The Most Expensive Line in the Budget Is What You Don't Know About AI

They're restructuring their pricing model away from billable hours toward automated, outcome-based delivery. 64% of organisations are now reporting measurable ROI from AI. Thomson Reuters found that while 40% of organisations say they've adopted AI, only 18% are actually measuring whether it's working.

These are all interesting data points. But the one I keep coming back to is from Daniel Kahneman's work on expert overconfidence: the phenomenon where the more senior you are, the more certain you feel about things you actually have no evidence for. In Thinking, Fast and Slow, he shows that experts are often worse at prediction than simple algorithms; not because they're inadequate or stupid, but because they're confident. And confidence, unchecked, is expensive.

Based on my work across industries - with senior boards and C-suite leaders across financial services, policy, retail, professional services, manufacturing - this is the story of AI adoption right now.

The unknown unknowns

In the month of March, I've been inside a multinational financial services organisation, a large education group, a manufacturing company, a leading global university, and a large private equity house. When we first sat down with each of them, they told me they were already "doing AI." High adoption figures. Copilot licences rolled out. A few people using ChatGPT. This is reported internally as a good thing.

And then, after a few conversations, a workshop, some hands-on time with the tools, their idea of "doing AI" and the reality of what sophisticated, configured AI can actually do in a professional setting turned out to be two entirely different things.

This is the pattern we're seeing everywhere right now. And this unknown unknown is the most expensive line item in your budget - it's just not appearing on any balance sheet.

Donald Rumsfeld took a lot of flak for his "unknown unknowns" line, but he was right. The things you don't know you don't know are the ones that hurt you. Nassim Taleb would say most organisations have built their AI strategy on assumptions they haven't stress-tested; that's fragility, and fragile things break suddenly, not gradually.

The danger right now isn't so much that leaders are ignoring AI. It's that they think they've got it covered. A Copilot licence, a lunch-and-learn, an "AI champion" in the IT department.

That's not a strategy though; it's a fig leaf.

The farmer and the combine harvester

My trusty farmer analogy comes out in most boardrooms I visit, and it seems to land:

For centuries, the farmer with a scythe developed a particular kind of strength. Forearms like rope, an instinct for weather, the ability to assess soil through smell and touch. The ability to work a single field with extraordinary precision and endurance. That strength was hard-won and valuable.

Then the combine harvester arrived.

And suddenly the most important capability wasn't physical strength. It was the ability to think in hundreds of hectares. To plan logistics across an entire harvest. To make decisions about crop rotation, soil chemistry, market timing, and supply chains. The muscles the farmer had spent a lifetime building weren't just unnecessary at this point of technological advancement; they were a distraction from the capabilities he now needed.

AI is doing exactly this to knowledge work.

The skills that made someone excellent at drafting a report, summarising a case file, analysing a dataset, or preparing a board paper; those are being automated, today, by tools that do it faster and, when AI is configured properly, to a higher quality. That's not coming soon or on the horizon, that's here already, in full swing, and if you think it can't reach your quality; that's a quiet sign of an unknown unknown.

The question isn't whether AI can do your team's routine cognitive work. It can. The question is: can your team think in hundreds of hectares?

Because the point of the combine harvester was never to automate the scythe. It was to extend the realm of what's possible.

What really matters to all of us though, is that the combine harvester doesn't drive itself to a successful harvest. It needs a farmer who can coordinate it with the fertiliser schedule, the workers in the field, the animals that need bringing in, the market timing, taking account of the weather, the soil, the uncertainty of each. The technology is powerful, but the human is the one who sets the direction, the goals, the purpose. Who sees the whole farm, not just the field.

Formula One works the same way. The car is a technological marvel; you cannot win F1 races without it. But the car doesn't win races. The seamless coordination of humans does; the driver's deftness and speed, the pit crew's accuracy, the strategist reading the race in real time, the engineers making split-second calls. Every one of those is a human capability amplified by extraordinary technology. Remove the humans and you have a very expensive car sitting in a garage.

So when people ask me "what can't AI do?", I tend to push back. AI can see patterns across an organisation. It can pre-empt regulatory shifts. It can synthesise information at a scale no human can match. The question isn't what AI can't do; it's who's driving.

How can you deploy agentic AI across your organisation if you don't know where to apply the agents? How can you have a human in the loop if you don't know what the loop is? How can you direct AI to do your bidding if you haven't worked out what your bidding actually is?

