Skip to main content

When AI Models Change Underneath You

AI models update on schedules you don't control. Building reliable systems on shifting foundations is the new normal, and it requires a different mindset.

Tim Clark

Tim Clark

Co-founder · 25 March 2026 · 4 min read

When AI Models Change Underneath You When AI Models Change Underneath You

There’s something about AI that’s different from other technology we’ve worked with, and I’ve been turning it over in my mind for a while now.

With traditional software, when something changes, you know about it. There’s a release, a changelog, maybe a migration guide if it’s significant. Someone on your team made a decision to update, and if the update breaks something, you can trace back to that moment and figure out what happened.

AI doesn’t work like that. In 2024 alone, OpenAI released over a dozen model updates, Google pushed five major Gemini versions, and Anthropic shipped three Claude generations. Each one subtly changed how existing integrations behaved.

The models we build on top of are maintained by someone else entirely. They get updated on schedules we don’t control, for reasons that aren’t always explained, and the behaviour shifts in ways that might be subtle enough that you don’t notice immediately. Your prompt that worked beautifully for six months starts producing slightly different outputs. Your automation that handled a particular task reliably now misses edge cases it used to catch.

Nothing changed on your end. The foundation just moved a bit.

Your prompt that worked beautifully for six months starts producing slightly different outputs. Nothing changed on your end. The foundation just moved.

Building on shifting ground

I find this fascinating rather than alarming, because it means we need to think about AI systems differently from how we’ve thought about software before. It’s less like building on solid ground and more like building on something that’s slowly shifting. Not unstable exactly, but not static either.

The Gemini models have been an interesting example of this. Google iterates fast, new versions coming out regularly, each one slightly different from the last. If you built something tuned to the behaviour of one version, the next version might handle things differently. Usually better overall, but “better overall” doesn’t always mean “better for your specific use case.”

We’ve started thinking about this in terms of baselines. Before putting something into production, getting really clear on what good looks like. What outputs do we expect? What’s the acceptable range of variation? How would we know if something had drifted outside that range? It’s the kind of discipline that feels obvious once you’re doing it, but it took us a while to realise it was necessary.

Nobody else is watching this for you

The vendors aren’t going to tell you this stuff, not because they’re hiding anything but because they’re focused on capability. They want you excited about what the new version can do, not thinking about how it might affect what you’ve already built. That’s not unreasonable from their perspective. They’re building general-purpose tools for millions of users. Your specific implementation isn’t their concern.

Which means it has to be yours.

Output drift over model versions — baselines help you notice when things shift

I’ve come to see this as just part of working with AI at this stage. The technology is moving fast, the models are improving constantly, and that improvement comes with change. Fighting that seems pointless. Better to build with the expectation that things will shift and have ways of noticing when they do.

A relationship, not an installation

It’s actually quite an interesting problem when you frame it that way. How do you build reliable systems on top of foundations that evolve? How do you maintain quality when the underlying capabilities keep changing? These aren’t solved problems. We’re all figuring it out as we go.

The people who’ll do well with AI in the long run, I suspect, are the ones who get comfortable with that. Not treating AI as a fixed tool that you configure once and forget about, but as something that needs ongoing attention. A relationship more than an installation.

Not treating AI as a fixed tool that you configure once and forget about, but as something that needs ongoing attention. A relationship more than an installation.

That might sound like more work, and it is. But it’s also more interesting. The tools keep getting better, which means the possibilities keep expanding. You just have to stay engaged with it.

Want to build AI systems that stay reliable as models evolve? Our AI Clarity Session helps you establish the baselines and monitoring that keep your AI working as intended. Or explore our real-world AI use cases to see how businesses are building with change in mind.

Related reading

Use Case

Smarter Meeting Summaries

Use Case

Customer Support Ticket Intelligence

Use Case

AI Capability Assessment & Training Measurement
Getting Your Data House in Order for AI Getting Your Data House in Order for AI

Getting Your Data House in Order for AI

AI is only as good as what you feed it. Research shows poor data quality costs businesses 15-25% of revenue. Most aren't ready for the AI projects they want to build.

Tim Clark

Tim Clark

Co-founder · 21 March 2026 · 4 min read

There’s a conversation I find myself having quite regularly with businesses who want to do something interesting with AI. They come in with ideas, often good ones, about what they want to build or automate or improve. And fairly quickly we end up talking about their data.

Not because I want to. It’s usually not the exciting part of the conversation. But it turns out to be essential.

AI is only as good as what you feed it. Everyone knows this in theory, but the practical implications take a while to sink in. Gartner estimates that poor data quality costs organisations an average of $12.9 million per year, and that’s before you try to build AI on top of it. If your data is scattered across systems that don’t talk to each other, if it’s inconsistent or incomplete or just not captured in the first place, then the AI doesn’t have much to work with. You can have the most sophisticated model in the world and it’s still going to struggle if the information it needs doesn’t exist or can’t be accessed.

The silo problem

The most common thing we find when we start looking at a business’s data is silos. Information sitting in different systems that were never designed to connect. The CRM knows one thing. The finance system knows another. There’s a spreadsheet somewhere that holds crucial information, except only one person knows it exists and they’re on holiday. Everyone has a partial picture and nobody has the whole thing.

Most businesses don’t actually know what data they have. They know the obvious stuff. We’ve got a CRM, we’ve got an accounting system, we’ve got customer records. But they don’t necessarily know what’s in those systems in any detail, or how it relates to what’s in other systems, or what gaps exist. And there’s often data they didn’t even realise could be useful sitting in places they never thought to look.

From disconnected silos to connected intelligence

This isn’t anyone’s fault. Businesses grow organically. Systems get added over time to solve specific problems. Nobody sat down and designed a unified data architecture because that wasn’t the priority at the time. The priority was solving whatever problem was in front of them, and they did that, and now they’re dealing with the consequences.

The temptation to skip

The temptation when starting an AI project is to skip past all this. Just get to the interesting bit. Build the thing we actually want. The data cleanup feels like a detour, a chore to get through before the real work starts.

The data cleanup feels like a detour, a chore to get through before the real work starts. But the data work is the real work.

But the data work is the real work, or at least a critical part of it. Getting clear on what information exists, where it lives, how it connects, what’s missing. That foundation determines what’s possible later. Try to skip it and you end up building on sand.

