Introduction

This talk, this isn't slide one, this is slide zero.

Why AI adoption is slow—and what you can do

So I thought I'd chat a bit about why AI adoption is so slow and what you can do about this.

The enterprise reality

So I'm really talking about the enterprise level here, which is so most of our clients at Geisten are slightly bigger. And at the enterprise level, AI adoption is absolutely terrible, to be honest.

Evidence from recent reports

So MIT released a report recently. Who's read it? Who's seen the headline? Yeah, seen the headline, not read the whole thing, right?

But one of the headlines is, 95% of enterprise gen AI projects show no measurable P&L impact. That is pretty horrendous.

And a 2025 McKinsey survey also said that only 1% of C-suite leaders say their AI deployment is truly mature.

Upcoming reports and events

I need to put in a plug at this point, these facts and many more will be available soon in the Guyston Tech Leadership Report that is coming out soon and will be delivered at another plug.

BTF Plus is like the week after next. Be there. There's going to be loads of cool stuff.

It's starting with the Leaders Summit on the Monday hosted by us and I'm going to go through all these facts and more. But anyway, that's plug over.

Identifying the core challenges

That's the challenge.

Security, privacy, and compliance concerns

This is the reason why I think it is the case, which is that within enterprise organisations, people are very worried about data security. They're worried about IP leakage, so secrets getting out into models and then out to other people. They're worried about privacy.

And they're worried about compliance with regulations. So those two come together. So PPI, for example, like personally identifiable information, is a big concern for lots of people.

Disconnected data and tooling silos

Another reason is that the data in their organisations isn't connected to their AI very effectively. So even if you're using something like Copilot and everything in your Microsoft Copilot and everything in your Microsoft ecosystem is connected, then probably your CRM isn't connected to it and your ERP might not be connected to it.

Various other things won't be all connected up, which really limits what you can do with it.

Lack of workflow embedding and licensing friction

And the other reason is that it's not really embedded in the way people work, in the processes that people follow and the tasks that they're trying to do. You kind of have to context switch to use it effectively. You have to move around. And it's very difficult to get consistency because it's not embedded in the processes.

And on top of that is an extra challenge that I think a lot of people that we're speaking to just don't want to commit to that like £25 a month Microsoft Copilot license for everyone in their thousand person organization when they're really not very convinced that it's going to be delivering significant value. So that sort of puts an extra barrier in there as well.

A pragmatic solution architecture

So the solution, slide three, we're nearly there.

An organization-specific AI platform

The solution that we have made for ourselves and that we're deploying for our clients is essentially a deployment, an AI platform deployment that's specific for an organisation. It sits within, sort of attached to their network, attached to all their data stores.

Team-based permissions, knowledge, and live data

It allows team-based collaboration, so permissions on a team-based basis. can augment the models with specific bits of data for each for each team based on exactly what they need and you can share prompts between teams and that sort of thing and and then the next step is sort of connecting live data sources so that might be your crm it might be your your finance system it might be some obscured bespoke system that you've got in your business

MCP connectors: standardizing integrations

Using MCP connectors to connect. Does anyone know what I mean by MCP connectors? Yeah, a few people.

Okay, essentially, it's the AI protocol for connecting stuff up to AI, which is just really starting to gain a lot of traction this year.

From digital transformation to AI-native operations

Live demo: the platform in action

So, this is the risky bit. When you get asked to demo AI live in front of an audience, AI is non-deterministic, so who knows where this is going to go?

Multi-model portal and side-by-side comparisons

Right, so this is a portal where you can ask various stuff of things. In its basic sort of form, it just gives you access to lots of different LLMs, different models from different providers. It's got a kind of a funky feature where you can run two models at once.

So like right here, I've got a query that's going to go to Gemini Flash and Claude Opus. And you can do all the sort of usual stuff.

So like I have an idea for a... squirrel gym in central London, which I think will draw crowds. Can't type again, crowds. What do you think?

And then you get a side-by-side response from two different models, which is quite cool. So Claw's come back first. It's giving some ideas. It says it's quite unique.

Here's some stuff you might want to think about. Might not be such a good idea. Maybe think about animal welfare. OK, so maybe I should do that.

Let's see what Gemini has to say. Much bigger answer from Gemini. Took it a little bit longer, even though it's the flash mode. Came back.

