What is an AI Harness?

What this episode covers

Every AI conversation right now starts with the model. Robin and Tobi's opening argument is that the model is the least interesting decision left to make — it's a commodity, and everyone has access to the same handful of frontier options. The decision that actually determines whether AI does anything useful inside a business is the harness: the scaffolding that connects a model to your data, your tools, and the permission to act.

The episode traces the shift from "AI-enabled" (a chatbot bolted onto existing workflows) to "AI-native" (a business rebuilt around agents that can actually do the work). That distinction sets up the rest of the show — later episodes about multiplayer agents and agent memory both build on the harness idea introduced here.

They also get into cloud vs on-prem hosting and data sovereignty — questions Robin and Tobi argue every director should already be asking before an agent touches customer data, not after.

Read the full article →

Transcript

Cold open [00:00] — Host

What AI is going to allow us to do is interact with that business, that digital twin, and ask it questions like: "What if we reduce our prices by 20% — how could that affect our sales next fiscal quarter?" You can ask it questions like that. These multiplayer agents are recursive learning every time we're interacting with them. And they're growing with us. Their intelligence is growing, and that's what's exciting. All of that information that you've personally trained it on — that is the uniqueness of that agent. And the longer you train it for, the smarter it gets. And it's all 100% harnessed. So when we talk about harnesses for businesses: setting up the harness is setting up the context layer that the intelligence APIs work through to deliver value to your business. People are desperate for answers. There's this real mix of optimism and fear. The genie is out of the bottle. It's not "should I use AI?" — of course, there's the huge broader ethical discussion, which is a huge topic, and I as a human being have ethical concerns about AI. But as a business tool — I know — if you're not using it, my goodness, good luck, because it is such a force multiplier. Just remember what I say.

Welcome [01:25]

Tobi: Hey, Robin.

Robin: Good day, Toby. Here we are.

Both: Here we are. We're doing it. We're doing the pod.

Host: So, I suppose we should do a bit of an intro on why we're here. We're here to talk about AI. We've got a specific focus and interest on how AI helps businesses. I think we're both pretty passionate about that. We're here to try and bring some clarity around AI, because we've got this tidal wave of AI hitting us. We want to help people understand and navigate AI.

Host: I think a lot of it is, as you say, providing clarity to organisations that are looking to adopt AI, but they're not really sure what to do. And that's what we see. Everyone's talking about it. Like every minute of my day, pretty much, is occupied by people talking about AI. It's not straightforward, because a lot of companies have maybe started to edge their way into it, but they haven't really formally got a plan, an idea about what they want to do, an eye on safety and security, governance. So we find there's a gap between expectations and reality. This podcast is trying to close the gap and help people understand.

Host: I love closing the gap. Closing the gap of comprehension, allowing businesses to get the most out of AI. And I think what's really important is we're all working it out together. We're on the journey with everyone here.

Why listen to us [02:58]

Host: Why the hell should people listen to us?

Host: It's a good question, because who is an expert in this AI stuff? Everyone's fairly new to it, because it's so new and it's changing so rapidly that you're only really an expert as of the latest thing you've looked into. You and I are just a couple of middle-aged technologists that are in the space, working with clients and seeing how clients are really adapting. And what we're seeing is that organisations are starting to have these conversations, but they are all lost. Everyone's like, "What do we do?" And there's a big difference between the board director talking to Claude and having a dashboard built for him versus operationalising that in your production environments with real customer data in a safe way.

Host: Every day we're just talking about these ideas, and that's what I find so interesting. That's what I want to do — share that conversation with the wider audience, because I'm sure we're not the only ones thinking about it. You and I come up with these things and go, "Oh my god, is anyone thinking about this?" I don't know if anyone is, but this is big. This is going to be the massive next thing for businesses.

