Anton : Kia ora koutou. Before we begin, I want to acknowledge that I'm joining you today from Naarm, Melbourne, Australia, and in the lands of the Wurundjeri people of the Kulin Nation. I pay my respects to their elders, past, present, and emerging. To those joining us today from Aotearoa, ngā mihi nui kia koutou. To those joining from across Australia and beyond, you're all very welcome.
Anton : I'd like to see us into today's conversation with a karakia called Whakataka te Hau. It's an ancient karakia from Tainui. It's a fishing karakia, but it's about being mindful of the cold winds from the south and the west, but optimistic about the day ahead and that it will be calm, clear, and bright. It's a karakia that is about arriving with intention, and I think that's a perfect way to start a conversation about leadership, and in fact, a series about leadership.
Anton : Whakataka te hau ki te uru. Whakataka te hau ki te tonga. Kia mākinakina ki uta. Kia mātaratara ki tai. E hī ake ana ta takura. He tio, he huka, he hauhū. Ko Ranginui e tu iho nei. Ko Papatūānuku e takoto nei tūturu whakamaua. Kia tina! Haumi e, hui e, tāiki e!
Anton : Well, tena tatou katoa. My name's Anton , the Chief Delivery Officer at Allen and Clarke, and your host for this leadership series. We've called this series The Quiet Lifting, reflecting both my personal philosophy about leadership and also the important work that leaders do in the background — the quiet aspects of leadership, the paying attention, the asking of the second question, carrying the weight that your team's not yet ready to carry on their own, elevating others. It's about the thinking, the planning that goes on in the background, grappling with the challenges of the day, and even those looming on the horizon.
Anton : The aim of the series is to support leaders to elevate their thinking and understanding by surfacing the evidence around critical issues that are not necessarily well understood. Each session takes one challenge that leaders in our region are genuinely grappling with, brings the evidence to it, shares the experience of other leaders, and hopefully will give you something that you can act on on a Monday morning.
Anton : So joining me today is Adam , our Chief AI and Innovation Officer. Adam spends much of his time inside the problem we're going to discuss today, helping organizations to close the gap between AI adoption and AI value. He's got strong views about what's actually getting in the way and strong views about what we should be doing about it. Kia ora, Adam. Welcome.
Adam : Kia ora, Anton, and kia ora koutou. It's great to have you all here. Yes, strong views, so maybe we'll get a little bit fired up later on. Let's find out. But I'm really looking forward to this first session in the series. Obviously, I'm looking forward to chatting about AI — could talk about it all day if needs be. But I'm really keen to get into some of the questions that people have submitted and really get into tackling, or helping others tackle, that gap — and action what they can do to get the most out of AI.
Anton : Fantastic. Thanks for that. Well, let's get into it.
Adam : Let's.
Anton : So I want to start by getting you to picture a scene. A senior leadership team sits down to review their AI rollout. Eighteen months in, licenses have been purchased, training's been delivered, the strategy's been published, but the productivity dashboard shows no measurable uplift. There's been no service improvement, no cost reduction. The chief executive looks around the room and says the thing that every senior leader in this position says: "The staff aren't using it properly." Now, that conclusion is wrong. It's one of the most well-evidenced things in the entire AI leadership literacy right now. And the cost of believing it is that your organization spends another year addressing the symptom while leaving the real cause untouched. So today, we're going to name that cause — where the gap actually comes from, what's sustaining it, and what you specifically can do about it.
Anton : So first, let's look at the scale of the AI opportunity. Anthropic published the "Labor Market Impacts" report in March of this year. For administrative and management roles, the theoretical AI capability sits around 90% of tasks. The share of those tasks where AI is actually being used sits around 10%. So that's an 80-point gap between what the technology can do and what it's actually being used to do inside most organizations. Not because the technology isn't good enough — it's more of a value capture problem, and it gets more expensive every quarter that you don't close it.
Anton : McKinsey Research found that as of 2025, something like 88% of organizations were regularly using AI in at least one business function. So that means that three years into the generative AI era, adoption is essentially universal. But only around 6% of organizations qualified as what we call high-performance, meaning that they could attribute more than 5% of their earnings before interim tax to AI. So Adam, that's quite a stark gap relative to the opportunity, of course. What do you make of that?
Adam : Yeah, I love this Anthropic chart. It's like my favourite — yes, I'm a nerd. It kind of clearly shows what they believe, or how capable they believe AI to be. Yes, they are trying to sell a product. So even if we say, okay, you halve it, right, you're still talking about a 40-point gap, which is still a massive opportunity. And so it's just a really great chart that shows how much of the capability we're collectively using. I think it's worth us all reflecting on where we sit within that chart.
Adam : Going back to your point, Anton, on how high-performers — those in that 6% — they're doing something specific, and so everything we cover today is about what it is that they are doing, and how do we collectively join them. And if you're wondering, like us, like we've wondered, why your organization isn't seeing the gains from AI, you're not alone. NBER published a working paper surveying senior executives across the US, the UK, Germany, and Australia. What they found was that nearly 90% reported no measurable productivity impact from AI over the last three years — which is pretty amazing — despite 70% of their organizations actively using it. So there is lots and lots of evidence out there of organizations like ours deploying these tools. But the leadership view is that we're not yet seeing the return from the additional cost that we've invested.