I’ve noticed that businesses don’t usually resist this step once they understand why it matters. They’re not opposed to doing the groundwork. They just want to skip it because it doesn’t feel like progress. It feels like preparation, and preparation is less exciting than building.

When good data clicks

The shift happens when they start to see what good data makes possible. When you can actually connect information across systems and start seeing patterns that were invisible before. When the AI has enough to work with that it starts surfacing genuinely useful insights rather than just reflecting your gaps back at you.

That moment is worth the effort of getting there. And once the foundation is in place, everything that comes after is easier. Each new AI project builds on what you’ve already organised rather than having to start from scratch every time.

Each new AI project builds on what you’ve already organised rather than having to start from scratch every time.

The boring work enables the interesting work. That’s true in most fields, and AI is no exception.

Not sure if your data is ready for AI? Our AI Clarity Session helps you map what you have, identify the gaps, and build a foundation that makes everything else possible. Or explore our real-world AI use cases to see what good data foundations enable.

AI Adoption: Meeting People Where They Are AI Adoption: Meeting People Where They Are

AI Adoption: Meeting People Where They Are

AI adoption isn't a technology problem. It's a people problem. 70% of digital transformations fail due to people issues, not tech. Successful businesses bring their team along.

Tim Clark

Tim Clark

Co-founder · 18 March 2026 · 4 min read

When businesses think about AI, they tend to think about the future state. What things will look like when it’s all working. The efficiency gains, the time saved, the insights revealed. It’s natural to focus there because that’s where the value is.

The bit that gets less attention is the journey from where people are now to where you want them to be. And that’s usually where things get complicated. McKinsey research consistently shows that around 70% of digital transformation efforts fail, and the primary reason is people, not technology.

I’ve watched projects that had everything going for them on paper struggle to get traction. Good technology, clear use case, leadership support, budget. All the ingredients for success. But the people who were supposed to use the thing just… didn’t. Or did reluctantly, badly, and then stopped.

Understanding the starting point

The people who were supposed to use the thing just… didn’t. Or did reluctantly, badly, and then stopped.

Usually when this happens it’s because nobody spent enough time understanding the starting point. What are people actually doing today? How do they think about their work? What are they worried about? What have their experiences been with previous technology changes?

That last one matters more than you might think. A lot of people have been burned before. New systems that made their jobs harder, not easier. Rollouts that promised efficiency but delivered frustration. Changes that felt like they were being done to them rather than with them.

So when AI comes along, there’s wariness. Another transformation. Another promise that things will be better. They’ve heard this before.

Sceptics become advocates

The interesting thing is that the people who are most cautious about new technology are often the most process-driven and detail-oriented people in the organisation. They’re not resistant because they’re stuck in their ways. They’re resistant because they’ve seen what happens when change is handled badly and they don’t want to go through it again.

Once you understand that, you can work with it rather than against it. Those same people, the ones who ask lots of questions and raise concerns and want to understand exactly how this is going to work, often become the strongest advocates once they’re on board. They bring rigour to the implementation. They think about edge cases that nobody else considered. They make sure things actually work properly rather than just looking good in a demo.

The path to adoption is never a straight line

Getting there takes patience. It means explaining not just what the AI does but why it helps with the specific things they care about. It means showing rather than telling, letting people see it work in their context rather than generic examples. It means creating space for scepticism rather than dismissing it.

A human pace

AI needs to understand how work actually happens before it can improve it. And that understanding has to include the people doing the work. Their knowledge, their concerns, their context. Skip that and you end up with technically impressive solutions that nobody uses.

The businesses that handle this well don’t just deploy AI. They bring people along. They treat adoption as something that happens at a human pace, not a technology pace. They understand that the dream state is only valuable if you can actually get there, and getting there requires meeting people where they are.

The dream state is only valuable if you can actually get there. And getting there requires meeting people where they are.

It takes longer. It’s messier than a clean technology rollout. But it’s the difference between a tool that gets used and a tool that gathers dust.

Want to bring your team along on the AI journey? Our AI Clarity Session starts with understanding where your people are now, not just where you want them to be. Or explore our real-world AI use cases to see how other businesses have navigated the human side of AI adoption.

When AI Surfaces Data You Didn't Want to See When AI Surfaces Data You Didn't Want to See

When AI Surfaces Data You Didn't Want to See

The most valuable AI insights are the ones that challenge you. Data that confirms what you believed is reassuring but doesn't change anything. Discomfort is a sign of progress.

Tim Clark

Tim Clark

Co-founder · 14 March 2026 · 4 min read

There’s a moment that happens sometimes when you surface data clearly for the first time. The client goes quiet. You can see them processing something.

What they’re looking at isn’t necessarily complicated. It might be something simple. Revenue they’re leaving on the table. Inefficiencies they knew existed but hadn’t quantified. Patterns that were invisible when the information was scattered but obvious once it’s brought together.

Research from MIT Sloan found that while 87% of leaders believe AI will give them a competitive advantage, far fewer are prepared for the organisational discomfort that comes with actually using data-driven insights to change how they operate.

The silence is interesting. It’s not confusion. They understand what they’re seeing. It’s more like the feeling when someone points out something you’d managed to not look at directly.

The reactions

People react differently to this. Some feel frustrated with themselves, like they should have caught it earlier, should have done something about it. There can be embarrassment, especially if others are in the room. A sense that this reflects badly on their management or judgment.

That’s a natural reaction, but it’s worth moving past quickly. Because the moment you can see something clearly is actually a good moment, even if what you’re seeing isn’t pretty. Before, you were operating without full information. Now you have it. That’s an improvement, not a failure.

The moment you can see something clearly is actually a good moment, even if what you’re seeing isn’t pretty. Before, you were operating without full information. Now you have it.

The harder reaction to work with is avoidance. When someone sees something uncomfortable and their instinct is to step back from it. To qualify it, to question the data, to find reasons why it might not be accurate or might not matter as much as it seems to.

That’s understandable too. People are busy. They already have more on their plate than they can handle. Seeing a problem clearly means having to do something about it, and where’s the time for that going to come from?

Where AI helps

This is where AI actually helps, though it takes a bit of faith to see it in the moment. Once a problem is defined clearly, once you can state what’s wrong and what better would look like, AI can take a lot of the load. It can monitor, analyse, suggest, handle the repetitive cognitive work that would otherwise eat up hours. The discomfort of seeing the problem is temporary. The benefit of being able to address it systematically is ongoing.