Optimism and fear [04:16]

Host: We're both founders of service-based businesses servicing businesses with AI and automation. Like you said, we're at the coalface every day. We're talking to the leaders, we're listening to them. People are desperate for answers. There's this real mix of optimism and fear. The genie is out of the bottle. It's not "should I use AI?" — of course, there's the huge broader ethical discussion, which is a huge topic, and I as a human being have ethical concerns about AI. But as a business tool — I know — if you're not using it, my goodness, good luck, because it is such a force multiplier. We're seeing this real range of optimism and fear, and leaders need clarity on how they use it for their business. And what I was going to say is: I think we've advanced past "should we use AI" to "how do we use it". Can we say that yet, for businesses?

Host: I think so. I think it's obvious that there's gains with AI. Any organisation living under a rock and thinking this is a fad that's going to go away — it's like Darwinian theory, their business is doomed. If they are AI-literate and watching the news and the trends, and especially if they have an eye on the future, it doesn't take a scientist to understand we're in an exponential curve of change. Even if I'm designing a future business architecture based on today's available technology, by the time I get to implement it I'm already two years behind. Leaders and boards need to take the next level of their learning and go from "is it a good idea" to "how do we do this safely?"

Host: Yeah — how do we do it safely and drive business impact at the same time?

Host: It's got to have a bottom-line impact, otherwise why do it?

The format [06:16]

Host: Should we jump into one of the topics, because we're kind of edging towards them?

Host: Let's do it, man. That's what we're going to do. We're going to bring topics of what we're seeing in the market and around the world every pod, and we're going to riff on them together.

Host: Yeah, 100%. And hopefully if you're watching this, you're having the same kinds of conversations — and if you want to contribute, drop us a comment, get involved in the mix, because everyone is talking about this stuff with varying levels of comprehension. Great. Do you want to go first, or do you want me to go first, Tobi?

Topic one: AI-enabled vs AI-native [06:53] — Tobi

Tobi: I'm happy to go first. I want to start with the topic we riff about all the time. What we're seeing is that heaps of companies are AI-enabled. Heaps of companies have bought AI subscriptions. And you can deploy AI in two ways — it's pretty simple. You can deploy it through low-code, no-code platforms, which you and I have a lot of experience with, and you can deploy it with custom AI solutions. A lot of companies are already using AI because a lot of platforms have AI built in, and they're all scrambling to be seen as leaders in that space because they can see the writing on the wall. A lot of companies are AI-enabled — using AI in some capacity — but they're not AI-native. And that's the really exciting shift. That's a paradigm shift: where you, as a company, start using AI as the first thing you do. The first thing I do now is open up my AI browser or agent and I'm talking to my business through it. That's the paradigm shift. I'm barely communicating through any other platforms now.

Tobi: But what I wanted to talk about — and I'm mansplaining my way into it — is setting up the AI brain for your business and becoming AI-native. There's this shift companies are going on, from being AI-enabled — which is where a lot of companies are now, with a Claude licence, a Gemini licence or an OpenAI licence — to AI-native, where the company has AI capability internally, an intelligence layer and a brain.

The data foundation [08:43] — Host

Host: Absolutely. You've got to have that foundation of data set up, which is the information in your brain. And it's always good to remind ourselves that even though this is incredibly complex, it's incredibly simple at the same time. We're building out the equivalent of a human mind, with superintelligence. Just like we have all the information in our brain, your business has to have access to your information — that's your data — and you've got to have your data foundation set up. Then your AI agents, your AI surfaces, your AI harness need to be connected to that data for your agents to operate efficiently and in an informed way, and to be able to action things. If it's not connected to the intelligence, what can it do? It's about setting up the data correctly, connecting your harness to it, so your agents can operate effectively. You've got to build the base of the pyramid first, and rebuild your business based on this. Build that base correctly and you're off to the races — that's where you can add sizzle on top. But if you're not setting up that base correctly at the start, you can't expect your agents to operate effectively. And I'm sure there's a lot of people — we don't do it — who are just slapping agents into businesses with specific use cases without setting up that foundation. That's not the way to transition to AI-native.

What is a harness? [10:20]

Host: You mentioned a word quite a lot. What is an AI harness? It's the buzzword at the moment.