Anton : Yeah. Right. So no measurable productivity impact over three years suggests that the gap is persisting. Does the evidence offer us any insights as to why that's the case?
Adam : Yeah. So another McKinsey study here. They did a global survey — it's a little bit old now, January '25. They surveyed over 3,000 employees across six countries, and that included Australia and New Zealand. And they asked both groups — the C-suite and the employees — the same question: "What's the biggest barrier to AI adoption in your organization?" And what they found was that the C-suite were more than twice as likely to name employee readiness — so that's skills, willingness, attitudes, et cetera — as the primary barrier. Whereas the employees themselves kind of indicated that they were substantially ready. And that from their perspective, what was missing was leaders steering fast enough, leaders visibly using AI themselves, and leaders creating the conditions to turn that adoption into value.
Anton : Yeah, okay. And that's interesting. So does that mean that the people most empowered to fix the problem are also the most likely to misidentify it?
Adam : Yeah. That's exactly it, by a factor of two, according to that study. And I think that kind of aligns with our lived experience through the clients that we work with, but also internally as well. And that's what we're going to chat about today.
Anton : Okay. No, that's good. And so it feels to me like that statistic's really important, and that it matters because it kind of sends the response in the wrong direction. So you potentially fall into the trap of commissioning more training. But in doing that, the real barriers go unaddressed and nothing changes.
Adam : Yeah, that's right. I'll just jump in there quickly, if you wouldn't mind. One thing that I would say is that my favourite analogy for learning to use AI is actually learning to drive a car. There's only so much training that you can put into learning to drive a car, right? Because you read the road rules, and then the best way to learn is by sitting behind a steering wheel and wiggling the wheel and so on and so forth, right? And it's kind of time on the road that gets you to the skill of being able to drive. AI is essentially the same. And so when we talk to organisations, we hear a lot about, oh, more training. And I'm kind of like, well, but you wouldn't say that if somebody is yet to learn to drive. You wouldn't say more training. You'd say, "I'll jump in the car with you and let's go for a cruise."
Anton : Yeah, okay. So I like the metaphor, and it's the structured training, right? So you've got a driving instructor, essentially.
Adam : Exactly. Maybe you're starting in the car park with Mum or Dad in the first instance. Graduating up from there.
Anton : Exactly, yeah. Okay, well — so what else do we know about the other things that are inhibiting the adoption, uptake, and appropriate use of AI?
Adam : Yeah. So over the past two years, we've been kind of mapping what we believe is slowing AI value capture down across — obviously this is in NZ and Oz — across the public sector, professional services, and kind of regulated private sector organizations. And what we've landed on is that there's kind of 15 frictions across four domains. So the full framework hopefully is on your screen, but you can download it afterwards. The point here is that there's three of those frictions that sit upstream of everything else, and they're leadership-level decisions. And they contribute to the other kind of 12 frictions that we've got on that slide. So I'll touch on a couple. We're not going to go through them all.
Adam : Friction one is that there's no clear answer as to what your strategic intent is. And not only do you not have a clear strategic intent when it comes to AI, but you also don't have an explicit risk tolerance that aligns to that strategic intent.
Anton : And this is, sorry, this is specific to AI?
Adam : Yes. Yeah. Right. Absolutely, yeah. So when you're thinking about your strategic intent, that might be efficiency, it might be service quality, it might be cost reduction. These are all very valid strategic intents. The point is that you need to have one. And you also therefore need to calibrate your risk tolerance to your strategic intent.
Adam : So I guess my point here is that most organizations or most leadership groups haven't clearly defined what they're optimizing for. So within friction one is a critical decision about who leads AI. What we see is that AI ownership often defaults to the IT team. And I'm not saying that's wrong. I'm not here to kind of hate on IT teams — they're critical. But the point is, you've got to think about how your IT team is actually measured. So often we see that it's around kind of security, risk management, this sort of framing. And as you can hopefully start to see, that produces a compliance frame around a tool that we see as a capability tool. So, if you contrast that with an operations-led ownership, what you're focusing on there is a capability and service delivery frame. I'm not saying that it's wrong. What I'm saying is you've got to make it as a deliberate choice, because that deliberate choice has implications, again, for the delivery of your strategic intent, risk management, et cetera.
Anton : Yeah, fantastic. Okay. So, if I think about that scenario, we see that phenomenon play out with other types of projects, transformation initiatives, et cetera. The leadership piece becomes quite pivotal in terms of how the approach is framed, but also the kind of value that is being generated from it. So I think in the scenario that you were just describing there — technology-led initiatives can tend to focus on technical performance, whereas another part of the business might place a greater emphasis just naturally on business value, for instance. So, I wonder then if the discipline of specifying the strategic intent of AI yet also serves as that natural prompt to identify and appoint or nominate the role or the function that's responsible for it. So, there seems to be a natural linkage between those two things. Is that what you were also saying?