I’ve started to think of uncomfortable insights as a sign that something valuable is happening. Data that just confirms what you already believed isn’t particularly useful. It’s reassuring, maybe, but it doesn’t change anything. The insights that matter are the ones that challenge you. That show you something you didn’t know, or didn’t want to know, or knew but hadn’t confronted directly.

From noise to clarity — uncomfortable insights are the starting point for real improvement

A learnable skill

Those insights are a gift, even when they don’t feel like one. They’re the starting point for actual improvement rather than just optimisation around the margins.

The businesses that do well with data aren’t the ones who never see uncomfortable things. They’re the ones who can sit with that discomfort, acknowledge what they’re looking at, and then figure out what to do about it. That’s a skill, and like any skill, it gets easier with practice.

The first uncomfortable insight is the hardest. After that, you start to welcome them. Because you know now that seeing clearly is always better than not seeing, even when what you see isn’t what you hoped for.

Ready to see what your data is actually telling you? Our AI Clarity Session surfaces the insights that matter, including the ones that challenge your assumptions. Or explore our real-world AI use cases to see how other businesses turned uncomfortable truths into real improvements.

Governance as Confidence (Not Constraint) Governance as Confidence (Not Constraint)

Governance as Confidence (Not Constraint)

Good AI governance doesn't slow you down. It accelerates you. Risk-based frameworks let teams move fast where it's safe and apply controls where it matters.

Tim Clark

Tim Clark

Co-founder · 27 February 2026 · 5 min read

When I mention “governance,” most people immediately think: bureaucracy. Slow. Red tape. “We can’t do this without filling out Form 47 and waiting for committee approval.”

That’s governance done wrong.

Good governance doesn’t slow you down. It accelerates you.

Let me explain.

The Governance Problem (Done Wrong)

Bad governance looks like: “No one can use AI without executive sign-off. All AI decisions go through a committee. All prompts must be reviewed. All outputs must be audited.”

This stops experimentation. It kills momentum. It makes your organisation move like a tank through mud.

So teams bypass it. They use ChatGPT in ways you don’t know about. They build AI systems outside your frameworks. They hide things because the official process is too slow.

That’s not safety. That’s the opposite. That’s fragmentation and risk. Deloitte’s research highlights exactly this problem. Unsanctioned AI deployments by individual teams create governance blind spots that are far more dangerous than the risks governance was meant to prevent.

And the numbers back this up: while 75% of organisations have established AI usage policies, only 36% have adopted a formal governance framework, according to the 2025 AI Governance Benchmark Report. Having a policy without a framework means no consistent roles, controls, monitoring, or enforcement. It’s governance theatre.

The Governance Insight

Here’s what I’ve noticed in organisations that are winning:

They don’t have less governance. They have smarter governance.

They ask: What’s actually risky? And what’s not?

Using ChatGPT to brainstorm marketing copy? Low risk. Let the team do it. No governance needed.

Using an AI model to flag fraudulent transactions? High risk. That needs governance. That needs an audit trail, oversight, validation.

Storing customer data in a third-party AI system? High risk. Governance required.

Processing internal meeting notes? Low risk.

The organisations that move fast don’t have no governance. They have risk-based governance. They apply strict controls where it matters and trust teams where it doesn’t.

What Risk-Based Governance Looks Like

Instead of one process for everything, you have levels:

Green (Low Risk): Team uses AI tools, minimal controls needed. Completion timestamp. That’s it.

Yellow (Medium Risk): Team uses AI, but there’s documentation of the prompt, the output is reviewed by a second person, there’s an audit trail. Takes an extra 30 minutes per project.

Red (High Risk): Governance committee review. Full audit trail. Compliance checklist. Formal sign-off.

What determines the level? Risk factors:

  • Is personal data involved?
  • Is this a legal or regulatory decision?
  • Could bad output cause real harm?
  • Are we trusting AI to make a decision, or just to inform one?
  • How transparent is the AI’s reasoning?

Most projects are green or yellow. A few are red.

And here’s the key: Teams know upfront which level they’re in. They’re not surprised halfway through. They know what controls they need and plan accordingly.

This is why the AI Native Platform includes a built-in governance framework — so your team can classify projects by risk level and apply the right controls without slowing down the low-risk work. Gartner’s research found that organisations with AI governance platforms are 3.4 times more likely to achieve high effectiveness in AI governance than those without.

Risk-based governance matrix showing low-risk tasks move fast while high-risk decisions get proper oversight

How This Enables Speed

Let me give you an example.

Tender analysis: Is it high risk?

Some organisations say yes: “We’re using AI to evaluate proposals. That’s a decision that could be wrong. Governance required.”

Other organisations think differently: “The AI reads the tender, extracts information, and makes a recommendation. A human reads the AI output and makes the decision. The human is responsible. The AI is a tool.”

Different risk assessment. Yellow governance, not red.

With yellow governance, you can:

  • Launch projects in 4 weeks, not 12 weeks
  • Scale across teams, not wait for approval
  • Iterate and improve, not freeze specs
  • Trust teams, not create bottlenecks

And you’re still safe. You still have documentation. You still have an audit trail. You still have review. You just don’t have bureaucracy.

The data supports this approach. The Cloud Security Alliance found that organisations with comprehensive AI governance are nearly twice as likely to successfully adopt advanced AI, with 46% adoption versus just 25% for those with only partial guidelines. Good governance doesn’t slow adoption. It enables it.

The Confidence Part

Here’s why this matters for your board and executives:

Bad governance makes it seem like you’re being careful, but you’re creating risk.

Good governance means your leadership can move fast with confidence.

“We evaluated tenders with AI support. The AI recommendation was reviewed by our procurement lead. Here’s the audit trail of the AI analysis. Here’s the decision. Here’s the reasoning.”

That’s auditable. That’s defensible. That’s confidence.

Compare to: “We used ChatGPT to help. I think it worked well. Here’s the outcome.”

That second one is risky. Not because you used AI. Because you can’t explain your process.

This is where structure matters, and it’s one of the reasons the AI Native Programme includes governance setup as part of the monthly partnership. You don’t need to build a framework from scratch. You need one that fits your risk profile and enables your team to move.

What This Means for Your Organisation

If you’re worried about AI risk, good governance is the answer. Not “no AI.” Not “secret AI.” Structured, risk-based governance.