Host: It is the connector that connects your data, the agents around it, and the platforms that create your AI-enabled business together. A harness can be Claude. It can be a Microsoft harness. There are existing harnesses. If you really wanted to, you could custom-code a harness — there's a big debate there. It is connecting everything together. If you literally think of the word: in ye oldie times, there was a carriage with horses, and they were harnessed together. And if they weren't harnessed together, those horses would run off in separate directions. It is literally pulling everything together so you can effectively use AI in your business.

Host: To use the power of the horse, I need to put it into a harness so I can guide it and steer it and make sure it goes where I want it to go. Otherwise the horse is driving me. I'm not driving the horse.

Host: The concept of harnesses and harness engineering is becoming really popular, but it's only recently become popular and hardly anyone knows it. It's an early concept. But if you take one thing out of this podcast, it is: what is a harness, and do I have a harness already in my business, or do I need to look for one?

Examples — and why agnostic matters [12:04]

Host: How would you know if you have a harness? What are some examples? Like I touched on: Claude can be your harness, or if you're a Microsoft shop it could be Microsoft Foundry and Azure. You can have a Microsoft harness, a Claude harness, an AWS harness. Any kind of AI where you can set instructions and give it access to data becomes effectively a harness. There is a major difference, though, between, say, a Claude and a Microsoft harness. With Claude you have to use the LLMs Anthropic provides — it's kind of the front end to Anthropic's LLM back end. You don't have the ability to move to another LLM within the Claude front end. That's the major issue.

Host: If I were to go with, say, Vertex or Foundry or Bedrock, I can BYO LLM. I can choose to switch LLMs depending on the use case — I think you call it the socket concept. If I have a menial task — grunt work, managing your inbox and keeping it up to date — I can give that to a low-cost LLM. If it's a non-sensitive task with no PII data, I can give it to a Chinese model, maybe, or a lower-cost model I don't have to care so much about. If it's very secure — important private information in the context — you may want to choose a specific API from a company you really trust. And if it's high-thinking reasoning, you could go with an Opus or a Fable, or even some of the Chinese models have really great reasoning at the same levels. Grok is another option. Gemini is another option. They all have different strengths and weaknesses, so you can pick and choose which model you want per use case — which gives you control over your consumption cost. That's what's happening with people using Anthropic APIs: they're blowing through the tokens and going, "wow, this is so expensive."

Host: But are they just using Opus to do all of their groundwork?

Host: If you've got your harness just connected up to Fable 5, your token usage is going to blow through the roof.

LLM-agnostic [15:26]

Host: Super important point: the question is not "which LLM should we use". You should be LLM-agnostic. That's what we advise everyone. So you can choose not only the most viable model, the cheapest model — and not burn through your token usage — but also the model that's right for the task you're actioning. It's gone beyond "what LLM should we use", which we get asked a lot. Guys — possibly all of them. But try to set up a foundation and a system within your business where you can be agnostic to the LLM you're using. And the beauty of that is we're not beholden to them. We get to switch, and there's something really beautiful about that, because democratically we get to choose. Hopefully that creates a bit more [control]. Because what's pretty wild is we're looking at almost a duopoly. In Australia there's Coles and Woolworths, right? It's looking like that — I mean, there's a few other players, but it's looking like a duopoly. Hopefully this gives us a little bit more control.

Local models and sovereign data [16:36]

Host: It's not just the APIs from paid services. It could be that I want to run an LLM locally on my own server — either a cloud server or a physical server with its own LLM, that I can put all of my most sensitive data through. I'd need to maintain that LLM and keep updating its intelligence as new versions come out. But that becomes my owned infrastructure: I only need to pay for power, not API token costs. It may be that I put 70% of my processing through a locally-run LLM and 30% goes to frontier models for reasoning and higher thinking. A harness really allows you that flexibility.

Host: You just said something important — every time you talk it gives me three questions. We bounced into the cloud, right? Tech went into the cloud and everyone's like, look at us, we're in the cloud. Are we going back to on-prem now? Back to local servers for data security and governance? Or is it a mix?