Adam : Yeah, that's exactly what I'm saying. So the strategy frictions must all kind of work in harmony. And so that's that one column on the left there for those who can see the slide. So they need to work in harmony — for example, if your organization has a strategic intent to increase the breadth or volume of services delivered, then you would need to accept a higher level of risk around AI use. Because what you are actually saying there is that the frontline staff are going to need to use AI systems, AI workflows, to deliver those services more efficiently and therefore be able to deliver a bigger breadth of services. Obviously, if you're giving them tools, then you have to accept that they're going to, at points, use it in ways that you probably didn't plan for. So that's really the point that I'm making — they all need to work in harmony with each other. And so your first job as a leader is to calibrate that stuff for your organization.
Adam : So, that's probably enough on that — moving on. The next one that I just wanted to highlight should be on the far right of the thing on your screen. It's friction 12. It's where leaders are not visibly using AI themselves. And if you think about how this works — if the chief exec or department head is not using AI, then their direct reports will use it sparingly or cautiously, and that has a cascading effect down through the organization. So, couple of implications. One, your adoption's not going to work as well as you would like it to. But also, and this is what we see when we kind of engage with organizations on this, shadow AI use is actually a really persistent theme across the organizations that we talk to.
Anton : Okay. So shadow AI use sounds incredibly ominous. I guess it's kind of like — for those of us who are not technology nerds or natives — what is shadow AI use, and is it dangerous?
Adam : Yeah. It kind of has cartoon bad guy in the alleyway vibes, doesn't it? So shadow AI is just essentially that your team, despite what settings you might have, that your team will be using AI to assist in their work. We find this super common. The point here is that that use obviously goes unmanaged from a workflow point of view, a QA point of view, et cetera. But actually, you also don't really know what they're putting into these systems, which again, if they're using publicly available systems, that's super risky.
Anton : Yeah. Okay. That's interesting. There was another friction I was interested in, so wearing my operating model design hat for a second — I'm interested in understanding more about friction eight on the slide, which is around the workflows are not redesigned. What can you tell me about that friction?
Adam : Yeah. So workflow redesign is super important. Basically what we see — and we put our hands up where we've done this as well — is you implement an AI system, kind of training, et cetera. But actually what we've done is we've just layered this technology on top of legacy processes. And there's a bunch of data on this, but one I'll point to is there's a McKinsey report that shows that AI-using organizations have done the actual work to redesign their workflows. And that redesign piece is key to the success of capturing that value. And the point here is that only leaders can manage the disruption, the temporary inefficiency, and the process alignment that this involves. And the process alignment thing is way trickier than you might think it is.
Anton : Okay. So, just recapping for a second then — so what you're saying is that leaders make the strategic decisions about whether you intentionally design the ways of working so that AI is integrated into the organization's operations and process flows. So, this is one we could talk about forever because this is at the heart of operating model design. But those decisions are investment decisions, obviously, because they involve funding, they involve resourcing, you're dedicating people to these tasks. Notionally, there's the possibility of needing to consider and even recalibrate organizational priorities. Yeah, big decisions.
Adam : Absolutely, yeah.
Anton : Okay. And then while we're still on the subject of frictions — I'm just going to introduce an interactive bit now. So, the poll on your screen is a quick friction assessment. Pick the friction you think is most blocking your organization right now. We're keen to hear about your experiences. So, we'll give the audience time to complete the survey. Any tips, or are you picking one of the frictions?
Adam : Yeah, I'm going to go for workflows. I just think, in our experience, even where you have an agreed process to do a thing, as soon as you start trying to align it with the team members that are applying that existing process, you very quickly learn that they're applying it in different ways. And AI will allow you to do that, to adjust for that, but not totally.
Anton : Okay. Interestingly enough, and I can see the early results coming through on the screen — because I was going to go with the no agreed strategic intent. Which currently is sitting at — by the way, thank you everybody for proving me right. All your bribes are in the mail.
Adam : Oh, I'm never going to hear the end of this. Okay, fantastic. So some interesting things coming through there, but I do think the strategic intent one — it's amazing, given that there is so much talk about the value that AI potentially can deliver, that the number of organizations that haven't as yet, or even leaders within teams, haven't as yet got to the point of specifying the value that they want it to deliver.
Anton : Yeah. And we see that time and again. And I think we've got to hold our own hands up here as well. We had a strategic intent. After implementation, our strategic intent has changed. So it's not a one and done, and I think nobody's perfect.
Anton : Yeah. Okay. Well, that concludes the diagnosis part of our webinar today. Now let's spend the rest of the session on what we can actually do about it.
What high-performing organisations do: four behaviours and tasks
Adam : Yeah. Agree. We've got two observable behaviors and two tasks that consistently separate the 6% getting meaningful value from AI from the remaining 94%. These are all leadership tasks, and they don't require additional tools or additional budget.
Anton : Okay. So, you can see the four elements that we're going to talk about on the slide, but the first behavior is that senior leaders visibly role model AI use. Before the webinar, you were telling me that the research shows that high-performing organizations were three times more likely to have senior leaders actively engaged in driving AI adoption. Which includes using AI tools. I thought that was interesting, and then you went on to tell me that leaders in particular that also share their experiences about learning AI and developing their own skill set — including the mistakes that they've made on that learning journey — they reduce the psychological barriers for their staff and team members and everyone else across the organization.
