Introduction

Thank you all for coming on a Tuesday evening in February. Thank you, Hannah. Thank you, Mindstone, for having me.

I'm actually on loan from the Bristol Mindstone chapter. So I've come up from Bristol tonight and gave this talk in January, but it's already out of date because that's the kind of world we're living in. But I'm also giving a talk this Thursday in Bristol.

So if anyone's in Bristol on Thursday, I'm giving a talk called did something big just happen which is a riff on matt schumer's talk or essay from last week and i think we are living through an extraordinary phase again for those of you who are familiar or been in the ai trenches for a while will realize that this week or this month again is proving

um seismic uh i'm running out of adjectives pivotal uh it's omg week every week at the the moment so part of the reason why I enjoy AI is just trying to keep up but what I'll try and do in 10 minutes is rattle through the principles that I use as a consultant for selecting AI use

cases going forward my name is Mark Riley I have a boutique AI consultancy in Bristol called Matheson and there is a prize of a book my book called the AI profit available on Amazon for anyone that can figure out where the name Matheson comes from by the end of my talk in 10

Who I Am and Why I’m Here

minutes no cheating but my background just in terms of credibility I ran the innovation team for Dow Jones in New York for 10 years which includes we published the Wall Street Journal so Victor's talk was super relevant to them and in 2016 my boss the CEO Will Lewis asked me to set up an AI lab and to do an audit of the building 10 years ago and find out what was going on of

Of course, not a lot was going on, but I went to Silicon Valley, met over 110 AI vendors and startups, and then had to figure out how do we adopt AI across Dow Jones globally and which use cases do we take forward.

So we probably did over 200 workshops and came up with a bunch of ideas that we then took into production. This was up to 2020.

Came back to the UK, went back to study AI at Oxford Said Business School, which I highly recommend and also did a 10 -week remote course at London Business School which I highly recommend and set up Matheson so a lot of my clients are media but we're now moving into kind of mid -tier medium -sized companies working for legal and risk companies as well so if we

Who This Talk Is For

have time I'm just going to go through 10 principles and then if anyone's got any ideas they want to bounce off me over pizza please feel free to do so so I'm going to talk about

But trying to dress for personas might not achieve that this evening.

But essentially, this is for the tinkerers. So these are people who just work from home or they're just trying stuff out. They're using LLMs for the first time, maybe dabbling their toes into Claude Code. They might start building product.

The AI wizard is potentially a CTO at your workspace. It might be the one guy or girl in the office who's extremely good at AI, who everyone's leaning on, trying to get some knowledge out of. you might be an AI consultant like me

or you might be the crazy the absolutely crazy people who are now AI entrepreneurs doing startups

so quick hands show of hands, who's a tinkerer who's setting out as a tinkerer

great, who's an AI wizard who's the office wizard

yay, any crazy entrepreneurs oh my goodness

and then anyone none of the above, anyone not classified classified. Okay.

And any venture capitalist investors, probably mess them off. Okay. VCs, these are the, we're the entrepreneurs, VCs here. Cool.

Quick Examples of Things I’ve Built

So just again in terms of credibility, as a tinkerer, I built a website, fun, purely for fun, called shelfimage .co .uk.

You can put it on your phone, take a picture of your bookcase, and it will give you a character analysis based on your reading history and it's quite snarky and then it'll

give you further reading recommendations so somebody built for fun we built a rag tool for American publisher called American Institute physics that talks to their archive so physics researchers can tap into their archive we built an

insurance contract review tool for a client in Bristol and I've got my own startup called Hannah which is a newsletter builder that works similar to perplexity it builds newsletters on any given topic I just want to look into

The Opportunity Space and Market Trends

to the opportunity space, this is from an excellent sub -stack called Rule of Thumb trying to highlight where the opportunity space is at the moment

and this is based on annual recurring revenue and we're seeing a lot around creative and voice 11 Labs, a great UK success story customer service and user research, health and bio and legal

but this is what's really really impressive to me and this is what makes this totally different to the dot -com boom

of 2020 is that in the dot -com boom a lot of the companies did not have revenue they were just raising on speculation and a promise whereas in the AI boom we're seeing massive AIR numbers

so we're seeing uh mid -journey and cursory 500 million AR within within two to three years and lovable already 100 million AIR after two years and these are phenomenal numbers so So this is for real.

And then this HBR report is slightly out of date already. I found this fascinating that the transition in use cases we're seeing across consumer -facing AI applications has gone from the 2024 much, much more around a kind of Google substitute.