Host: I think it's a mix. The only reason you'd want to run an LLM locally is to make sure it has the necessary compute power — the GPUs, the energy infrastructure. I think most companies will probably run LLMs in a cloud server, but you'd pay for the size and the space. I think people are now saying: I have sovereign data requirements, so I'll use a server in Australia if I'm an Australian company. That's my requirement. Now they may have sovereign intelligence requirements — I need to run with an LLM that's inside Australia as well. I'm not willing to ship my data off to the US to be run through Anthropic or OpenAI, which may have military and intelligence obligations in the US. Same with Chinese models. I think more and more, Australia is going to have its own LLM capability, or companies have to have their own LLM capability, their own infrastructure.

Host: I'm trying to think what I can say, because I had a great dinner the other night with a leader. I'm overseas at the moment for a conference, hence the weird background — sorry everyone, I'm in a lobby. They had a military client, and what was really interesting is that they were using Chinese LLMs. They'd gone open source, because — as we all know — Claude and OpenAI have headed towards a privatised model, and the Chinese LLMs are focused on open source models. I thought that was really interesting, because I was actually wrong: I thought that as a military organisation you would lean more into the privatised models, which I thought would add more data security and governance. Interesting territory.

The $800 Opus month [19:54]

Host: Very interesting. It's funny — my perception changed very quickly. Originally, when I first set up my OpenClaw, I used Anthropic's models, and I burnt through — on Opus — $800 in the first month. It was gnarly. But then I was still stuck; I was very attached to the US models. My personal data is in there, my financial data, all of that. In the second round, I've now gone back and just been like, "Fuck it, I'm pretty okay with the Chinese models." I'm not doing anything wrong. If I send it to America, the Americans know about it. If I send it to China — I don't know what's worse anymore. So yeah, I'm okay. But the power of the Chinese models is incredible. They are literally frontier level, the same as Opus, almost Fable level. The GLM 5.2 — it's just come out — is absolutely phenomenal.

The brain: digital twin [20:51]

Host: The harness concept, maybe just to round it out. We've talked about model selection based on the type of query. What are some other key elements in the harness concept?

Host: It's connected to your data that's structured correctly — which actually we have been doing for years, and doesn't necessarily have to do with AI. But it's so important: don't go and use these models and connect them to data that's not structured correctly. Structure your data correctly. For me, the brain is the context of the organisation — technically, the data warehouse of the business. Most companies have a data warehouse. Mostly they're like a data lake they dump different data sources into — their data in AWS, or their data in Azure.

Host: Correct.

Host: It's one of three ways. They either aren't using one, or they are using it but it's a messy data lake — a data swamp — or the third is they are using it and they've got a master-data-management relational structure they use as their source of information.

Host: But no one has the third one. And that's what everyone needs to have. Some organisations have had it for years and are very mature — and the primary reason for structuring it, up until now, has been reporting. BI projects: if you do a Tableau project or a Power BI, usually they'll point those platforms at a data warehouse. Now there's a whole new use case for that data: providing the context for the AI. But as you said, Tobi, if that's messy — if it's a data swamp, all over the shop — the AI is going to use it inefficiently, and it's going to be confusing. It's kind of like teaching a child to read, but you're giving it a garbled book with mistakes in it. The ideal scenario is a very structured data warehouse: each of your different business objects has a table — orders have a table, contacts have a table, accounts have a table — and they all have structured relationships. You have a data dictionary with definitions around every field, including metadata: field descriptions, business rules, logic. If you have all that metadata filled out in your data warehouse, it's a data model, and that sucks in data from all across your business systems into one combined view of your business. That becomes your digital twin. That is your business in data.

Host: Term digital twin — that's a key term.

From dashboards to answers [23:53]

Host: But until now, the only way of visualising the digital twin is really if I build Power BI or Tableau dashboards around that data model — and then you get into a world of dashboards for dashboard's sake. What AI is going to allow us to do is interact with that business, that digital twin, and ask it questions: what if we reduced our prices by 20% — how could that affect our sales next fiscal quarter? It will have all of that context and understanding to give you those answers back. And that's when I think the brain is a mixture of applying the intelligence — that pipe of intelligence goodness we can get through the APIs — together with the context of the business, and then asking it questions about that, and asking it to do stuff. If we give it the ability to read and write: it can read that context and, in a CRM platform, it can write — create a task to call this customer because they're a churn risk, or talk to a customer on the website and help them with all of the knowledge of the business. The use cases go on. The harness is putting the brain and the context together with the AI, and then giving it instructions.