It lets you move confidently. Your team knows what’s expected. Your executives know what’s auditable. Your customers (if relevant) know you’re being responsible.

It’s not bureaucracy. It’s confidence.

And it’s not “no governance vs. full governance.” It’s “smart governance that enables speed where it’s safe and controls where it matters.”

That’s how you actually win with AI.

Keen to talk about what a governance framework might look like for your organisation? Get in touch. It’s usually a shorter conversation than you’d expect.

Knowledge Capture Matters More Than Automation Knowledge Capture Matters More Than Automation

Knowledge Capture Matters More Than Automation

Your domain experts' frameworks, decision patterns, and judgment calls are more valuable than any automation project. Here's how to capture them before they walk out the door.

Tim Clark

Tim Clark

Co-founder · 27 February 2026 · 6 min read

You have a domain expert. Maybe it’s your emergency response director who’s been in the role for 15 years. Or your most senior lawyer who knows every nuance of contract interpretation. Or your procurement manager who understands supplier relationships across your entire network.

This person has knowledge that’s critical to your organisation. And almost all of it lives in their head.

They’re not leaving tomorrow. But they will eventually. And when they do, 15 years of pattern recognition, judgment calls, and contextual understanding just… walks out the door.

Most organisations think this is inevitable. “That’s why we need to hire and train replacements.”

But there’s something you haven’t considered yet: What if you could capture that knowledge before they leave?

Not all of it. You can’t download 15 years of intuition. But the frameworks? The decision-making patterns? The rules of thumb? The common mistakes to avoid? The client relationships and history that inform decisions?

That’s capturable. And it’s way more valuable than most automation.

Why This Matters More Than You Probably Think

Let me give you a real example. A government agency needed to evaluate tenders. The evaluation was complex: technical capability, cost, track record, financial stability, cultural fit. The decision required judgment.

There was one person, call her Sarah, who was brilliant at it. She’d been doing it for 8 years. She understood the industry, knew the vendors, had a nose for where corners were being cut.

When Sarah evaluated a tender, you got 95% accuracy. People trusted her judgment.

The organisation started thinking: “Can we automate this? Can AI learn Sarah’s rules?”

They built models. They tuned prompts. They got to maybe 70% accuracy. It was helpful, but it wasn’t Sarah.

Then they tried something different. Instead of trying to replace Sarah’s judgment, they tried to distribute it.

They recorded Sarah evaluating tenders (she talked through her reasoning). They extracted the frameworks: What questions matter? What red flags indicate problems? What indicators suggest the vendor is financially stable? What does good cultural fit look like?

They built training material from this. Not a checklist. A framework for thinking.

Then they trained other people using Sarah’s frameworks. Within 3 months, three other people could evaluate tenders at 85% accuracy.

When Sarah eventually retired, the knowledge didn’t walk out the door. It stayed in the organisation. Distributed across three people.

The automation was useful. But the knowledge capture was transformational.

Research from Deloitte found that 42% of enterprise skills and knowledge exist only in individual employees’ minds, with no documentation or formal transfer mechanism. When these people leave, organisations lose an estimated $31.5 billion annually in knowledge-related losses, according to Panopto’s Workplace Knowledge and Productivity Report.

Why Organisations Miss This

Most organisations are obsessed with automation. “How can we do this with less human effort?”

The automation question is valuable. But it’s often the wrong first question.

The first question should be: “What does excellence look like? Who embodies it? How do they think?”

Once you understand that, then you can decide: Do we automate? Do we distribute the knowledge? Do we do both?

Organisations that skip the first question end up with automation that’s technically clever but doesn’t capture the reasoning that made the original process work.

Knowledge flowing from one expert to multiple team members through structured frameworks

What Knowledge Capture Actually Looks Like

It doesn’t mean recording someone for 40 hours and transcribing it. That’s not scalable.

It means: In the context of your work, as you’re doing it, capture the frameworks and reasoning.

  • Voice sessions where your expert talks through a complex decision. The platform captures this, extracts the frameworks.
  • Structured interviews that get at the reasoning behind decisions: “When you saw X in the tender, why did that matter?”
  • Documentation that happens alongside the work, not after.

This doesn’t take extra time. It happens during the work.

For emergency response: If your director handles a critical incident and talks through the decision-making in real-time, that’s knowledge capture. The platform listens, extracts the frameworks.

For contract review: If your lawyer walks through a contract and narrates their thinking (which they’re probably doing anyway), capture it. Extract the frameworks and concerns.

For procurement: If your procurement manager talks through vendor relationships and red flags, capture it. Extract the decision-making criteria.

All of this is happening already. You’re just usually not capturing it. It’s also why Expert Knowledge Capture is one of our most popular use cases. Organisations realise the value as soon as they see a framework extracted from a single voice session.

The Real Value

Here’s what knowledge capture enables:

Onboarding speed: New team members learn from your expert’s frameworks in weeks, not years. This pairs naturally with New Employee Onboarding Automation for a complete onboarding solution.

Consistency: You get consistent decision-making based on proven patterns, not individual variation.

Scalability: Your expert doesn’t have to do everything. They’ve distributed their frameworks.

Resilience: When people leave, the knowledge stays.

Automation: Once you understand the frameworks, you can automate the routine applications. You’re not replacing judgment. You’re automating the judgments that are already clear.

Confidence: Your leadership knows how decisions are made. You have an audit trail of the thinking.

How the Platform Helps

The AI Native Platform has Voice Sessions specifically for this.

You’re having a critical conversation: leading a decision, handling a complex situation, working through a thorny problem. You record it. The platform listens. It extracts:

  • The frameworks you used to think through the problem
  • The decision criteria you applied
  • The insights and nuances
  • The next steps

You get a structured summary of your own thinking. You can share this with your team. You can use it to train others. You can refer to it when similar situations come up.

This is happening automatically. Not slowing you down. Just capturing the thinking that’s already happening.

The result: Your knowledge becomes scalable. Your team can do more. Your experts become multipliers, not bottlenecks.

What This Means for Your Organisation

Think about the critical knowledge that currently lives in specific people’s heads.

For a services firm: Client relationships, vendor capabilities, past project lessons, proposal quality benchmarks.

For government: Policy interpretation, precedent, stakeholder relationships, evaluation frameworks.

For emergency response: Incident decision trees, resource allocation logic, risk assessment frameworks.