No matter how powerful the technology becomes, the human driver is the one who sets its goals, direction, and purpose. That will always be true. And that's why the most valuable capability in your organisation right now isn't technical; it's strategic.

What happens to your brain when you hand it a combine harvester

There's a growing body of research in cognitive load theory and neuroeconomics that I find fascinating, and it has practical implications for every hiring decision you make in the next twelve months.

Sweller's cognitive load framework, developed over three decades of experimental research, shows that when routine processing is offloaded, working memory doesn't simply idle. It redirects capacity toward what psychologists call "germane load": the deep, schema-building thinking that creates expertise. Pattern recognition. Creative problem-solving. Strategic reasoning.

Back to the farm. The farmer swinging a scythe all day could certainly think while he worked; but the nature of that thinking was constrained by the task consuming his attention. His cognitive bandwidth was tied up in the immediate, the physical, the repetitive. Give him a combine harvester, and that capacity redirects from processing and drudge toward strategy and pattern recognition; the higher-level thinking we excel at, and which is needed to apply AI effectively within complex organisations. Where to add AI into processes; how and when and under what structure; in ways that keep people productive, boards informed, skills current, and the organisation moving forward. That's complexity. And complexity is a human speciality.

The same thing is happening in offices right now. When a financial analyst offloads the routine parts of their work to a well-configured AI, they don't become redundant. They become more strategic, more creative, more valuable. Their brain redirects capacity from processing to pattern recognition.

This has implications for hiring, too. The skills profile that made someone an excellent analyst five years ago; meticulous, fast, good with detail; may not be the profile you need now. You might need someone who thinks in systems, makes unexpected connections, holds complexity, and communicates with clarity. The combine harvester didn't just change the work. It changed the kind of farmer you needed.

The question for leaders is whether your people are being given the space and the tools to make that cognitive shift, or whether they're still swinging the scythe because nobody has shown them the combine harvester exists.

From the field this month

This month has been a masterclass in the gap between perception and reality.

At a large education group, we deployed 20 personalised AI assistants in a single hour. Each one configured for a specific person's role, responsibilities, and development goals, all aligned with their organisation's context, strategy, and guiding principles. Not a generic chatbot; a thinking partner that knows their context and shapes their professional development pathway. The reaction wasn't polite interest. It was the kind of visceral excitement that tells you something fundamental has shifted. Sounds like "what? No, did it really just do that?" and the ever-present "ahaaaaa" of recognition emanate from the room.

At a leading European university, MBA students came to ask me afterwards whether I ran a course they could sign up for. Several were ready to pay over $3000 for an online version that calls itself practical but in reality relaxes into the technical. These students wanted something real; something that would actually change how they work - tools they can leave with configured to their purposes that can start doing work for them right away.

A manufacturing company signed up for their first AI transformation programme because their MD realised in a single conversation that his mental model of what AI is and what it actually does were two completely different things. His unknown unknowns turned out to be enormous. And he's a smart, experienced, senior leader. That's the point.

What your board should have noticed this month

Thomson Reuters: 40% AI adoption, but only 18% measuring ROI. Kahneman would recognise this instantly: the planning fallacy applied to technology investment. Organisations are buying AI, deploying AI, and not measuring whether it works. 40% don't even know if ROI is being tracked. That's not adoption; it's procurement without ownership and accountability.

Microsoft launched "Zero Trust for AI." They're treating AI agents as a new class of enterprise identity requiring governance, audit, and access controls equivalent to human users. If Microsoft is building an entire control plane for AI agents, they expect those agents to proliferate at a scale requiring governance infrastructure. Your board should be asking: do we have that infrastructure? Or are our people using AI tools nobody is monitoring, nobody is securing, and nobody is accountable for?

Accenture acquired UK AI firm Faculty. The enterprise AI capability arms race is real. Large consultancies are buying deep applied AI capability and selling "safe, secure, outcome-driven" AI transformation at scale. For everyone else, the window to build capability in-house rather than renting it from Accenture is narrowing.

PwC said AI sceptics have "no place." And I think they're right; though not quite in the way they mean it. The question isn't whether to use AI. That question is settled. The question is how; how to use it responsibly, strategically, and in a way that makes your organisation and your people better rather than just faster. That's a question of capability, not procurement; and it's the question almost nobody is asking properly yet.