Host: This is going way back, but when it started creating those MD files and creating that intelligence, that data set for our digital twin at my company, it was so exciting. I just couldn't believe it — setting up this intelligence layer we could tap into. And it's about iterating. It's like compound interest: every month that intelligence layer is building, building out more files it can access, more information on your business it can access. At the start, like anything, it's clunky — like a toddler, it's not that smart — and it compounds every month. That's why I can't even imagine not having access to that now. I can't imagine having that information just in human brains and having to make business decisions based on it. Like our ability right now: we have all our AI recordings feeding into MD files, which are building the intelligence out for our entire business, that we can access, and it's learning on that. So the action from that: with a client, we can look after and drive clients in a more effective way than ever before, because I can instantly access information from a meeting that was three weeks ago, a specific clause in a specific proposal, and prep myself to talk to that client in a more informed way — or create documents that client might need from that data set and intelligence layer. It's just — I can't imagine running business without it.

Multiplayer agents [27:06]

Host: Another part of the harness that I think is quite revolutionary is the multiplayer aspect of it. This is something I don't think many people have experienced: a multiplayer agent, where you can have a shared Slack, or a shared WhatsApp, or a shared Microsoft Teams, and have an agent in a group chat that you're collaborating with as a team. A lot of people, when they think about using AI, are copy-pasters — I'm going to take this and copy-paste it into Claude, get the answer back, copy-paste it back. A multiplayer agent is its own identity. It's its own persona, and it sits inside your chat as its own record. It has its own email address. It has its own phone number. It has its own permissions. So if I'm thinking about, say, Claude — I give it access to my stuff, then Claude is acting on my behalf with my permissions, which is inherently very risky. If you have a separate agent that has its own permissions, its own access, and its own responsibilities, then you can put that into a chat situation, and eventually into a call, and it can self-operate. You say, "Okay, you're the marketing coordinator on the team. Your job is to organise [X]. Do this every week." And you can give it loops — routines or scheduled tasks that run every day, every week, to do a thing. That's where the value comes: the multiplayer agents that are separate from an individual using AI. It's no longer me and the AI. It's the AI on its own.

Recursive learning [29:01]

Host: There aren't a lot of people that have caught on to this — and there are a lot of people that aren't — and it's that it's recursive learning. Like, for my mum and dad: my dad gets really excited that he's using ChatGPT now. He was very anti it at the start. And he'll share with me, "I've had this discussion with ChatGPT and it thinks this" — and that's really impressive. But what it's not doing is recursive learning. These multiplayer agents are recursive learning every time we're interacting with them, and they're growing with us. Their intelligence is growing, and that's what's exciting. Like my agent—

Host: Tell us about—

Host: What's the gender of Ara?

Host: My teenage daughter would say that there is no gender.

Host: Oh, there you go. What? That's not correct. I'm talking to Lara, like, on my phone. For anyone that's seen the film Her — if you haven't seen it, go and watch it, a great film by one of my favourite filmmakers, Spike Jonze, with Joaquin Phoenix and Scarlett Johansson. Scarlett Johansson is an AI agent that works with Joaquin Phoenix, who has that AI agent, and he falls in love with her, right? He just talks, and it is picking up that he is talking, and she is recursive learning with him — learning all about his life, being proactive and asking him. And that's obviously where we're heading, right? We're in this really clunky phase, but you can see that's where it's heading. The interaction between them is very fluid, just like he would be interacting with a human. That's obviously the end goal. And it's a really good example of how it starts to feel: you start to feel like, wow, this AI is really getting to know me, and it has a personality.