For any organisation: Your domain experts’ pattern recognition.

If you could capture that before they leave, what would that be worth?

Not in automation dollars. In resilience, consistency, scalability.

That’s often more valuable than any individual automation project.

Keen to talk about where the critical knowledge lives in your organisation? Get in touch. That’s usually the first place to focus.

Why 95% of AI Projects Fail (And It's Not the Technology) Why 95% of AI Projects Fail (And It's Not the Technology)

Why 95% of AI Projects Fail (And It's Not the Technology)

After working with dozens of teams on AI adoption, the pattern is clear: the technology works fine. The failure isn't technical. It's organisational discipline.

Tim Clark

Tim Clark

Co-founder · 27 February 2026 · 5 min read

You could hand the world’s best AI tools to your organisation and you’d still fail 95% of the time.

I know that’s a bold claim. But after working with dozens of teams on AI adoption, I’ve watched the pattern repeat: The technology works fine. The failure isn’t technical.

Rand Group research puts the figure at 87% of AI projects never making it to production. Gartner estimates that through 2025, at least 30% of generative AI projects will be abandoned after proof of concept. Whether the number is 87% or 95%, the root cause is the same: it’s almost never the technology.

The Real Problem

Here’s what actually happens:

A leader reads about ChatGPT or Claude and thinks, “We need this.” A team gets excited. Someone sets up an account. They run a few experiments. And then… silence. The project lives in one person’s head. There’s no measurement of what worked. No training for the wider team. No governance framework. No documentation of what was built. When that person leaves or moves to something else, the whole thing collapses.

Or they implement something bigger: a workflow automation or document processing system. It works brilliantly for three weeks. Then nobody maintains it. The prompts drift. The outputs degrade. No one notices because there’s no monitoring. Three months later, someone asks, “Whatever happened to that AI project?” and nobody remembers.

The pattern is consistent: Technology success + organisational chaos = project failure.

Why This Happens

It’s not stupidity. It’s that AI is new enough that most organisations haven’t figured out the discipline required to make it sustainable.

When you implement a new CRM system, you have processes. A CRM vendor gives you implementation guides. Your team gets trained. There’s a change management plan. There’s governance about who can access what. There are quarterly reviews of adoption metrics.

With AI, we’re still in the Wild West. People experiment in isolation. They build something, it works, and then they expect it to just… keep working. Without maintenance, measurement, or documentation.

MIT Sloan Management Review found that 91% of top data managers cited team challenges and change management, not technology limitations, as the main barriers holding them back with AI. McKinsey’s research reinforces this: companies with a formal AI strategy report an 80% success rate in AI adoption, compared to just 37% for those without one.

The constraint isn’t whether ChatGPT can analyse documents. It’s:

  • Does your team have a framework for finding AI opportunities systematically?
  • Once you build something, who owns maintaining it?
  • How do you measure whether it actually delivered value?
  • How do you scale from one team experimenting to the entire organisation moving systematically?
  • What happens when someone leaves and nobody knows how the system works?

These aren’t technology questions. They’re discipline questions.

Organisational chaos versus structured AI adoption pipeline

What Actually Works

The organisations that successfully adopt AI do something different:

They treat it like any other critical business capability: discovery, planning, execution, measurement, governance, training.

They don’t ask, “Can we use AI?” They ask, “Where will systematic AI use save the most time, reduce the most risk, or improve the most?” They map opportunities. They prioritise. They run pilots with proper measurement. They scale only what works.

They build documentation as they go. They train their teams. They establish governance that enables fast, safe decision-making rather than blocking it.

They measure ROI clearly. Not to prove AI works, but to understand where it’s creating value and where it’s not.

They treat the first person who builds something as a proof point, not the final solution. They systematically spread that knowledge across the team so it survives when people move on.

This takes discipline. But it’s the difference between a single successful AI project and systematic AI adoption that compounds over time. It’s also why we built the AI Native Programme as a structured monthly partnership rather than a one-off engagement. Sustainable AI adoption requires ongoing structure.

How the Platform Helps

The AI Native Platform is built around this exact insight. It’s not trying to replace ChatGPT or Claude (those are tools you’ll use within the platform). It’s infrastructure for the discipline part.

The AI Native Platform dashboard showing Discovery Canvas, Team Training, Governance, and ROI tracking

  • Discovery Canvas: systematically find where AI matters most in your business, rather than guessing
  • Roadmap & Prioritisation: rank opportunities by impact and effort so you invest in the right projects first
  • Task Management: keep AI project execution visible across your team
  • ROI Measurement: track actual returns, not assumptions, so you know what’s working
  • Governance Framework: move fast without fear of chaos, with built-in risk assessments and policy guardrails
  • Training Modules: build team capability so knowledge doesn’t walk out the door

Essentially, the platform exists because we realised that the bottleneck for most organisations isn’t the AI tools. It’s the discipline, measurement, and structure around them.

If you want to see what this looks like in practice, the AI Readiness Assessment is a good place to start. It helps you understand how AI-literate your team is, what tools they’ve already used, and how they feel about using AI in your organisation.

What This Means for Your Organisation

If you’re experimenting with AI today, you’ve probably got ChatGPT, Claude, maybe Copilot. Different teams using different tools. No shared framework for what’s working. No measurement of impact. No governance. You’re probably part of the 95%.

The path forward isn’t more tools. It’s systematic adoption. Discovery, measurement, training, governance, scaling.

That’s not sexy. But it’s what actually works.

Interested in understanding where AI could actually matter in your organisation? Get in touch. I’m happy to chat through your biggest opportunities and where the real bottlenecks are.

Using AI Right Now: A Business Guide Using AI Right Now: A Business Guide

Using AI Right Now: A Business Guide

A practical guide for New Zealand businesses on which AI tools to use and how to get real business value from them right now.

Tim Clark

Tim Clark

Co-founder · 7 July 2025 · 8 min read

The landscape has shifted quite a bit in the last few months - it’s less about finding the “best” model and more about finding the best overall system for what your business needs. Good news is picking an AI is actually easier than ever, though understanding how to use these increasingly complex tools effectively for business outcomes is the real challenge.

It’s less about finding the “best” model and more about finding the best overall system for what your business needs.