Adam : Yeah. That's right. So, there's a little bit of vulnerability, I guess, that in this space, you're asking the leaders to bring to it — sharing those mistakes, et cetera.
Anton : And conversely, you mentioned to me that leaders who perform with certainty about AI can sometimes create resistance instead, and that the ideal target that they're trying to create is what you call "enthusiastic skepticism."
Adam : That's right, yes. Which, if I understood it correctly, is about confident use combined with rigorous verification.
Anton : Yeah. Does that fit?
Adam : Yeah, that's exactly it, yeah. Okay. So, what we're saying is that when leaders model openly, they give everybody alongside and below them permission to do the same thing. Which I think makes complete sense, but — how do you do it? Yeah, one of the things I'm curious about is — you're the Chief AI Officer at Allen and Clarke. What are some of the things that you model in terms of those AI adoption behaviors?
Adam : Yeah. So, I'll put one example here. I have set up an AI assistant, what I call a board of advisors. And it's basically an agent that kind of channels different personas. And the great thing with that is, I can openly consult it in conversations with the team, et cetera. So we can kind of have it listen into our conversation, ask it a question, and it will kind of come back — and mine is designed to be very blunt and highly critical on purpose. The point though is that, that is showing that I use it all the time as a thinking partner. Right. And the key here — why I like this — is because everybody knows AI can punch out some text. It can punch out a doc, it can create a PowerPoint, whatever. The point here is that you can use it as a thinking and challenging partner, and that's a use case that people kind of don't necessarily expect. But it's also a use case that's super easy for people to kind of engage and react with, because then they're literally able to be like, "Oh, and what do you think about this idea?" And it will give them a response.
Anton : Yeah. Okay. Interesting. I like that idea and I like the open use in front of others. I seem to recall that you hosted a previous webinar which spoke about the board of advisors, and you can kind of calibrate the composition of the board as well.
Adam : Yeah. So I have Anton in my board of advisors. So yeah, you absolutely can calibrate it. As you touched on in a previous webinar — which I'm sure the team will send out and give you a link to — there is literally step-by-step instructions. There's a video tutorial on how to set it up. All the data is there to set one up for yourself. And the AI team here also at Allen and Clarke are happy to jump on a quick call and help you set it up quickly for yourself so you can experience it.
Anton : Fantastic. I'm not sure that you would've got any useful advice from the Anton representative on your board. But so Adam, once again, the second behavior on the list has piqued my delivery and operating model interest. So behavior two is about changing your team's perception of QA — quality assurance and the human verification. So what's the rationale behind this behavior?
Adam : Yeah, so the reason that I've kind of picked this one is it plays across a couple of the frictions that are on there. Obviously, it plays to the workflow redesign — we all have QA processes, et cetera, like existing QA workflows. So it plays to the workflow thing, but it actually also plays to the workforce anxiety friction as well. And so the point here is that even the most advanced AI agents, they still need humans to direct them, to verify that the work is of expected standard, and make the judgment calls that the complex work that we all do often needs.
Adam : The issue here is that many organizations that we see are using the wrong mental model for the human's role in this process. So often I hear things like, "Oh, it's your job to spot whether the text is AI-generated," or to check the outputs. But the research essentially does not support that at all. The research shows that only a human expert reading content from their domain is actually only able to detect accurately what text is AI-generated or not about 55 to 75% of the time.
Anton : Oh, okay. So now obviously everyone listening is like, "Holy smokes." What that means is that AI-generated text is kind of going through your QA process at the moment.
Adam : Yeah. Now that doesn't mean it's an issue, but the challenge is that you don't know to what extent it is an issue, and that's the core challenge. We're actually almost — not really unfortunately — building a training game based on this research that basically shows people how hard it is, increasingly how hard it is, to spot AI text. And give them tips and tricks and that sort of stuff of how to do it. So yeah. Obviously, we'll share that with everyone once it's set up and running. We're doing that for our internal purposes. But obviously, we'll share it.
Anton : So what I'm taking from that is you're saying that the human role is not to become a detector of AI as such. The human role is to, I guess, own the judgment and run the verification of the judgment elements. The reasoning, the logic, the decision-making recommendations that are going in. Those are the things that you want the humans to be responsible for.
Adam : Absolutely. Yeah. Exactly. So that's a very different kind of human in the QA loop. And at the moment, your QA process is what I call a bit of a vibe check by a senior expert. It's not a criticism — ours was exactly the same. And what they're checking is: does this feel right? Does it sound right? And the challenge with that is that in the AI world, "does it feel and sound right" has no bearing on factual accuracy anymore. And I guess what I would say, lastly, is that what you're shifting to is a verification. Right? So is the content generated accurate to the original source material? And that is both a mindset shift, and it's also your first workflow shift. So your QA process sits across everything you do. And you're not training people to spot AI text. So the quality comes from designing the workflow so that the claims, the sources, et cetera, are all verified.