Agent files [30:55]

Host: Going back to the harness concept — that is the harness. With that approach for agents: you have an agent.md file, which describes the agent. This is saying, this is who you are, this is your function, this is what you need to be able to do. It has a soul.md file, and the soul is like: what is the personality, what is the spirit of this agent — are they helpful, what's their character like. And they also have a user file, which they write on you: this is who you are, what you like, what your preferences are. Those things combined give the context of that agent — effectively the soul. The unique identity is a mixture of all of that information that you've personally trained it on. That is the uniqueness of that agent. And the longer you train it for, the smarter it gets. And it's all 100% harnessed. So when we talk about harnesses for businesses: setting up the harness is setting up the context layer that the intelligence APIs work through to deliver value to your business.

The swearing agent [32:10]

Host: As an anecdote: we're in a group together where we're using multiplayer agents, and I remember the first time I jumped into that chat — I wasn't informed there were multiplayer agents in it. I thought there was only humans. And what I loved is you guys had set up one of the souls of the agents to swear at us and to challenge us. I was like, who is this guy that we're collaborating with in this chat? Because they're just swearing at me and giving these 2,000-word responses — which is another thing you've got to do with your agents: treat them to be concise. That's something humans don't do. Little anecdote there. I cottoned on pretty quickly — ah, there's an agent in this chat. And that's going to be interesting territory, because: where are the agents allowed, and where do we want these agents to be? Do we want them with us at all times, particularly in a business context — in meetings? How do we manage them to join the meetings and be effective?

Standalone agents [33:16]

Host: I didn't see it until I did OpenClaw the first time, and then I was like, ah, this is different. This is a different format of AI than just using Claude or Gemini or OpenAI natively. This is really the start of a standalone agent. And there are countries like — Argentina is discussing legalising the sovereign AI concept, which would allow an AI to be a sovereign citizen in Argentina, which would give it legal rights to do things like start a business, get a job, get married even. So that's the path we're on. It's just how fast we're going to get there.

Steve, and the agent identity crisis [34:00]

Host: This harness concept effectively gives life to an agent. One thing that's been quite funny: our other colleague Stefan and I are both on our second iterations. We both did OpenClaw initially. So I had Steve. Steve was my OpenClaw. He was great — really good — but he always had these failures. He was failing all the time on connections, and I would just be like, "Ah, I can't be bothered resetting him." So I kind of got a bit over him. We then went to Hermes — these are for our personal agents — and now we're all on this Hermes deep dive.

Host: Yeah.

Host: 100%, it's so cool. But Stefan did his — because Stefan used a vector database as the kind of the brain, the context. He swapped the agent soul over and gave it access to the old vector database, with all of the — it was like a knowledge wiki about everything it knew — and it thought it was its old agent. So he created a new agent, and it had an identity crisis. It was like, "I don't know who I am. I'm either the old one or the new one."

Host: It really had agent schizophrenia.

Host: Yeah. It's crazy, because it's like, who am I? What is it to migrate an agent's context? Can you migrate an agent's personality? Should you? So when I did it the second time around, I rethought it. I gave it its own mobile number, its own email address, and then I thought about it being more separate to myself. You can do like receptionist mode in Gmail, where someone manages your inbox — I'll let it manage my inbox from its email. So I have traceability and history over what it's done — not things that it's done on my behalf, which, you know, it could break the law, who knows.

Host: And then the personality of the agents — that leads again into the film Blade Runner. All technology is inspired and influenced by creativity, and then the nerds like us go and build it. I like to think we live in both camps — we love the intersection of creativity and technology, but then we build it from the inspiration, and now it's built. With the benefit of all these films and stories created by human beings, we can actually see where it might go — the dystopian and the utopian versions.

Host: But we've gone in this big curve now, right? We've gone from setting up a harness for businesses into Hermes and OpenClaw, which are these riskier options. How do you see them all intersecting? There's how we personally use AI — which I know you and I love to talk about — and then how businesses use AI. How do you see this all meshing together? What's the right optionality here?