Copilot to the shops, or a thousand hectares?

Most boards are still asking: "Are we using AI?" But the research, the evidence, the real life case studies show us: that is fundamentally the wrong question.

Someone logging into Copilot once a week is not the same as someone who has fundamentally changed how they work. Riding a combine harvester to the shops is not the same as strategically planning the harvest of thousands of hectares across three dimensions. Usage is not adoption. And adoption without measurement is just hope.

The right questions are harder.

What kind of adoption are we measuring? Safe adoption? Responsible adoption? Productive adoption? Or are we just counting licences and calling it progress?

What does our AI asset register look like? Do we actually know what AI tools are in use across the organisation, by whom, and for what purpose?

Are our people and our business thriving through this transition? This is the question almost nobody asks. I'm currently working in the IEEE's Global AI Flourishing Initiative (alongside Iliana Grosse-Buening and an extraordinary group of researchers and practitioners) with a view to developing international KPIs for measuring exactly this. Because every major governance framework measures whether the technology is trustworthy. None of them measure whether the people and the business undergoing AI transformation are actually thriving, producing, delivering value. It's not technology OR people. It's technology AND people, for better. That gap of how is what needs filling.

If you can deploy AI brilliantly and make your people miserable or redundant, you may be suffering a lack of imagination about what's actually possible now. The point of the combine harvester was never to automate the scythe. It was to transform what one person could achieve.

Teaching someone to drive a specific car is not teaching them to drive

One more thing, because I see this mistake everywhere and it costs organisations a fortune.

Training people on a specific AI tool is not the same as building AI capability. And confusing the two is an irresponsible investment of time and money.

Tool training has a very short lifespan. The tools are changing so fast that not only will your training be insufficient - there is a genuine shortage of people who deeply understand these tools and their application to your context - but it will be out of date by the time you finish delivering it. The platform you're training on today will look entirely different in six months. The features you're teaching may not exist.

The key is building transferable AI capability: the skills that work across every tool and every platform. Configuration. Contextual knowledge-base design. Prompt architecture. Responsible, governed use. Critical evaluation of AI outputs. Workflow redesign. Understanding what AI can and cannot be trusted to do. Knowing how to build guardrails that protect your organisation without strangling innovation.

These are the capabilities that travel. These are the ones that compound over time rather than depreciate. These are the ones that make your people AI-capable rather than temporarily AI-trained.

Training someone to use Copilot is like teaching someone to drive a specific car. Building AI capability is like teaching someone to drive. One expires when the model changes. The other lasts a career.

The category error, the Ferrari, and the three-legged stool

After working across dozens of organisations in the past year, the pattern is remarkably consistent.

First, before we arrive, they have treated AI as a technology project. It isn't. BCG's research is unambiguous: 10% of successful AI transformation is the technology, 20% is the process, and 70% is the people. Yet most organisations spend 90% of their budget on the 10%. This is the category error, and it's why the failure rate remains stubbornly high.

Second, before we arrive, they have trained individuals but not built organisational scaffolding. You can give someone the best AI skills in the world, but if the organisation punishes experimentation, restricts tool access, or has no clear strategy for how AI fits into the business, those skills atrophy within weeks. Individual capability without organisational infrastructure is like giving someone a Ferrari and no roads.

Third, before we arrive, they have skipped to tools without understanding skills and use cases. "We bought everyone Copilot licences" is not an AI strategy. It's a procurement decision - and an expensive one. Capability requires three things working together: the skills to use AI effectively, the right tools for the job, and clear use cases tied to real work. Miss any one of those three and the whole thing falls over. Most organisations we see have addressed one at best, and are now lamenting the lack of ROI on their AI investment.

Three questions for Monday morning

Ask your team:

First: what AI tools are people in this organisation actually using right now, and do we know?

Second: if AI could eliminate one task that wastes the most time for the least value, what would it be?

Third: how would we know if our AI adoption was actually working?

If you can't answer those confidently, you've found your unknown unknowns. And now you can start doing something about them.

I'm Mackenzie Howe, Co-Founder and CEO of Atheni.ai. We build AI capability in organisations across industries, from financial services, education, manufacturing, retail, to government. I sit on AI Advisory Boards, and in particular on international committees developing global standards for measuring human flourishing through AI adoption. If any of this resonated, I'd love to hear what you're seeing in your own organisation.