It's for research, it's for looking at analytics. And what we're seeing now is a lot of people using it for therapy, counselling, support and friendship, which you can interpret how you like but it's interesting that the behavior is moving that way

quite rapidly and then in terms of market trends in industry the first use cases out the blocks two or three years ago was sales and marketing which is essentially just a chat bot sitting on your website we're seeing big advances in health care legal and advisory some very high value

startups like Harvey are taking away legal work finances things like fraud detection education for personalized learning media which is my world for everything from an article for research through to distribution logistics risk and your supply chain and your suppliers supply chain and robotics is coming up

Principle 1: Stay in Your Lane (Start with Domain Strengths)

fast but one of the key messages I want to get across if I can is if you're choosing an AI project to do, try and stay in your lane. In other words, lean into what you're good at, bring your domain knowledge with you, don't lead with the technology,

lean with your background and your expertise.

Principle 2: Fall in Love with the Problem, Not the Technology

So first of my kind of 10 principles is fall in love with the problem, not the technology.

So we're all here for an AI session, great. We're not here to sort of discuss problems necessarily, but really figure out where the problem is either with your own workflow or your colleagues workflow or your potential customers

workflow and the exercise we do with our clients is to first sit them down for a week or a day long workshop and do a forensic analysis of their workflow and really figure out where the bottlenecks

are and it's kind of one of our golden rules that boring is beautiful so the more boring the problem the more likely that a human being will want to offload that problem onto a machine rather than taking away the exciting bits so if you can identify a really boring problem that's holding people up in their day -to -day life that's a great one to start with and then again this is

Principle 3: Start with Data (Proprietary, User, or Public)

reinforcing the message of stay in your lane stay in your domain strengths to do a startup in ai or a project you've got to start with the data everything starts with getting the data organized getting it out of its silos, getting it AI -ready.

The most valuable data anyone can get is proprietary data. So this is data that sits within your organisation that only you can access.

So, for example, Insistro uses proprietary lab -generated biological data sets.

If you don't have proprietary data, it's not a problem. You can use user data. So we're seeing a lot of companies lean on the data they've already generated the exhaust they've got from their customers it's

already giving them enough data to build products and duolingo to learn off user behavior and previous courses if you haven't got user data then you're gonna have to go for public data and there's plenty of public data out there there's

all sorts of data sets you can buy you can beg you can borrow or you can steal and there's plenty of companies I won't name them that have gone and got data data without asking permission.

But again, the value comes from combining the data and the domain knowledge with the skills.

Principles 4–7: Build for the Future, Legal Readiness, Unit Economics, and Moats

And then this is a Sam Altman message from a couple of years back. His message was that build for where the models will be in two years, not where they are now.

Because if you're just building what they call an AI wrapper now that's just leaning on the current technology, chances are you're going to get wiped out out in the next two years because it's going to do the models the LLMs will catch up so try and think ahead to where they'll be in terms of memory in terms of reasoning in terms of intelligence

and head off to get there a fixer is a very cool UK startups raised 30 million in series b it's got a 70 80 million valuation but it does email automation for you so it reads your emails and suggest replies gemini it's plugged into gmail now how long's flick fix are going to last so you've got to think ahead in terms of where you're going to be where you can defend your idea so this

is the sam altman quote i think there are two fundamental strategies to build an ai startup right now you either bet the technology is about as good as it's going to be or you bet the

technology is going to get massively better for uh legal and compliance frameworks we used to build product in Dow Jones and then six months after we built the product we used to go and check if it was okay with the legal team that didn't wash 1so you need to get your legal ducks in a row before you start building and in AI that means is it aligned around your policy

your company policy can you mitigate for hallucinations and there are technologies now that can do that with automated fact checking and then is there a human in the loop so nothing gets put into production without human oversight and then a lot of people still

think we're working on SAS revenue models so software as a service revenue models you've got you've got HubSpot you've got Dropbox you've got all the good SAS models out there phenomenally good business models once they built the technology their cost of sales cost of service was practically zero they could

add customers without seeing any incremental cost so they had operating margins of around 90%. It's a big mistake to think that you can run an AI startup on similar margins.

If you're lucky, an AI startup will see 40 % margins because you've got to bake in the inference costs. So as you increase your customer base, you're increasing the load on the chatbots and on the reasoning, on the inference, and you have to scale that into your

business so AI business models at the current standing of cost is quite hard to do to break even and then get to profitability this is a whole presentation in itself so I'll try and skip

through this quickly but I mean I get a bit annoyed now when VCs ask me what's your moat because I'm not I'm going to have the discussion I'm not even sure that anyone's got a moat anymore more. I'd argue that even the hyperscaler, just the open AI, I could argue that open

AI doesn't have a moat. But if you do want a moat, these are some of the broad themes I'd go for.

