When AI Gets Expensive, the ROI Question Gets Honest

What I said on RTHK Radio 3 this week — and why the subscription shift is forcing a conversation most organisations have been avoiding

Carolyn Wright opened our conversation on RTHK Radio 3 this week with the question every board in Australia is now sitting with: how are companies actually going to deliver a return on investment from AI?

It’s the right question. And the fact that it’s arriving now, not twelve months ago, tells you something important about where we are.

A year ago, most organisations were treating AI as an experiment. The tools were largely free or nearly free, the business case was speculative, and the board conversation was mostly “are we using it?” Now the tools are moving to subscription. Now the bills are landing. And now, finally, the question is “what are we getting for this?”

The honest answer is: not as much as the business case projected. And the reason isn’t what most people think.

The fee changes the question

There’s a particular dynamic that happens when a technology transitions from free to paid.

When something is free, the cost of getting it wrong is low. You experiment, you check, you hedge, you layer human oversight on top because the oversight doesn’t cost you anything compared to what you’re not paying for the tool. The whole relationship is low-stakes enough that imprecision is tolerable.

When you’re paying for it, that calculus shifts. Now the checking isn’t free either. Now the time your team spends verifying AI output is time they’re not spending on something else, and you’re paying for the tool and the verification simultaneously. The ROI calculation changes, and not in the direction the business case assumed.

Most AI ROI models have two numbers: what the technology produces, and what it costs to licence and deploy. The first number looks good. The second number is real but manageable. What most models are missing is the third number: the time your people are spending checking the output before it goes anywhere.

I’ve been calling this the Checking Tax. It’s the invisible cost nobody put in the business case. And it’s sitting between the two numbers that did make it in, quietly absorbing the productivity gain that the tool was supposed to deliver.

The 2026 AI Trust Report found that senior leaders are spending almost as long verifying AI output as generating it. Gartner’s research tells you only one in five AI projects delivers against its ROI expectations, with less than two per cent of executives reporting returns of twenty per cent or greater. A National Bureau of Economic Research survey of nearly 6,000 executives across the US, UK, Germany, and Australia found that eighty-nine per cent reported no measurable labour productivity impact from their AI investment.

Those numbers aren’t because the tools are bad. They’re because the checking is absorbing the gain.

Why we check

Carolyn asked me why people are still checking everything when the tools are so capable. It’s a fair question, and I think the honest answer has two parts.

The first is genuinely about trust, and it’s more complicated than it looks. Most people using AI professionally right now are in a state I described on air as a little bit like being a teenager. You want to look like you know what you’re doing. You’re using the tools, claiming the outputs, presenting the results. But underneath, there’s a layer of suspicion. “Is this right? Did I ask that correctly? What if I missed something?” You’re checking not because you’ve been told to, but because you don’t quite trust it yet and you don’t fully trust yourself using it.

That’s a developmental stage. It’s real, it’s normal, and it’ll pass. The search engine went through it. We didn’t trust early search results; we’d verify them against other sources, we’d go to the seventh or eighth result because the first few seemed too confident or too generic. Over time, we learned to search better. The technology got better at surfacing what we actually needed. We found a working relationship with it. AI is going through the same process, and it’ll take time, and that’s not a crisis — it’s just where we are.

The second part is more structural, and it’s the part that doesn’t fix itself over time.

Seventy-seven per cent of leaders hit what The 2026 AI Trust Report calls a structural trust ceiling every week. Not personal doubt, though that exists. Structural doubt: the absence of institutional permission. Your organisation has deployed the capability but hasn’t built the governance that tells people which decisions AI can produce without a human review, and which ones still need one. In that absence, the rational default is to check everything.

That’s not a people problem. That’s an architecture problem.

The sliding scale nobody’s built

One of the clearest things that came out of our conversation on air was this: the question isn’t “do you trust AI or don’t you?” It’s a sliding scale. Some tasks, you can hand over completely. Others, you can’t, and probably shouldn’t. And most of the useful work lives somewhere in the middle where the appropriate level of human involvement depends on the task, the stakes, the context, and the quality of the governance you’ve built around it.

I’ve been working with a framework I call HUMAND, a way of thinking about what work is best done by Humans, Machines, AI, or some combination — and the key insight it surfaces is that the sliding scale isn’t fixed. It moves. The task that needed full human oversight twelve months ago might be fully delegable today. The task that seems obviously safe to automate might carry consequences that only become visible after you’ve handed it over.

The right analogy I used on air is the one for a human employee. When you bring someone new into an organisation, you don’t give them full autonomy on day one. You give them access appropriate to their skills, their experience, the job they’re doing, and the stakes of the decisions they’re making. You expand their authority as they demonstrate judgment. You pull it back when the risks change. You never give them the keys to everything.