Which AI to Use for Business

For businesses wanting to use AI seriously, you’ve got three excellent choices: Claude (Anthropic), Google’s Gemini, and OpenAI’s ChatGPT. All three give you access to advanced and fast models, voice mode, document handling, code execution, decent mobile apps, and the ability to create images and video (though Claude can’t do images yet). Some features are free, but you’ll generally need to pay around $45-55 NZD/month per user for the full feature set your business needs.

There are other options - Grok if you’re big on X, Microsoft’s Copilot through Windows, or DeepSeek’s free Chinese model - but honestly, just stick with Gemini, Claude, or ChatGPT for most business applications.

Business Impact: Where AI Actually Helps

I spend considerable time helping businesses actually use AI to get things done, and the complexity can be overwhelming. Let me start with the practical business applications that are working well right now.

Deep Research: Your New Business Intelligence Tool

Deep Research is brilliant for producing high-quality reports that genuinely impress professionals I work with - lawyers, accountants, consultants, market researchers. This is where AI can immediately impact your business.

Practical business uses:

  • Market analysis: “Analyse the current state of the New Zealand construction industry, focusing on residential building trends, key players, and regulatory changes affecting small to medium builders”

  • Competitive intelligence: “Research our top 5 competitors in the Auckland accounting services market, including their pricing models, service offerings, and recent client wins”

  • Regulatory compliance: “What are the recent changes to New Zealand employment law that affect small businesses with 10-50 employees? Include practical compliance steps”

  • Supply chain research: “Find alternative suppliers for industrial packaging materials in the South Island, focusing on companies that can handle our minimum order quantities”

  • Client research: “Research this potential client’s business model, recent news, key personnel, and industry challenges” (brilliant for sales prep)

Deep Research reports aren’t error-free but are far more accurate than just asking the AI directly, and the citations tend to be correct. Each system works slightly differently - Claude and o3 with web search enabled work as mini Deep Research tools, while Google offers options like turning reports into presentations or infographics you can use with clients.

Document and Content Creation: Scale Your Communications

All three systems can produce professional documents, presentations, analyses, and communications. This is where businesses see immediate productivity gains.

Business applications:

  • Proposals and tenders: Give it your previous successful proposals and brief details about the new opportunity

  • Client communications: “Draft a project update email for our client explaining the delay in the Auckland office fit-out due to supply chain issues, maintaining professional tone while being transparent about revised timelines”

  • Policy documents: “Create an updated remote work policy for a 25-person Wellington tech company, incorporating recent changes to employment legislation”

  • Training materials: “Develop onboarding materials for new sales staff covering our CRM system, client communication standards, and territory management”

  • Financial reporting: Upload your data and ask for executive summaries, trend analyses, or board presentation materials

To get Gemini or ChatGPT to create professional documents reliably, select the Canvas option. Claude handles document creation well on its own.

The Right Model for Business Work

Each system offers multiple AI models - think of it like choosing the right tool for the job. You’ve got three tiers: fast models for casual tasks (Claude Sonnet, GPT-4o, Gemini Flash), powerful models for serious business work (Claude Opus, o3, Gemini Pro), and ultra-powerful models for complex analysis (o3-pro, which can take 20+ minutes to think through problems).

Fast models are fine for brainstorming or quick questions. But for anything business-critical - client proposals, financial analysis, strategic planning, legal document review - switch to the powerful model. Most systems default to the fast model to save computing power, so you need to manually switch using the model selector dropdown.

The free versions don’t give you access to the most powerful models, so if you don’t see these options, that’s why you need the paid version for business use.

I use o3, Claude 4 Opus/Sonnet, and Gemini 2.5 Pro for serious business work. For most business applications, stick with these powerful models for anything important.

Privacy for Business: Claude doesn’t train future models on your data, but Gemini and ChatGPT might (unless you’re using business/enterprise versions). You can turn off training features in ChatGPT without losing functionality. For sensitive business information, this matters. Andrew wrote a good post on this on LinkedIn.

Voice Mode: Perfect for Mobile Professionals

Voice mode is brilliant for busy professionals. The best implementations are in the Gemini app and ChatGPT’s app and website. Claude’s voice mode is weaker.

The killer feature for business isn’t just the conversation - it’s sharing your screen or camera. Point your phone at contracts, spreadsheets, equipment, or site issues.

The AI sees what you see and responds in real-time. I’ve used it to:

  • Review contracts while travelling
  • Get quick analysis of financial data on screen
  • Identify equipment issues on job sites
  • Translate documents or signs when dealing with international suppliers
  • Get instant feedback on presentations before client meetings

Most people use voice mode like Siri - you’re missing the business applications.

Working with AI: Business Best Practices

The most recent AI models can often figure out what you want without complex prompts, so just approach it conversationally rather than getting too worried about exact wording.

Key principles for business use:

Give business context: Most AI models only know basic information and the current chat. Provide context: company background, industry specifics, previous successful documents, client requirements. Upload files, images, or detailed briefs using the file upload option.

Be specific about business outcomes: Instead of “Write a marketing email,” try “I’m a Wellington-based accounting firm targeting small e-commerce businesses. Write a cold outreach email addressing their GST compliance challenges and our specialist e-commerce accounting services. Here’s our service details: [paste]”

Ask for business alternatives: The AI doesn’t get tired. Ask for 20 different approaches to a client problem, or 10 ways to improve a proposal. Then push the AI to expand on the approaches that resonate. But remember, treat it like you would any human. Be nice and give it context.

Use iterative refinement: All systems let you edit prompts after getting answers, creating conversation branches. This is perfect for refining business documents - you can explore different approaches and compare results.

Common Business Pitfalls

Hallucinations: While much improved, AI still makes confident mistakes. Answers are more reliable from the powerful models that do web searches. Always use AI for business topics you understand, and verify important facts, especially for client-facing work.

Not magic: The best AIs perform like very smart consultants on many tasks, but can’t provide miraculous insights beyond human understanding. If something seems impossible, it probably isn’t actually doing that.

Always engage: Have back-and-forth conversations. Don’t just ask for responses - push the AI, ask for alternatives, request refinements. This is where the real business value comes from.

Verification: For business-critical work, click “show thinking” to see what the model was considering. Not 100% accurate, but helpful for understanding the reasoning behind recommendations.

Getting Started: Your First Business Applications

Pick a system and invest the $45/month per user (free versions are demos, not business tools). Then test these three business applications immediately:

First: Switch to the powerful model and give it a real business challenge - a client proposal, competitive analysis, or strategic planning question. Provide full context and have an interactive discussion. Ask for specific outputs and refine until you’re happy with the business quality.