Go deep, not broad, to get very good at one thing, rather than trying to go horizontal and own the browser. As I've said before, try and find proprietary data. And

now you are your brand. So your brand, which is you and your beliefs and your relationships and your network is your moat. And then distribution.

So product is easy now. 12 -year -olds can build product. It's very, very easy to create IP. The harder bit now is distribution.

So if you can crack distribution, and you know how to do that, which is essentially marketing, you might have a moat from there.

Personalization is a good one, especially in the media world. If you can create a product that's highly personalized to the user,

chances are they won't leave. if you can create a feedback loop so it learns more and more about your user and becomes more personalized as a result and then one thing we're seeing with the llms is building in memory

so you probably notice now that if you're on chat gpt or perplexity it's starting to learn about you it's learning about your background and your hobbies and your habits and your interests it's quite hard to well you can export that knowledge but it's quite a good mode for them for you to feel that you have a sense of a relationship with that model and then finally

the bit that people often forget is that fine tuning your prompts is the hard part so if you're building an LLM that you're fine tuning

with post training it can take days if not weeks to get the prompting to a standard where it's hitting evaluations on a

human capability a lot of people will try and build a product it gets to 60 % human and they'll just give up, what you've got to do is

keep going and going and going and measuring against evaluations, your evals to get 100, well, 90 % is a good target. If you can hit 90 % against human performance, you've probably got a moat at that point.

Principles 8–9: UX Matters and How to Pick High-ROI Use Cases

UI is as crucial as AI. People think they can just plug in AI, but it's not UI. The user interface, the design,

This is a Gamma product. This is Gamma deck. This is a Gamma UI, perplexing UI.

And then this is an exercise we do with our clients in workshops. So we figure out what your definition of success is for AI. That can be a new product creating new revenue, but chances are it's an efficiency play.

So how are you measuring that efficiency? And then plot that against effort and then come up with your effort impact analysis.

And we try and steer our clients into the big ROI returns on high impact, high effort. and the other thing we try to do with our clients

is steer them away from thinking of AI as purely an efficiency tool and start to think of it as a revenue generator, incremental revenue

a new exciting product that does something new that will bring new revenue into the business and I think that's ethically important as well because

if you go in and just try and sell AI as an efficiency tool that to most people means job cuts so

who's going to lose their job but if you go and sell an AI tool as an incremental business, as a new idea, it's much better to sell it that way.

A Simple Framework: The Three Boxes of Innovation

So the tool I use a lot is just the three boxes of innovation. So box one is an easy problem. I think of it as assisting a human.

Box two is a hard problem to solve. It's probably the same data, but it's a new product.

So it's probably replacing a human. And then box three, which I also called the WTF box is superhuman so we're talking about robotics and quantum

Final Principles: Build/Buy/Borrow, Disruption, and Ikigai

two more to go if you're looking to do a product build something new you can either build it you can buy it or you can borrow it so if you want to build it that's great most people look tools to do that now you can buy product off the shelf or you can rent it using a subscription and there's a fantastic

fantastic website called there's an ai for that uh she's got over 40 000 um tools that you can look at to see if anyone's chances are someone's done it before right so it's good idea to go and look on there and see if it's happened and then this is the last thing i don't know if i was going

to include this but i will because it's kind of worth talking about i think is what's your tolerance for disruption as a builder as an ai practitioner and it's a serious question because if you do something well you're going to assist people in their workflows you're going to make their lives better if you do something extremely well if you're highly successful

and you disrupt an entire industry it will have huge impact on people's careers and lives so what is your tolerance for disruption to play in this field and you have to take that into account when you're building because these are dangerous dangerous technologies so it's worth a little bit of reflection.

And then finally, if none of that makes any sense, you can always come back to Ikigai. Who's heard of Ikigai? Great.

So it's a very cool matrix to figure out. Do something you love, what the world needs, what you can be paid for, and what you're good at. And if you can

crack those four criteria, you end up with your Ikigai. That's it. Do what excites you and what will you build?

Wrap-Up and Q&A

That's me. If you want opportunity workshops I'm available to come in on -site, work with you and your colleagues.

I've got a newsletter called MediaMorph and I've written a book called the AI Profit. Does anyone know where the name comes from?

Who said that? Well done. Thank you very much.