AI should be treated the same way. The question isn’t “how much can AI do?” It’s “what has this organisation consciously decided it’s authorised to do, in what contexts, with what level of human oversight?” That’s the conversation most organisations are having implicitly, through individual judgment calls, instead of explicitly, through governance design. And the individual judgment calls default to checking everything, because that’s what rational professionals do when nobody’s told them where the accountability sits.

Where you draw the money line

The most revealing version of this question came up when we were talking about AI and financial transactions. Carolyn raised a scenario that I think is the sharpest test of how much you actually trust AI: would you allow it to make payments on your behalf?

Your bank already asks questions when you make an unusual payment. Who are you paying, why, have you verified this is legitimate? Those friction points exist because payments are high-stakes, high-consequence, and difficult to reverse. The bank doesn’t trust you unconditionally, even though you’re the account holder, because the cost of getting it wrong is significant.

Now layer AI into that. If you instruct your AI to handle routine payments, fine: low-stakes, low-consequence, fully within the zone you’ve defined. But if your AI starts making payments in response to context it’s inferred, or based on instructions that weren’t fully specified, you’ve entered territory where the architecture of trust matters enormously. Who authorised that? On whose behalf? Under what instruction? What happens if the inferred context was wrong?

The financial payments question isn’t just about payments. It’s the clearest illustration of what the governance conversation is actually about. When the stakes are high enough to be irreversible, the question of what you’ve explicitly authorised becomes unavoidable.

And here’s what I find interesting: most organisations are having that conversation about their most consequential decisions, while leaving the medium-stakes decisions to individual judgment. But the medium-stakes decisions are where the volume lives. They’re where most of the Checking Tax accumulates. And they’re where the governance conversation is most worth having.

The guardrail model

Carolyn offered a good frame in the conversation: it’s like a GPS.

A GPS guides you along the best route it’s found. It warns you if you’re about to turn the wrong way. It recalculates when you deviate. But it can’t physically stop you from doing something dumb. The guardrail is informational, not mechanical. If you decide to drive into a field, the GPS notes your position and waits for you to reconnect with the road.

That’s a useful way to think about AI governance. The point isn’t to build walls that prevent AI from touching certain domains. The point is to build guardrails that make the right choice clear, that create friction when you’re approaching something the organisation has decided needs human judgment, and that log what happened so you can revisit the call when the context changes.

Organisations that are doing this well aren’t doing it because they’re more worried about AI. They’re doing it because they’ve figured out that the guardrails are what let you give AI more room to run, not less. When your people know what AI is authorised to do, they can hand it off. When they don’t know, they check. And the checking is the cost.

This is what the Decision Trust Zones framework is designed to answer: what has this organisation explicitly decided AI can decide autonomously, what requires AI assistance plus human judgment, and what requires human authority only? The organisations with clear answers to those questions score significantly higher on trust maturity metrics. The ones without clear answers are paying the Checking Tax.

We’ve been here before

I said something on air that I want to expand on here, because I think it’s important context for how to feel about all of this.

We’ve always used tools. This isn’t new. From the moment our ancestors picked up a stick to extend their reach, the history of human capability has been the history of tools. Every tool raised the same questions: how much do we trust this? What do we use it for? Where does it stop and we begin? We’ve been navigating that relationship for as long as we’ve been human.

What we also know, looking at the last few decades, is that we tend to overshoot with new technologies and then correct. We over-indulged in social media. We treated the early internet as though all information was equally trustworthy. We consumed digital music in ways that broke an entire industry before a new model emerged. Human nature. We see the capability, we lean in, we find the edges of what works, and then over time we find the equilibrium.

That doesn’t mean the overshoot is fine. The consequences during the correction phase can be significant. What it means is that the correction happens, and the people who navigate it best are the ones who started asking the governance questions earlier rather than later.

Right now, most organisations are in the overshoot phase with AI. They’ve leaned in, sometimes past where their governance is ready to support them. The subscription model transition is the forcing function that brings the ROI question into focus. And the ROI question, if you follow it honestly, leads you to the Checking Tax, which leads you to the governance architecture, which is where the real conversation needs to happen.

The thing that doesn’t change

Here’s what I find genuinely reassuring about this moment, and I said it on air too.

The technology is extraordinary. What AI can do now that wasn’t possible three years ago represents a genuine capability expansion. Some of what it’s made possible in research, in synthesis, in pattern recognition, in communication, is going to compound over decades in ways we can’t yet fully see. I’ve been thinking and writing about technology futures for thirty years and I don’t say that lightly.