Second: Try Deep Research on a business question where you need comprehensive information - market analysis for a new service line, competitive intelligence, or regulatory research relevant to your industry.

Third: Use voice mode during your commute or between meetings to think through business problems, review presentations, or brainstorm solutions to client challenges.

The tools are genuinely impressive when used well for business applications. The key is understanding what they can actually do for your specific business needs and industry context.

Cheers, Tim

Want hands-on help applying these tools to your business? Our AI Accelerator programme gives business owners personal coaching and group workshops to build practical AI skills that deliver immediate results. If you’re leading a team, explore how our AI Native Programme can upskill your entire organisation. And for a look at real business outcomes, browse our AI use cases library.

The "AI Cake Fallacy" - Why NZ Businesses can't have their AI Cake and eat it too The "AI Cake Fallacy" - Why NZ Businesses can't have their AI Cake and eat it too

The "AI Cake Fallacy" - Why NZ Businesses can't have their AI Cake and eat it too

Stanford research reveals 41% of AI investments target the wrong areas. Learn the four zones of AI automation and why people-first adoption delivers better results.

Tim Clark

Tim Clark

Co-founder · 25 June 2025 · 6 min read

I’ll be honest with you - I’m living this contradiction every single day.

In my role as a business owner, I’m constantly looking for ways to cut costs and improve efficiency. When I see a process that could be streamlined or a role that might be automated, there’s this immediate mental calculation: “What would this save us monthly? Could AI handle this?” It’s almost automatic now, this lens of optimisation and cost reduction.

But here’s the thing - I’m obviously a huge AI user. And when I use AI tools, I’m not trying to replace myself. I’m trying to get rid of the tedious stuff so I can focus on what actually matters. I want AI to handle my scheduling, draft my initial emails, and organise my notes so I can spend more time on strategy, relationships, and the work that genuinely energises me.

The contradiction hit me recently when I realised I had been looking at my team’s roles through that “efficiency lens” while protecting my own work from the same scrutiny. I wanted AI to free me up to do more meaningful work, but I was unconsciously viewing their roles as potentially replaceable.

Last week, I stumbled across this Stanford research that perfectly captures what I call the ‘AI Cake Fallacy.’ Everyone wants their slice of the AI productivity gains, but nobody wants to give up the parts that matter most to them. And the data proves this disconnect is real - and it’s killing AI implementations.

The Problem We’re All Dancing Around

The Stanford study surveyed 1,500 workers across 104 occupations and found something that should make every business owner pause. Yes, 46% of workers want AI automation, but only for specific tasks. They want AI to handle scheduling, data entry, and file management. You know, the stuff that’s genuinely dull. They want to focus on strategy, creativity, and meaningful work.

Makes perfect sense when you think about it. Why would anyone want to spend their workday doing stuff they don’t enjoy?

But here’s where it gets interesting. The research shows that 41% of current AI investments are going to what they call the “wrong places” - either low-priority areas or “Red Light Zones” where workers actually resist automation.

We’re literally investing in the stuff people don’t want automated while missing the opportunities where they’d welcome it with open arms.

The Four Zones That Change Everything

The researchers identified four distinct zones, and understanding these may shifted how you think about AI implementation.

There’s the “Green Light Zone” where workers actually want automation and AI can deliver. Think automated scheduling, basic data processing, routine admin work. Companies operating here see both productivity gains and employee satisfaction. It’s the sweet spot we should all be aiming for.

Then there’s the “Red Light Zone” - high AI capability but low worker desire. This is where we often get seduced by what’s technically possible rather than what’s actually wanted. Creative tasks, strategic planning, client relationship management. Force AI into these areas and you’ll create resistance and disengagement faster than you can say “digital transformation.”

The study found that Arts and Design workers only wanted 17% of their tasks automated. Editorial roles consistently rated as requiring essential human involvement. The pattern is clear - workers fiercely protect tasks involving creative expression, human relationships, strategic thinking, and work they find genuinely enjoyable.

The Reality Check I Needed

There’s a cycle that is all too familiar. We saw it with failed digital transformation projects as well. Initial excitement about AI capabilities, followed by the slow realisation that implementation isn’t working as expected, then confusion about what went wrong, followed by the frustrated “just make it work” directive that leads nowhere.

We skip the critical step of understanding what our people actually value about their work before trying to automate it away. We assume that because AI can technically do something, it should do something.

What Workers Actually Want

The most telling finding from the Stanford research is that workers prefer equal partnership with AI, not replacement. They want collaboration, not elimination. This isn’t about being resistant to change - it’s about being protective of meaning.

When you dig into what creates the strongest resistance, it’s the human-centred stuff. Customer service conversations where relationship building matters. Sales interactions that require reading between the lines. Training and mentoring colleagues where empathy makes the difference. Any work requiring genuine human judgment or emotional intelligence.

Workers will gladly hand over the scheduling, the data entry, the routine administrative tasks. But ask them to surrender the creative problem-solving, the strategic thinking, the work that makes them feel human, and you’ll hit a wall.

The Skills Revolution We’re Missing

The research reveals something important about where we’re heading. There’s a clear shift happening from information-processing skills toward interpersonal and organisational competencies. The traditionally high-wage work like data analysis is becoming less valuable, while relationship building, creative thinking, and strategic judgment are becoming more important.

This isn’t just about redistribution - it’s about fundamental redefinition of value. The most valuable human skills are becoming the ones that are most distinctly human.

Getting This Right

The companies that succeed with AI won’t be the ones who automate the most. They’ll be the ones who understand that AI works best when it amplifies human capability rather than replacing it.

This means having real conversations with your team about which tasks they’d actually want automated. It means understanding what they find meaningful about their work before you try to optimise it away. It means designing AI implementation with them, not to them.

The Bigger Picture

The Stanford research proves what many of us have suspected - the real opportunity lies in that collaboration zone where AI handles the tedious while humans focus on what they do best: thinking, creating, relating, and innovating.

As business owners, we need to resist the temptation to see AI as primarily a cost-cutting tool. Yes, it can reduce costs, but its real value lies in freeing our people to do more valuable work. When we get this right, we don’t just save money - we create more value.

The question isn’t whether your business can have its AI cake and eat it too. It’s whether you’re smart enough to share the cake in a way that everyone wins. Because the research is clear - when you get this alignment right, both productivity and satisfaction go up.