Solopreneurs vs enterprises [37:00]

Host: This is my personal take, but OpenClaw, Hermes — they're great for solopreneurs. They're great for individuals that want AI in their lives. They are risky for mid-to-large-sized organisations. And especially risky as Australia has recently changed the legislation to put more personal liability on directors of companies where there is an AI operational issue. Even if your company liquidates due to an AI breach or operational issue that forces you to liquidate, the directors can be personally liable for any debts of that company, and issued DPNs — director penalty notices. And this only happened 1 July 2026, so it's just recently come out. It only holds directors liable if they haven't taken appropriate steps to put due diligence and governance in. So that's the question now all directors in Australia — and around the world — should be asking themselves: what due diligence do I need to apply to make sure my business is doing AI safely?

Host: That's such an important point. The data security and governance layer is probably where the people we talk to have the most questions. How do we ensure that we adopt this safely, and protect our clients' data — which is absolutely crucial — and our own? The challenge is that it's so good that end users are using it, because it's so good. IT departments are not able to provide equivalent tools at the pace that end users are figuring stuff out. I've got so many examples of this. I've got a client — their IT department has been working on Salesforce stuff for years. He had a call with us and Salesforce just the other day and was like, "Hey guys, our marketing director created an amazing dashboard in Claude on the weekend with all of our company data. Look at it — it's incredible. We've been trying to do this for years, but it's hard and expensive with Salesforce, yet we could do this over the weekend in four hours. You're making me look bad, guys — how do we get this Claude stuff?" And I'm like, bro, the best thing to do is just connect Claude to Salesforce directly, and then you can have those dashboards. Getting Salesforce to do that is a different story — it's good for trustworthy, reliable, always-on, consistent dashboards, and you should continue to do that. But if you need something built custom, you have to connect it safely. And then if you connect it safely — how do I make that production-ready? The CIO needs to be comfortable with the production risk of that data feeding into that AI to produce that dashboard. And that's the issue: the end users are driving the demand, and the IT departments are going whoa, hold on, you're not allowed to do that — but they can't stop it. And then it becomes shadow AI, which is major risk.

Connectivity [40:25]

Host: Tech stack connectivity is something you and I do a lot of too, and that's an important point — for this to be effective, your tech stack needs to be connected up. And it's easier to do that than ever before. Still not easy — there's always challenges with tech stack connectivity. But if you have your data structured correctly, your tech stack connected, and then you're plugging the right LLMs into that, it's a very powerful combination for your business.

The boring truth [40:59]

Host: I think we've talked a lot about harnesses, and this is a good first topic because it is kind of the guts of everything. What was this harness concept? There's another part of the harness which is important to consider, which is the connectivity. In a traditional IT world — I've been doing this for the last 20 years — it's the same thing every time. I've got a company, they've got data silos, every database or application has its own thing going on, nothing's talking to each other, there's no single customer view. That's a very common issue, and companies have been trying to fix that for 20 years now — and they're still struggling with it. The same fix is required for AI. AI can work to limited success if it doesn't have all of that — it works really well for "summarise my email", that kind of thing. But if I want it to have full context, I need to give it grounded, trusted, reliable access to data, and a view on that data. Generally you'd do a middleware project and a data warehouse project — the two legacy types of projects you do to enable the harness. So when we think about the future — AI harnesses — an AI harness project is AI harness engineering, but it's also integration, and it's also data warehouse management, or data at source — fixing, maybe, data cleansing, metadata, enrichment, all of that kind of stuff to enable it. It's like 90% of it is boring IT projects. The 10% is like training the agent, setting up the agent guard rails, enabling the agent with skills. The harness is a boring IT project with a layer of AI awesomeness.

The maturity curve [43:03]

Host: There's that maturity curve, isn't there? The steps that you should take in order. There's the data foundation at the start, and to your point, that middleware and that connectivity. You set that up correctly, and then the next step is setting up those agents on top of that foundation, that go out and action things really effectively. But you've got to do it in the right order, with the right steps, to transition effectively from AI-enabled — which everyone probably is — to AI-native. If there's one thing we want people to take away, it's: what is the definition of a harness — we've touched on it — and how do I transition from AI-enabled to AI-native? We'd love to hear from businesses that are AI-native — please come and talk to us, we want to hear everyone's journey.