And the deepest truth about it is also the oldest one: we’re tool-using beings. We’ve always extended what we can do through the tools we build. The question was always how much authority to give them, what judgment to delegate, where the human has to remain in the loop. Those questions didn’t arrive with AI. They were there with every tool that came before. AI just makes them more consequential and more urgent to answer.

The organisations that’ll close the ROI gap aren’t going to do it by having better AI. They’re going to do it by having clearer governance. By defining what AI is authorised to decide without human review, what needs AI assistance plus human judgment, and what requires human authority only. By running a verification audit that tells them what their actual trust maturity score is, in their organisation, with their people. By assigning someone to own and update those definitions as the technology evolves.

None of that requires a new platform. It requires a conversation that most organisations haven’t had yet, because it looks like a technology question when it’s actually a governance question. And governance conversations are harder, slower, and less satisfying than capability announcements. But they’re the ones that produce the return.

Three things worth doing in the next thirty days

Carolyn asked me what practical advice I’d leave people with. Here’s what I said.

First, run a simple audit this week. Ask a handful of your people to track two numbers alongside their AI use: how long generating, how long verifying. Do it for five days. The ratio you get is your organisation’s Checking Tax in its rawest form. Before you design any solution, you need to see the problem.

Second, have the explicit conversation about your three highest-volume AI use cases. For each one: has your organisation decided whether the output needs human verification before it moves? Not assumed. Decided. Documented. Communicated. If the answer is “not really,” that’s the gap. The fix isn’t to tell people to check less. It’s to tell them what checking is actually for, and when output can move without it.

Third, assign ownership of this question. Somebody needs to hold the responsibility for updating the zones as the capability evolves. As the tools improve, as your use cases expand, as the regulatory environment shifts, the answer to “what is AI authorised to decide here?” will change. Someone needs to own that question, or it’ll calcify and the checking will creep back in.

That’s not a massive programme. It’s three conversations. But they’re the right three.

What the subscription shift actually gives you

There’s something I want to close with that I didn’t fully get to on air.

The transition from free AI to paid AI looks like a cost increase. And in the short term, for some budgets, it is. But it’s also a gift, in a strange way. Because it forces the question.

When AI is free, the Checking Tax is invisible. You’re not tracking it, because the tool it’s checking isn’t on the balance sheet. You’re not asking whether the ROI is there, because there’s no cost to justify. The question never comes up.

When you’re paying, the question arrives. And when the question arrives honestly: not “are we using AI?” but “what are we getting for it and why is the business case not closing?” — you’re finally in a position to find the real answer.

The real answer isn’t that the tools aren’t good enough. It’s that most organisations are paying the Checking Tax without knowing it, and they haven’t built the governance that would let them stop.

That’s the problem worth solving. It’s also the most solvable problem in the AI ROI conversation right now.

Not a technology problem. A governance architecture problem. And unlike the capability questions, where the pace of change is genuinely hard to keep up with, the governance question is one we already know how to answer.

We just haven’t answered it yet.

Choose Forward.

Morris Misel spoke with Carolyn Wright on RTHK Radio 3 Morning Brew on Tuesday 21 July 2026. He is a foresight strategist based in Melbourne, working with boards, leadership teams, associations, and government on the human and organisational dimensions of technology change. More at morrisfuturist.com.

FAQ

Why isn’t AI delivering the ROI organisations expected?
Most AI ROI models count what AI produces and what it costs to licence, but miss the third number: the time people spend verifying output before acting on it. When an organisation hasn’t defined what AI is authorised to decide without human review, the default is to check everything. That verification overhead, which I call the Checking Tax, can absorb most of the productivity gain the tool was supposed to deliver.

What changes when AI moves from free to subscription pricing?
When AI is free, the Checking Tax is invisible, as verification time doesn’t show up against a licence cost. When you’re paying, the arithmetic changes. Now you’re paying for the tool and for the time spent checking it simultaneously. That’s when the ROI question gets honest, and when the governance conversation that should have happened at deployment becomes unavoidable.

What is the HUMAND framework for AI decisions?
HUMAND is a decision framework for determining what work is best done by Humans, Machines, AI, or some combination. The key principle is that the right level of AI autonomy is a sliding scale, not a binary. Like a human employee, AI should receive authority appropriate to the task, the stakes, and the governance structure around it: expanded where trust is earned, constrained where consequences are significant.

What’s the first practical step for addressing AI ROI gaps?
Run a verification audit for one week. Ask a cross-section of your team to log two numbers alongside their AI use: how long generating, how long verifying. The ratio reveals your organisation’s actual Checking Tax. Most leaders are surprised by the result. Once you can see it, you can address it — starting with explicitly defining what AI is authorised to decide without human review in your three highest-volume use cases.