And honestly, isn’t that the kind of transformation we actually want to be part of?

Ready to align your AI investment with what your team actually wants? Our AI Clarity Session helps you identify the green-light opportunities where AI delivers both productivity gains and employee satisfaction. Or explore our real-world AI use cases to see practical examples of people-first automation in action.

Leading AI adoption with a clear vision Leading AI adoption with a clear vision

Leading AI adoption with a clear vision

Companies with a formal AI strategy see 80% adoption success vs 37% without one. Learn how to define, craft, and share an AI vision that brings your team along.

Phil Vinall

Phil Vinall

Co-founder · 23 June 2025 · 4 min read

Does AI spark excitement or unease in your team? As a business leader, your clear vision is needed to ensure smooth AI adoption.

For any AI effort to really hit its stride, it needs a clear purpose, right from the very top. A recent McKinsey report outlined research where not having a clear vision can cause “significant division” in organisations. This kind of disconnect, where only 45% of staff feel good about AI adoption compared to 75% of leaders, shows a clear gap in understanding and alignment. Some employees might even start using their own AI tools on the sly (“shadow AI”) if they’re not happy with, or can’t get to, the official ones. That’s a recipe for unmanaged risks for your business.

However, companies with a formal AI strategy report an 80% success rate in AI adoption, significantly higher than the mere 37% for those without one. This significant 43 percentage-point gap highlights how vital a clear vision is to successful AI adoption.

Companies with a formal AI strategy report an 80% success rate in AI adoption, significantly higher than the mere 37% for those without one.

Defining the vision

Successful AI adoption begins with envisioning what an AI-Native version of your business looks like. It’s about imagining what your business could become in the next few years with an AI-first mindset. This isn’t just about the technology you’ll use; it’s about seeing how AI fundamentally changes how your business operates, how decisions are made, and how problems are solved. Crucially, it’s about understanding what new expertise, culture, and mindset your people will also need to make that vision possible. Start with some of your key frustrations and aspirations - what do you wish you could do away with, and what do you want to do more of? What would your business look like if you started it today with AI running at the heart of it?

Crafting the vision

Once the vision is becoming clear, it’s essential to ground it in what’s achievable today - by deepening your practical understanding of AI’s capabilities (and limitations) within your unique business context. AI isn’t a magic wand - it requires thoughtful planning, thorough testing, and expert human oversight. As a result, AI works best as a collaborative partner, not just a tool. Think “humans multiplied by machines”, genuinely augmenting your team’s abilities rather than replacing them. Generative AI can typically boost performance by as much as 40% in a variety of areas for highly skilled workers. This allows your team to focus on higher-value, more engaging (and satisfying) work.

Sharing the vision

Finally, and perhaps most crucially, it’s important to effectively communicate this vision to your team. Transparent communication is vital to ensure your people feel included and informed, rather than threatened by change. It’s about painting a picture where they can genuinely see themselves thriving alongside AI, not being replaced by it. Leaders who proactively educate employees, involve them in decisions, and maintain transparency can transform resistance into advocacy. This empathetic approach helps bridge the gap between management and employee perceptions, fostering a culture of trust and shared success. Remember, successful AI adoption isn’t about rushing into technology; it’s about bringing your people along thoughtfully.

What’s your initial thought or biggest question about defining a powerful AI vision for your business? We’d love to hear from you.

Need help crafting your AI vision? Our AI Clarity Session is a focused 2-hour workshop where we help you define exactly what AI could mean for your business and align your leadership team on a clear direction. For a deeper dive into the leadership skills required, read our article on essential leadership skills for AI adoption.

Essential leadership skills for successful AI adoption Essential leadership skills for successful AI adoption

Essential leadership skills for successful AI adoption

McKinsey says leadership is AI's biggest hurdle. Discover the essential skills NZ business leaders need to drive successful AI adoption and empower their teams.

Phil Vinall

Phil Vinall

Co-founder · 1 June 2025 · 3 min read

The biggest hurdles when adopting AI aren’t about the tech. They’re actually about you - how will you lead AI adoption in your organisation?

The biggest hurdles when adopting AI aren’t about the tech. They’re actually about you.

A report from McKinsey said, “The biggest hurdle to success [with AI] is leadership.” And it’s a common theme in other findings. MIT Sloan Management Review (April 2025), found that 91% of top data managers said that “team challenges and managing change” were the main problems holding them back with AI.

For Kiwi businesses, getting AI working well isn’t just about buying new software or hiring a tech guru. It’s about how you, as the leader, tell the story of AI, get your people ready, and weave it into your unique business.

Your role in leading AI adoption

Sharing the vision: If you don’t have a clear idea of why you’re bringing in AI, or what you want it to achieve, your efforts can end up all over the place. As Codewave pointed out (May 2025), “AI actually helps the business when the top leader leads the way.” Your team needs to understand why and how AI is coming into your company.

Turning worries into wins: It’s fair enough for people to worry about AI taking their jobs. But staff are often less scared of AI itself, and more worried about how it will change their daily work, especially when things aren’t clear. This is where your leadership is crucial. When you show that AI is here to help people and not replace them - it changes everything. For example, Microsoft’s own studies (April 2025) found that staff using an AI Copilot, saw a 10-20% boost in getting things done and 68% felt happier at work. It helped free them up for more interesting, important (and valuable) work, and boosted their confidence in their future.

Leading the journey: Bringing in AI means some changes in the workplace. You need to be the kind of leader who can guide your team through this. You can do this by setting clear expectations, letting them try out new things, and building trust. When you, as the leader, actively share exploration of AI and its benefits, you create a workplace where your team sees it as a helpful partner, not something to be afraid of.

Adopting AI is less about a tech race and more about a thoughtful, people-focused journey. For Kiwi businesses, this means looking beyond the gadgets and truly focusing on your team and your style of leadership. When you, the leader, are strong and understanding, your AI plans will have a clear purpose, get results, and grow right alongside your business.

What’s your biggest question or fear about leading AI adoption within your business? We’d really love to hear from you.

Want to develop your AI leadership skills with expert guidance? Our AI Clarity Session gives you a structured 2-hour workshop to build your AI vision and roadmap. For ongoing support, the AI Native Programme helps you lead your team through the full adoption journey. You might also enjoy our related article on leading AI adoption with a clear vision.