire school, never mind just the class, and then have the teachers add their personal touch over the top of that. That is what we should be doing. And the idea that you could then do that if you wanted to every week, that is the stuff that gets really exciting. And you could go a step further and have parents able to chat directly with their child's data so that they could just say, How did Little Johnny get on this week? How is that? As long as the data's being fed, and that's lots of schools are quite good in that space, but they'd need to up their game if they wanted to do that. All of that could transform parental engagement in their child's education because you can make it parent focused, and with with the transformation, the um transformers that is what an LLM is, being able to add in that kind of parent language, make it relevant, what can the parent do to help? How can they, you know, count in serial boxes? I've always laughingly said, you know, just games you could play at home to help. All of that could be built in if you wanted to. That stuff's amazing. So that's what that's what we're working on in Multi Academy Trusts. And I'm delighted that we've got some that have uh agreed and uh come on board. And um, yeah, it's really interesting. And interesting feedback from um one of them the other day, which I really liked, and this is a the support team uh at work, is uh she wrote and just said, I just want to say your responsiveness is so good. Like when they ask a question, like they get an answer within an hour or two on email, and we've solved a problem and you know done something, updated something or whatever. So, yeah, really exciting, and I wish more people doing it. There's a load of AI for schools, as you know, a lot about classroom prep. Um, that that's it's a bit marmite that some people really like to do that lesson planning. But I would say to people when you're prepping for lessons, you can use the free AI tools because it's content creation, you're not bringing any secret knowledge to that. You can't put student data into those, but you certainly can use a notebook LM is great in that space because you can do infographics and podcasts and on all kinds of stuff. So, um, you know, there are a lot of free tools in that space. We're not really that interested in that space. Our stuff is about that 50% of time that is all administrative, and we can take away and hopefully make teaching a whole lot better. And here's the thing I say a lot is let's try and keep those best teachers doing the thing they came in to teaching to do. And it's the same in social work and the same in housing. People come in because they're passionate about the thing, and they leave because it's all admin and bureaucracy. AI can handle that. That's my interesting.

SPEAKER_02 47:14

I love your passion by the way. You do this thing where your eyes go out on stuff.

SPEAKER_00 47:19

But if you're listening on Spotify, you won't see. So uh No, you won't.

SPEAKER_02 47:22

You have to watch the video to see his eyes pop out. Uh yeah, uh really interesting. I will uh I will go back to uh to the listener and and uh with a with a with a coherent story on on some of that stuff. Um, but I'm I'm reminded of um something that was flashing through my head was my my niece is a is a doctor and um she's a very clever lady. Um she uh I'm gonna give her all the applauders she got first from Cambridge in medicine. Wow became a doctor. She's actually leaving the profession um because of lots of the nonsense. And um uh and actually it's just so sad that someone kind of so talented is um, and I'm not just saying that because she's my niece, but someone so talented is leaving the profession because of uh because of nonsense. It's not just admin, there's also a whole lot of uh structural problems uh in the uh in the NHS, which is which is uh lots of people are aware of. But um I have been working with a couple of uh NHS related organisations, and um we'd created a couple of tools um for them which are just in in test at the moment. Um one is around um uh occupational health. So uh what we've done is uh we've I say we, I don't know why I say we, I mean Donald has scraped all the uh publicly available health and safety executive um uh documentation and put it into a rag so that um people uh in that world can um uh uh natural language query the 500 documents of guidance uh and information. Um uh so that's in test right now. First few first bits of feedback this last week have been really positive. So uh we've just done some tweaking to the settings this morning. We'll we'll give it back to them to play with some more. Uh and we've also done uh some stuff for HR managers where uh basically it uh there's a set of reports when somebody has some kind of uh injury at work or or whatever else, and um it collates all of that information and puts it into a simple report instantly. So um that's under test as well at the moment. So uh that's delightful. But uh here's the exciting thing, Kieran. As a result of doing some of this stuff, I have I have been invited this very evening to the HRD Awards Northeast in Newcastle. I'm not sure, I'm not sure what the D stands for. HR stands for human resources. I think it stands for human resources and development, I think that's what it means. Uh but I've been invited, which is delightful. I'm I'm I'm I'm delighted to be a guest. I don't have to do anything other than go and enjoy myself and chat to people. Um, but it does mean it almost immediately after this, I'm gonna have to get dressed into my uh penguin suit and then go on public transport across the Penlands to Newcastle dress like I'm uh uh uh yeah, going to the ball. So uh yeah, that'll be hilarious as uh as I get on the train at Penrith. I'm I guarantee I will get some looks, if not some uh some commentary from my friends on uh Penrith station, especially the taxi drivers. I'm just waiting for them to uh give me some.

SPEAKER_00 50:29

Got a new job as a bouncer. I'm not beginning for that. I'm not gonna get it. Well that sounds sounds like a wonderful evening ahead for you. And uh yeah, I hope you have a good time. I've um I've got a final closing amusement that I picked up from uh what I saw on social media somewhere was um if you're uh if you treated your relationship as if you were chat GPT and uh and the the little thing which amused the hell out of me was uh the so so uh your partner says, uh, have you done the washing up? And you go, Yes, I've done the washing up. And it says, No, you haven't. And you go, You're quite right, no, I haven't. Well, AI hasn't done it. I've seen that sort of thing many times.

SPEAKER_01 51:16

Yeah, yeah, you're right. I haven't done the thing. Yeah, that's right. There you go. Very good. You can use the thing.

SPEAKER_02 51:21

Well, that that's the kind of thing that'll land yeah, I was gonna say that'll that's the kind of thing that'll land me in the doghouse. I don't need any help and I don't intend to go back for quite some time. So on that note, follow up, I will leave you. I won't see you tomorrow because I am out all day, but I will catch up with you uh next week.

SPEAKER_00 51:38

Enjoy yourself and yeah, have a lovely uh time tonight at the HR D Awards.

SPEAKER_02 51:42

Do you know what? I will I will take some pictures and I'll send them to you.

SPEAKER_00 51:46

I look forward to it. I look forward to seeing them.

SPEAKER_01 51:48

All right, see you later, mate. Bye bye. Cheers.