Anton : That makes a lot of sense. And again, just to recap. So, you're saying that the human verification shouldn't be framed as we need to catch AI out. You're instead framing it that we know that AI can help with these different tasks, drafting, synthesis, the structuring side of things. But it's important that the humans remain accountable for the judgment elements, the reasoning, the logic, the decision-making recommendations that are going in. Those are the things that you want the humans to be responsible for.
Adam : Absolutely. Yeah. Exactly. So practically, this means replacing your old review-by-reading model with two things we'll touch on quickly. So the reading check still needs to happen, but what I recommend is also layering on open questions. And you're going to ask those of the drafter, because what you're testing for is to what extent do they understand the content that went into the final product that is being QA'd. So a couple of examples: describe how you came to these recommendations; what alternatives did you consider, and why did you disregard them? And you can imagine if we were here having that chat, that if you didn't have access to the document or your AI assistant, and you didn't know the detail of that document, very quickly I'm going to detect that that's the case.
Anton : Okay. So that detection is pointing to calling out and highlighting the degree to which the drafter has really engaged with the underpinning logic and content. That's right. And then creating, in turn, a picture as to, I guess, how much reliance there has been on AI to do the thinking bit, which is the bit that we need the people to be accountable for. Gotcha. Okay.
Adam : Brilliant. And then the second thing is that we recommend using a source verification tool. This is not an AI-checking tool. AI-checking tools are hopelessly useless at detecting AI text — check the research out, they're just very unreliable. What you're doing here is that you're using a verification tool that is more designed to assess to what extent are claims made within a document supported by the evidence that underpins it. So this is just an AI-assisted QA step within your QA process. It just highlights — it doesn't say that things are wrong. It just says, "Hey, I can't find the evidence. From the data you've provided me, I can't see where this has come from." So you as the QA person can therefore go, "Oh, mate, where did we get this finding from?" And in most instances, there's going to be a good explanation — "Oh, yeah. I didn't put it in the data," or, "Here it is here in my Excel sheet and the tool missed it," or whatever.
Two tasks: strategic intent and workflow redesign
Anton : Okay. So that's really helpful. Thanks, Adam. Those are two leadership behaviors. But we also had two tasks on our list. The first task is to find your strategic intent and your risk appetite for AI. I know that the audience are really interested in that one because it came in second on the survey. I'm assuming that this doesn't necessarily need to be a massive document or a big complex or complicated undertaking. Nationally, could you just start with ensuring all the leaders have the same answer to the question of, what are we trying to achieve with AI? And I guess the second, the corresponding question with that is, what are we not willing to risk to get there?
Adam : Yeah, exactly. So we typically do this as a workshop — get the leaders together, get the current AI system owners together. And what you're doing here is you're trying to debate and challenge internally to land on that cohesive view of what your strategic intent is. And then what you're also trying to do in that workshop is to find the upper bounds of your risk acceptance. And we have a little tool that we use to do that. It uses real world scenarios, and so leaders can give their own perspectives of where the line is drawn. And then we create an org-wide aggregate of that view. And that then, as you can imagine, cascades down into your policies, your implementation, et cetera.
Anton : Okay. Yeah, fantastic. On the risk side of things — that tool that you just described and the different scenarios — I'm curious about, presumably everything's sequential, right? So you have to have done the intent setting part in the first instance.
Adam : Yes, yeah. And then you're calibrating the risks from there. So the intent, yeah. That's right. And then what you'll find is that out of that process, once you've got your intent and your risk appetite, what drops out of that is a pretty clear indication of what area of your organization should be leading the AI implementation. Because, again, going back to our earlier example, if your intent is more services, then you are likely to land on an implementation that enables frontline staff, et cetera. Therefore, who manages the frontline staff probably becomes a question of do they lead the AI implementation. And what you can see here is that that person doesn't need to be an AI expert. They have the wraparound support of the IT team, et cetera, but they're the ones embodying the strategic intent and the risk appetite.
Anton : Okay. Excellent. So I guess one of the other elements that I was interested in is around the governance side of things and the composition of the AI governance team. In our earlier chat, you talked about four key roles. Someone with legal and privacy knowledge, someone with technical expertise who understands the sort of core work the business does, someone from a business area already using AI, and then the named lead who's accountable. And IT should also form part of that.
Adam : Yeah. Okay. And then the lead who's accountable, I guess, for a number of different things, including what happens when AI goes wrong unexpectedly. And as you said before, it doesn't need to be someone from IT — it can be from the right functional area or in the right role for the organization.
Anton : That's right, yeah. So here's the diagnostic test and task that you can run this week: When an AI output goes wrong in a way that matters, who does your team take it to? What's the actual escalation pathway? Governance that no one can locate in a moment of pressure is not governance, it's paperwork that gives an organization the feeling of safety without any substance.
Adam : Mm. Yep. So the second task is about redesigning workflows. Yeah, sorry — I know we're running behind time, and the marketing team is screaming at us — but just one quick note on the governance thing.
Anton : Sure.
