Global management consultancy Bain & Co has made a name for itself in recent years with some astute analysis of enterprise AI adoption trends, contrasting strategic aims with operational realities, and documenting the gap that sometimes appears between them.

This week sees the publication of its Global Technology Report 2026, which positions AI as “the most consequential technology of our lifetimes”. Doubtless that is true, but it is also an ambiguous statement, because what are those consequences?

One is that there is a gulf between the AI industry’s hardware needs and its software revenues, which Bain & Co suggests is the biggest challenge of this era – beyond even last year’s focus on ROI. The reasoning is stark: according to the report, the sector will need to generate revenues of $6 trillion a year by 2031 just to fund its planned infrastructure buildout.

But existing consumer and enterprise AI applications might generate only $1.2 trillion to $1.8 trillion by that point, it suggests, meaning that vendors will have to find up to $4.8 trillion in new revenue. But from where?

But even that is an optimistic assessment, according to other recent figures. Statista and MarketsandMarkets are among the research firms estimating that the current value of the AI software market is in the region of $600 billion. So, to reach the necessary $6 trillion – just to cover its infrastructure Capex commitments – means the industry finding ten times its current revenue in just five years. For comparison, that’s roughly six times the value of the entire cloud computing sector today.

No wonder the likes of OpenAI, Anthropic, and Meta are shovelling out new apps, assistants, and platforms like there’s no tomorrow, and no wonder the likes of NVIDIA CEO Jensen Huang are exhorting us to spend, spend, spend our token budgets. Simple reality - the industry may implode if we don’t.

In the meantime, AI companies are resorting to creative accounting to fund the buildout. In September, the FT reported that US hyperscalers are having to turn to residual value guarantees to back their colossal buildout debt, “issuing up to $300 billion in commitments in less than a year,” while “recording little of that exposure on their balance sheets”.

By doing so, vendors are guaranteeing a minimum future value for their chips and data centres: a high-risk move if the market shifts to smaller AI models that can run on phones and laptops. Put another way, they are kicking their troubles into the long grass and hoping they will be awash with cash by 2031 – which is a polite way of saying they are shifting all their risk onto their investment partners.

All of which explains other Bain & Co findings from its industry snapshot. For example, “hardware and semiconductor stocks grew at a 24% compound annual rate from 2020 to 2026 vs just six percent for software”. No surprise there. After all, which would you invest in? NVIDIA is currently worth $5.5 trillion – that’s roughly four times the market cap of Bitcoin.

But the report doesn’t ignore enterprises’ existing challenges. Take productivity, for example, which everyone links to AI advantages such as vibe coding. The report warns:

Faster AI-powered coding is not automatically translating into equivalent end-to-end productivity: developers complete roughly 21% more tasks while review time rises by approximately 91%.

This proves that the Productivity Paradox and the Jevons Paradox, with which we should all now be familiar, are once again in play. AI’s promised easy wins fail to materialize in any broad, transactional sense because the technology creates a raft of new problems and costs, which eat into the process gains.

Meanwhile, the industry cliché that AI automation would free us all from boring, low-value tasks is challenged by another finding this week, this time from PwC. It notes from its own AI adoption research that:

Greater use of AI does not automatically mean an easier workload or job for employees. Almost half (45%) of UK AI users say the complexity of their role has increased, while 44% report an increase in their workload.

In other words, AI hasn’t freed us to be more creative or strategic. Instead, it has merely freed many of us to do more of the same, but with ever-compressing deadlines and rising expectations of delivery.

All of which brings us back to Bain & Co. The crushing irony of the AI era is that most enterprises rushed into AI adoption for three tactical, rather than strategic, reasons. At least, that’s according to the 100-plus enterprise reports that have accumulated on my laptop since the launch of ChatGPT four years ago, from management consultants, universities, Big Four firms, services giants, and more.

To re-cap, those reasons were: AI promised to enterprise slash costs; it promised to make organisations more productive; and everyone else was adopting it, so therefore you should too (aka FOMO). Other advantages, such as making smarter decisions from trusted enterprise data were nowhere on that list of drivers.

So, why do I describe that as a crushing irony? The clue is in another Bain & Co finding – not in the annual report, but from separate research this month in which it forecasts that organizations’ IT costs are likely to increase by a staggering 75% over the same timescale in which vendors will need to find $4.8 trillion in new revenue.