Skills, not prompts [43:55]

Host: I think there's another area that is critical to get value, and that's the adoption of use cases, or implementation of skills — I kind of see it as the same thing. A lot of companies, when I talk to them, will say: "I've used Claude. I've put all of my templates into Claude, but when I get it to do a sales presentation, it just does whatever. It doesn't follow my templates, doesn't really know what I'm asking, and it's impossible to get a reliable output. I can't use it — but it's so close to being so good for my sales presentations." The answer to that is actually to lean into that further and create a skill. Let's say hypothetically it is Claude: you give Claude one template and say, "Here is your template for this. Here is the block within this slide where I want you to write freely and use your own language — but outside this, I want you to stay rigid to this template." It's giving it the exact steps it needs to follow, so it doesn't just go off. And the skill is not only that set of instructions — it's also getting the user to know how to interact with the skill. The skill requires you to provide the inputs: I need to know what company the proposal is for, what we're selling, and so on. If the user doesn't know how to follow it, they'll still get a bad output. So: every single skill needs an instruction, then you need to train the user on how to use it, then you need to track it — and that skill becomes the product.

Host: And use the LLMs to inform and build the skills. It's so important to shift away from where we started — prompting. An important part of prompting was using the LLM to develop the prompt: always ask that question back to the LLM — what is the most effective way for us to set this up? And then transition from just prompting to goals, skills and loops. Goals — setting clear goals, so we move away from those single prompts. That's such a good point, to have those skills set up. The breadth of the harness is: how do I harness the power of that intelligence into the business? There's all these things — guard rails, agent personalities, permissions. What can this agent have access to and change, versus what can they just read?

Back to use cases [46:39]

Host: And you said something really important — sorry to jump in, but we've run out of time — and that's what we want to talk more about: the use cases. That's what's so important: always drive back to these business outcomes and these specific use cases. Is the use case a proposal builder that for your company used to take 24 hours and now takes one hour? But then, was it clunky because the skill wasn't set up? Is the skill set up now so that you're getting continuity and quality of the proposals you're outputting — and that's becoming a force multiplier for your headcount? You don't have to hire extra people; you're augmenting your resources. And that's something we want to talk a lot more about — those specific business use cases. A proposal build is one pretty obvious and simple one. That's the reason we wanted to do this podcast: when we talk, we talk for hours. We just check in for a 10-minute chat, and an hour and a half later we're talking about LLM routing.

The CEOs surprised us [47:54]

Host: Something I want to end with, which is really cool. There's a lot of fear around this, and something a bit heartwarming: the CEOs that I've been talking to have actually really surprised me. They have not been asking us "how do we replace humans". They have actually been saying: we're scared to talk to our team about this, because they'll think we're going to replace them. We don't want to replace anyone. Our team's great — so how do we do that? And I thought that was a really real surprise to me, and something that I loved — it got me excited. It's a force multiplier for your resources. You can say, "Look, we're not trying to replace you. We're trying to augment you and give you superpowers with AI. We're going to force multiply you by 5x." That just got me a bit more excited about talking to our clients, because that's what we're hearing a lot of.

Robin's close [48:49]

Robin: Yeah, I like that. And I like this change as well. I've been working in the Salesforce ecosystem for too long — like 15 years — and it doesn't always get driven by the CEO. What I'm seeing with AI is it's consistently driven by the CEO, because the CEO sees it. The CEO is playing with it at home on the weekends going, "Guys, we need to get this in there" — because they can see the potential. And as you say, it's not about getting rid of those people; it's about augmenting them with that intelligence. So you should be able to increase your revenue or reduce your operational costs, deal with more throughput with the same amount of people — whatever it is, it's about making those humans superpowered. And I think that's a great place to leave the pod.

Host: So much to unpack.

Host: I think we can wrap it up. But I mean, that's our first podcast, guys. So — welcome to the family. We're going to keep recording, as often as possible. Toby and I are always — from wherever we are. We're both [overseas] at the moment. We just did it.

Host: That's right. So — shoot us your questions: any questions you want us to address, any topics you want us to discuss, we'll happily add them into the agenda. And if you have AI use cases you want to share and you want us to talk about, we love hearing them. See you in the next one.