Adam : What's super interesting about AI is you can build your governance into your system. So let me give you an example. AI will confidently express te reo Māori kind of capabilities. But obviously that's dubious at best because of the training data. The point here is that we have a kind of a blocker in our system that basically says, "As a user, you've reached a point in which you've tipped over from adding te reo into your document, into an area where AI is not actually capable of assisting you appropriately." You need to speak to these named individual internal people who will help you.
Anton : Okay. So, I guess in that example, you're talking about the appropriateness. So the appropriateness, and I guess in that cultural setting, really important, right? So a real risk of misappropriating or using incorrectly indigenous concepts and language where you shouldn't be. And again, coming back to the alternative, which is that the human expert is the place to go for that kind of input.
Adam : Exactly. So that's an example of governance built within the system. And there's all sorts of ways that you can, from a governance level, report on how many instances are being blocked. And then you can obviously then target training, whatever it might be. But you can build your governance into the system. And that's something I'd really encourage people to do.
Anton : I like that. The idea of integrating the controls and guardrails, keeping people safe, but also preventing those unwanted risks. So integrating that into the AI system and use. Okay, excellent. We're going to have to go fast now. Good interruption, but time pressure now. We'll see how we go.
Anton : So the second task is about redesigning workflows. And again, this was of prominence in the survey. And admittedly, you picked it. As we heard earlier on, there's friction redesigning workflows — absolutely critical. The research shows that high performers in this space are two to three times more likely to have fundamentally redesigned their workflows. And I think, and you mentioned not simply adding an AI layer to the legacy ones.
Adam : Yep. Mistake that we made as well.
Anton : Yep. So, we also talked about that that's a leadership task, because only leaders can authorize that level of disruption, any investment that's required, and reprioritization. That's right. And so how do you do it? What task are we giving people out of this?
Adam : Yeah. So the task is — on your screen is our kind of method to prioritize your workflows. What you're looking for here is tasks that are repetitive, low risk, judgment light, with minimal change and sovereignty requirements. So when I say tasks, if you think of your workflow as a series of kind of headings — so, in an evaluation sense, it's data collection, data analysis, et cetera — what you're looking for here is, okay, under data collection is a series of tasks. Some of those will be suited for AI assistance and some of those will not. And so it's kind of prioritizing those tasks within the workflow using this model.
Anton : Okay. Brilliant. And so, once you've mapped the tasks, the redesign follows a three-stage logic. You should be able to see it on your screen. Again, we're kind of whipping through this pretty quick, but you can download it afterwards. So, the point is that it's actually the embed step where you have your AI team work with the team currently doing the work to co-develop that new AI assisted workflow. But you can only really do the embed stage once everyone has a basic understanding of AI systems use, et cetera, which is the equip stage in our framework.
Adam : Okay. Yeah. Brilliant. The co-develop piece feels to me to be — I mean, from an operating model design perspective, co-design, co-development — you want to have business units, the actual users, in the mix, because they're the main experts, right? Well, obviously I lead the AI team here. What do I know about evaluations, right? We've got veterans who know everything about evaluations. Their input into that process is absolutely critical, because if you ask the AI team, "Oh, can we create this workflow or AI assisted workflow?" They'll say, "Oh, we can automate the whole thing." And it's like, hmm. Can isn't should, and that doesn't mean that it'll be done well.
Anton : Yep. Okay, fantastic. So, we've experienced that things don't always go smoothly. We ran this process ourselves. We ran a few projects, actually, where we had the human-only activity running in a parallel stream with the AI assisted process. Took some lessons from that, and that again emphasizing the importance of co-development.
Anton : Okay. We promised everybody something that you could act on by Monday. Not a framework to implement, just three specific things. Commitment one: name a single accountable person for AI governance in your team or function. Not a committee, but a person. And if that person already exists in your structure, write their name on a slide and put it in front of your leadership team this week. If they don't exist yet, you've got your first decision to make, and hold Adam's point in mind when you make it. The question of who is accountable and the question of what value are you looking to create — ideally, you're not making those with a massive gap between them. They should flow one from the other.
Adam : Absolutely, yeah. That's right.
Anton : Okay. Commitment two: use AI visibly in at least one piece of work this week, and then tell your team about it. Share the experience. Adam, what's your advice for leaders here?
Adam : Yeah. Obviously, board of advisors is a great place to start. So give that a crack if you fancy it. I think an alternative is produce a report or draft something using it, but that makes sense and it doesn't have to be a big activity, but do something where you're demonstrating to others that you've used AI.
Anton : That's right, yeah. Okay. And then commitment three: pick one workflow — not a tool — and then start the redesign conversation. So nominate a workflow, and then get together with the team that work on that flow and walk it through with them. Not how a process map says it happened, but follow those people through what they actually do in that process flow. Ask pivotal questions along the way. Where does it slow down? Where is effort duplicated? Which tasks are repetitive, slow, or — and importantly — cognitively light? And that's where the AI question should start. So tasks first, and then the tool second.
Adam : That's right. Yep.
Anton : So three commitments: one person named, one piece of visible AI use with an honest debrief to the team afterwards, and one workflow conversation with the people who actually do the work.