Put those figures together and suddenly everything makes sense - AI vendors are desperate for us to spend more money with them, because a multitrillion-dollar house of cards will collapse if we don’t. Just feed your tokens into the instant-answers machine, folks! (Don’t worry, you’ll hit the jackpot eventually and that’s a promise from Messrs Altman, Amodei, Musk, Zuckerberg et al)

Hence the crushing irony I mentioned above: enterprises rushed to adopt AI in the belief it would slash their costs, but it has done the exact opposite. In many organizations, AI spending is now out of control, so much so that we see reports of annual budgets being burned through in a single quarter.

Danielle Burgs Escobar is a Bain & Co Partner, and head of the firm’s UK Enterprise Technology practice. In her view, the huge uptick in enterprise IT costs demonstrates AI’s success, not its failure. She tells me:

Partially it's that people have realised AI can be effective at solving problems, and AI companies have realised they can charge for that! So even though we're seeing individual token costs go down, we're seeing a lot more consumption of those tokens, and of the layers in the technology stack that are needed to enable AI. You know, orchestration and visibility, observability. Not to mention extra cybersecurity.

So even though we're seeing the individual token costs drop, we're seeing broader adoption, broader use cases, and some of that is cyclical. As the token cost comes down, the cost per task on something that used to be way too expensive becomes achievable. As a result, you get more and more adoption, and your costs go up.

Indeed, that is a perfect expression of the Jevons Paradox. At-scale adoption ends up recreating the problem you wanted to solve by using a technology in the first place. She adds:

So, it’s almost the success of AI, showing that the cost curve will continue to increase. But ideally, that's because it's adding value elsewhere in the organization. Sometimes it's value directly into the tech function, but often it's driving efficiency in, let's say, Finance or Customer Support. But at a significant cost to the IT budget.

I put the point about the Jevons and Productivity Paradoxes to her, given that the main drivers for many AI adopters were cutting costs, increasing productivity, and FOMO, and yet survey after survey finds that – against those simple metrics, at least – AI is just not delivering yet. She counters:

I challenge that a bit because sometimes they are still getting the value. There's more usage, right? More people are using AI, more workflows are being redesigned, and more products are embedding AI, so that's driving usage too. There are heavier workflows, and more complex agentic cross-checking. So, the consumption is not just for the initial ask, but also for retries, context, and multi-step reasoning.

But then you've got model creep. Newer frontier models come out. They're awesome, so people want to use the most awesome one – regardless of whether they actually need the awesome one to solve their problem. Then unfortunately, as part of that, there's also waste.

Quite so, but this is also an expression of users being drawn into an enterprise-wide dependency on AI, one that simply didn’t exist before in most cases. She goes on:

There are pockets of heavier adopters, which recognise the cost impact and waste, and they are getting much savvier. Some of them are also asking the question: ‘Is AI the right answer for this?’ But where adoption is not so well measured, I don't think executives have caught on as much to the cost impact and the necessity to be more thoughtful.

A lot of people are on the maturity curve. And a lot of them are asking ‘How can I look at my tokens, my credits, and my requests to work out how much this is actually going to cost me?’ Not just to do this one little piece, but the whole end-to-end workflow and compute of that task. So, ‘What's my cost to complete a useful task?’ is the important question, rather than bitty, one-off kinds of queries.

An excellent point, which begs the question: do those decision-makers know what the task used to cost before they started doing it with AI? This is perhaps the most important thing to ask in the enterprise today, because if you have no idea what a task or function used to cost, then how can you possibly measure what the ROI is of doing it with AI? Escobar acknowledges:

That’s a good point! Having a baseline of what good looks like, and knowing how long this task used to take, and how accurate it was when you had humans doing it. That's essential. You need to know what your starting point is to know if what you're achieving now is better, or as good.

All of which puts me in mind of yet another enterprise AI report which landed on my desk this week, this time from an outfit called Culture Amp, which describes itself as an employee experience and performance platform.

The firm surveyed 112,000 workers in 123 organisations worldwide and found that 85% of respondents say their organisation actively encourages exploration and experimentation with AI, and yet 42% have no idea why they are being asked to use it.

What are the costs of that?