Conclusion. Let's see what it says. Immense potential to be hugely popular. They really think... Honest opinion, go for it. This is really why you should not trust AI, right?

But yeah, so you can do that sort of usual stuff that you can do with any sort of model. But this is where it starts getting funky.

Augmented knowledge and bespoke models

So you can set up in here knowledge. So this concept of knowledge is so you can attach things to your model.

So here I've uploaded a few files, which is some case studies that we've got in the company that I've just uploaded and badged them as a knowledge thing. And then you set up a model in this context.

Oops, I didn't mean to do that. I meant to edit it.

It's basically an augmented model.

Model configuration: prompts, knowledge, and tools

So you give it a base model, so in this case I've got Claude Opus. You give it a visibility, so I've just got it for this demo group that I've given, so that's how you control it around Teams. It's got a system prompt, so it's kind of giving it a framing for everything that goes on in this model.

And you can configure some prompt suggestions, which I'll come to in a minute. So these are like things that you can just click on and do once.

And then you give it knowledge. So this is the case studies I uploaded earlier. These are a couple of other bits of knowledge floating around in the system.

And then these tools, these are MCP connections. So we've got it connected to HubSpot here. We've given it a bunch of capabilities and it's got some default features. So we've got that all set up.

Shared prompts for teams

And then you can also set up some prompts. So in these cases, you can create a prompt that everyone, so this is shared prompts that everyone in a team can use.

So this prompt, for example, says something about looking up some list of contacts that we have at a certain company based on what's in our CRM, and then find me some other contacts in the company that we might be interested in. And you can see the double parentheses here that's showing where you can enter a variable, and you'll see that happening in a second.

Putting it to work: sales assistant walkthrough

So let's see some of this in action. I'll remind myself of what I was actually going to do.

Right, so if I do a new chat, let me just do one model and we'll use this bespoke model. enhanced one so we've got our sales assistant and let me click on this so these are these suggested prompts that come in so let's see what's guys since public profile looking like today suggested prompt so now it's going out to the web is searching for some information and it's coming back with some answers let's see what it says so and what's it going to tell me uh appears strong and active there we go uh a bit of an overview oh it's founded by me that's nice to know uh um uh we won a best place to work award well done us uh we've got some of our size uh it tells what we're focusing on oh the state of tech leadership 2025 report it knows about that that's another plug for it and our current focus and a continuing to champion our values of innovation collaboration and excellence so there we go um

We've got some LinkedIn engagement too. Brilliant. So that gives me a quick flavour of our profile and what it's looking like. So that's just sort of a standard prompt.

Now let's try using one of those preset prompts that I put in place. So I'd put one about relevant experience. Here we go. This one's going to ask me about a company.

So we've done some work for one post. And basically this prompt says, with reference to your case study in order to give me a summary of the work we've done. I don't see a web search on this, so I'll just do it like that. And now this is going and looking at that augmented part of the model, so it's specifically digging into that.

It's telling us that OnePost, what their client, what they do as an organisation, the challenge, a little summary of what we did for them, and what we managed to achieve. All good, quite useful. If you're like a salesperson, that's a really handy kind of way just to dip into all the reference stuff that you might need.

CRM lookups via MCP and real-world variability

But let's see what happens if we're doing something a little bit more complicated, and this one may or may not go well. So this one... Sorry, I forgot the slash. So let's try and find some contacts.

So I know that we've done a little bit of stuff with BMW, so let's find some stuff on BMW. I'm going to ask you if some contact deals are going to come up in here. Please forget that you ever saw them for privacy reasons would be good. And hopefully they'll come up if it works.

So this is now using the MCP connection to HubSpot.

and also looking on the website, combining all the results and hopefully giving me a useful answer to the question of finding me the contact details we've got and look up some other people who might be interested in knowing about Geisten and the things that we do.

So we've got a nice long list of all the people that we know at BMW and now it's giving me a list of

Oh, well, this time it's not being particularly useful. This time it's just telling me how to look up these people and saying at the end, would you like me to craft an outreach message? Hasn't actually found specific people.

This is the variety of LLMs, right? Most times when I've done this, it's actually found specific people.

Conclusion

But it's saying, like, these are the sorts of people you might want to go after and then suggesting, like, an extra prompt that you can do afterwards.