Anton : The through line of everything that we've talked about today: the gap between what AI can do and what your organization is getting from it isn't a technology problem. The evidence is clear — it's a leadership problem. The frictions that sustain it are upstream from your staff, and that means the job of addressing and closing the gap is yours as leaders.
Anton : The substance of session one — that's the first in our new leadership series, "The Quiet Lifting." Thanks for staying with us through the harder parts. We're excited about the upcoming episodes that we've got in the series. We have some gnarly topics, such as: leading across boundary systems in a whole of government world; doing the right thing in public view — ethical decision-making and behavior; trust, the key to the authorizing environment and social license; and timeless wisdom — Te Ao Māori and indigenous leadership in a Western system. We'll also be joined by some amazing senior leaders from across the public and private sectors from Australia, Aotearoa, and the Pacific, so keep your eye out for an email and details of episode two.
Adam : Mm-hmm.
Anton : If you want to continue the conversation before then, find us at allenandclarke.com. We're pretty easy to find.
Adam : And easy to talk to.
Anton : That's the intent. Before we jump onto questions, Adam, any final thought?
Adam : Yeah. I think you've covered it beautifully, mate. Sorry, I've chewed up quite a lot of time. So, do we have time for questions? Okay. So what we'll do is I think we'll stay around. And we'll actually go through and we'll answer all the questions, so that the video contains the questions. Because we'd really appreciate people submitting, taking the time to submit questions and challenges. And if we haven't answered them directly in the content, then we can address those now.
Anton : No, I think it's good. People want to hear from the experts, so that's fantastic. Look, everybody, thank you for joining us today. We appreciate it. If you want to hang around for questions, please do so.
Adam : Tēnā koutou katoa.
Anton : Let's go to questions.
Anton : Okay. Fantastic. So a question from Svetlana. "Workflow redesign feels like the primary task. The longer it takes, the wider the cultural and process divide grows within and between organizations. What's your take on centralized versus decentralized models for AI adoption?" That's a cracking question.
Adam : That is a fabulous question. So, from my experience, to achieve the kind of speed that you're alluding to, Svetlana, decentralized is really the only way to go. And as I mentioned earlier in the session, about your governance being built into the system — that is the enabler for your decentralized mode. And what I would say is we have a mechanism by which, once a team or team member creates a process for themselves or within a small group, we have a system by which they then let the AI team know, and the AI team come along and say, "Hey, this is awesome. You've done a brilliant job here. Let us take this and systematize it." You can then test it because you've built version one. We're building version 1.1. And you then as the group test it, use it, we tweak it, and maybe we get to a version two. The version two gets rolled out across the whole organization. And then what you've done in that process is you're getting this ground-up bunch of tools, systems, et cetera. You still have the AI team, as it were, managing the overarching architecture of your setup, but you get so much value from those individuals that it's impossible to ignore. And the thing I would just add to that is your AI team don't know enough about the frontline work. So, for me, waiting for the AI team to come along and hand you something is just not at all going to achieve the speed that you need.
Anton : Yeah. Okay, fantastic. We also have a question from Pavana: "When dealing with large volumes of data, are there suggested ways to check for reliability and accuracy?" This is your domain, mate.
Adam : Thank you for this question. This is a real challenge for everybody, including ourselves. So I'll try to address it very quickly. What happens with your AI tool when you put too much context into it is you get what's called a "lost in the middle" effect. So AI kind of forgets the stuff in the middle of your dataset. So if this is your dataset, then the middle obviously is shallow. But as you extend your dataset out, the middle gets bigger and bigger and bigger. And obviously that causes problems with accuracy, et cetera.
Adam : And the other thing — so if you're familiar with some tools like NotebookLM. So NotebookLM, you can load tons of stuff in, and it feels amazing at how it can kind of accurately recall all the text. That's using a system called RAG. And what that actually is doing in the background is searching — kind of like a Google search — of the content that you've given, pulling out chunks from the text, processing it through an AI system, and then giving you an answer. What is hidden within that process is that the system, when it searches, obviously, one, you have no control over the search criteria. Two, you have no control over to what extent it is drawing the line between relevant and irrelevant content. And so you're still getting an averaging of your data.
Adam : The best solution — kind of the only solution at the moment — is to chunk your data into smaller pieces. So we use some tools to do that ourselves. But you can do this for your own kind of AI system. So if you think about an evaluation report, for example, right? Generally, your report will go, "KEQ1, here's kind of my findings," et cetera, in there. What you want to do is you want to go, okay, what data is irrelevant to KEQ1? Exclude that from your dataset and then only include the relevant data, and that's your chunk for KEQ1. And then again, if you think about this as all your data, you might chop half of it out because you know definitely half of it is irrelevant.
Anton : Okay. All right. So putting some criteria about what in the big dataset, what data we actually want to interrogate or point AI towards.
Adam : Yeah, so it's kind of like the reverse of how people think about it, because historically you would find the data that is relevant. In this instance, you're going the other way around. You're saying what data is definitely not relevant, and then everything that's left is what you're using.
Anton : Nice clarification. Okay, fantastic. We've got another question here, this time from Alex: "There's a low appetite for risk among senior leaders to be innovative. How would you navigate this?"
Adam : It is super tough. I don't have an immediate answer. Do you have any views on this?
Anton : Well, I think a couple of thoughts that I think are worth throwing into the mix. So from what I liked about the frictions and the commentary you put around the frictions earlier on is that there's a relationship between risk and the value that you're trying to extract. So I think in any of the conversations where you're trying to influence those senior leaders, we need to be coming back to the value part of the equation. And I think sometimes we talk in too general a sense when we're talking about that value. But as soon as you start specifying and, where it's possible, quantifying what the opportunity is — and the opportunity cost that that risk is walking away from or preventing from being realized — I think that's a more provocative type of conversation to possibly be having with senior leaders.
Anton : I also wonder whether there's an opportunity sometimes to bring into the mix a validation process around the risk. So once we start picking it and asking people to be specific about what the risk is that's causing that nervousness, we can then have a conversation around some of the mitigations that are available — or validate, in fact, whether the risk is real versus a perceived risk. I do feel that it is a real challenge, though, and I think that we have a number of organizations struggling and grappling with it, and those permission settings are key. So the conversations with senior leaders in the organization are the ones to keep trying to influence, I think, and those are some of the techniques possibly that we could be using.
Anton : I think the other opportunity that occurs to me is that sometimes there are the exemplars out there, right? There are the other organizations in the same sort of space who are doing things and doing it really well. And so creating opportunities to sort of showcase what others are doing and create that understanding about how they're managing risks and what upside they're delivering as a result — those are some of the opportunities to explore.
Adam : Yeah, I like your framing around the risk thing, because we see this in our conversations with our clients quite often. At the moment, while we're kind of in this transition phase, clients will say to us, "Oh, I don't want you to use AI on this project," or whatever. And then the kind of follow-up question is, okay, what is it that you're concerned about? And what the answer often is, is our policy or whatever it might be is that we don't put our data into an AI system. But you can still use AI and still meet that requirement. Because not all AI use requires that specific, in our case, client's data. So you can use synthetic data — which is AI-generated data, which mimics the real stuff. You can use placeholder data — like if you're building a dashboard or something like that, you can use placeholder data, use AI to build the dashboard, then disconnect the AI from the system, and then put the real data in. Hey presto, your dashboard's built. So there's all these kind of nuances. So that kind of goes to the second question. I'm not sure that was that helpful. Sorry, Alex. It's a tough one.
Anton : Yeah, it is tough. So last question, this one is from Mark: "How do I hold senior leaders to account when they don't understand the capability and possibility AI presents?" This is, in some ways, a little bit similar to the question from Alex — a little bit of overlap, not entirely, but some. Do you have any thoughts about this one?
Adam : I do, yeah. So this is something that we experience when we're talking to clients. And that's kind of why I go back to that Anthropic chart quite often. Because that Anthropic chart is just such a great visualization of what the gap is, and then you move the conversation away — before I move on, so there's that gap, then there is so much data out there, market research papers, et cetera, that shows the trajectory of AI capability, and nobody will be shocked, number go up. And so it's relatively easy for us, in our experience, to paint a picture that this isn't going away. The systems are extremely capable. And we like to show some live uses of our system as well, to be like, here's how our system will do a task that would align to them. And that gets you to the realization of how capable the systems are, and then you're immediately moving into the next conversation, which is natural — that people then go to, oh, but, and here are the reasons why we can't do it, or we can't capture that value. And now you've moved the conversation on from a "I don't believe that the capability is there or the possibility is there." You're talking about the frictions that are blocking the organization from unlocking that capability, and the frictions have solutions to them.
Anton : Yeah. Okay, fantastic. As you were explaining that one, one other thing that just occurred to me is that through any strategic planning process, right, there's a contextual analysis piece. We look at the environment. There's normally an internally-focused sort of contextual analysis, but also an externally-focused one, and people use a range of different techniques to do that context setting. Whether it's using a PESTEL-type frame to have a conversation and do the analysis, whether you're engaging in some sort of scenario concepts and planning. The point is, though, that periodically we're looking at what's going on in the environment, and then trying to understand what the impact of that might be on our organization, so that we can calibrate what we're doing and what we're planning to do appropriately. There's not many organizations I wouldn't have thought where, if you're doing that contextual analysis, AI doesn't feature in some significant way.
Adam : Absolutely, yeah. And I wonder whether that's also an opportunity for Alex and Mark in their conversations to sort of quietly suggest — or looking at the next planning cycle — to include AI, to get that included into the agenda
Anton : Okay. Well, that's probably a good place to wrap up. Thanks everyone for staying on and for all your thoughtful questions. They were fantastic. I know there are a few others that we've received, and we haven't quite got to them. But we will do our best to address those and come back to you. But this is the kind of conversation that makes these conversations, I think, valuable for everyone. So remember, if any of today's discussion has sparked some ideas for your own work or some more questions, we're always happy to continue the conversation. Thanks again for joining us. We really appreciate it.
Adam : Thanks, everyone. Have a great rest of your day. Ka kite anō.
Anton : Ka kite. See you around.
