The AI Bear Case: Capable enough to attack a bank, not reliable enough to run a shop

This month, four of the biggest US banks found AI agents operating inside customer accounts. Most likely they were there because customers had setup their own agents to do things their bank didn't think to build. Meanwhile, a company called Andon Labs has spent most of 2026 letting Claude run a real shop in San Francisco. Every model they've tried has lost money, including the newest one.

So AI can now be pointed at your bank account, but it still can't run a corner shop at a profit.

This is the first of three posts. This week I'm going to make the best case I can against AI, or at least against the idea that it's about to transform business. Next week, the case for. Then after that, where myself and my co-founder John-David Wuarin actually think the truth lies. We build these systems for a living, so we're hardly neutral, but I'm going to do my best to put forward the arguments both for and against.

In this essay against AI I'm not going to argue that it doesn't work, because it obviously does. Instead, I'm going to attack AI on the economics and with a very simple question: why, three years and over a trillion dollars in, has so little turned up in anyone's profits apart from the companies selling it?

My argument, which I'll try to build up through this post, is that the problems holding AI back aren't separate. Cost, reliability, regulation, power and finance are all links in one chain. The payoff, if there is one, needs all of the links in the chain to hold at the same time. Right now, there is reasonable doubt about each and every link in that chain. Taken as a whole, the logical conclusion is that the chain will not hold — that the promise of AI transforming our lives has been severely over-sold.

Where's the money?

Torsten Slok, chief economist at Apollo, has spent this year pointing out that AI shows up everywhere except the economic data. Outside the tech sector, employment, productivity and inflation look much as they would have done without it. In a note published last week he went further. Revenue per worker and profit margins are rising at the seven biggest tech companies, but are flat across the other 493 in the S&P 500. Productivity shows no sign of an AI boost. Further, mainstream companies are starting to see token spending as a cost that eats into margins, rather than a source of growth. His conclusion is that the AI story will be decided in the 493, not the Magnificent Seven, and so far there's no evidence it's working there.

CEOs seem to agree. When Oliver Wyman asked them in May whether AI had met expectations, 27% said yes, down from 38% a year before. Yet 43% plan to cut junior roles anyway. They're restructuring for returns they haven't seen, possibly because "the street" expects them to and not conforming can be taken as a worrying sign.

Even Sam Altman of OpenAI admitted in August that he'd expected GPT-4 to shake up the software industry almost straight away, and it didn't. He put it down to people carrying on buying from the same companies and working the same way. He's almost certainly right. Businesses change at the speed of procurement, retraining and trust, and a new technology doesn't speed any of those up. IBM still sells mainframes, a server technology that most engineers would agree was technically redundant many decades ago, but procurement processes and trust keep the mainframe alive. Things in regular businesses nearly always move much, much slower than in Silicon Valley.

Cheaper tokens, bigger bills

The usual answer to the concern about the pace of change is that we need patience. Costs are falling, so the returns will come. This is the part of the sceptic case I find hardest to argue with, because we deal with it all the time.

The price of AI is falling fast. Epoch AI reckons the cost of a given level of capability halves every quarter. But we've changed how we use it. A chatbot used to answer a question and stop with a single call to the model. Now, an agent goes off and does real work in order to answer that same question: it calls tools, reads files, checks itself and tries again. That additional work has a price tag that can easily be a 10x or more increase in the number of calls to the model. The price per token falls, but the number of tokens rises much faster. The result is that invoices businesses are receiving are going up, not down.

Even the companies keenest on AI are now rationing it. Uber used up its whole 2026 AI budget in four months and capped spending at $1,500 per person per tool per month. Its COO said the link between that spend and new features "is not there yet". Accenture, which sells AI transformation, has cut central funding for Claude to $29 per person a month after some staff spent $1,000. KPMG surveyed 2,145 senior leaders and found nearly half had delayed or scaled back agent projects because they cost more to run than they were worth.

Today's prices aren't the real prices either. OpenAI spent about $34bn last year to make about $13bn. Anthropic's prospectus, filed this week, shows up to $567bn of future compute commitments, including up to $120bn with SpaceX. That's more than 30 times its revenue for the first half of the year. Yann LeCun says the labs have to cut costs and put prices up or the whole thing bursts.

People point to the labs' revenue as proof it's all working. But Anthropic's $65bn run rate is $65bn coming out of other companies' budgets. It shows businesses are spending, but whether they're getting value back is a separate question.

Some of that spending isn't buying anything new at all. AI has made attacking companies cheap, so defending them is turning into a running cost. Bank of America hired 1,000 contractors to fix flaws an Anthropic model found in its systems, and Anthropic's own security chief says things will get worse before they get better. Sam Altman went further last weekend, saying he no longer thinks the industry can patch all the bugs AI is finding. If attacks are automated with AI and never stop, defence has to be automated and never stop either, and the people best placed to sell it are the labs with the most capable models. AI-powered security defences could easily become a monthly bill rather than a one-off project, and one you can't cancel without leaving the internet. That money counts as revenue for the labs and cost for everyone else, but it doesn't make anyone more productive. It's the price of standing still.

Fine in the demo, wrong in production

Suppose the costs did come down. You'd still need agents good enough to take real work off people's hands, and they aren't there yet. When researchers test agents on the kind of work they need to do in order to replace humans, the scores are a long way off. A UC Berkeley test built by more than 300 industry experts had the best model passing 24%. On analysis of real spreadsheets and documents, the work most companies want to hand over, the best model manages 54%. Microsoft Research had models work through long document-editing tasks, the sort you might leave running overnight. By the end, even the best had quietly corrupted about a quarter of the content.

Silent errors are the real problem. Broken code usually fails and someone notices. A wrong number in a spreadsheet or a made-up citation in a report just sit there until someone notices. Judges are now regularly telling off lawyers for citing cases that don't exist, which isn't the ideal way to discover a hallucinated piece of content. Somebody has to review the output, and their time rarely makes it into the business case, partly because a lot of the time the checking never happens.

Banks don't want creative

Most of the money in enterprise IT is in banking, insurance, healthcare and government. Those sectors need a decision to come out the same way twice, to be explained afterwards, and to survive an audit. The technology industry calls this "determinism" and it's something that language models don't do out of the box.

When the model gets it wrong, it's also your problem, not the AI vendor. Meta's new business terms make advertisers solely responsible for mistakes made by its AI tools, because no technology provider would survive if it were responsible.

Look at what actually works in regulated settings and the AI part keeps getting smaller. The US state of Utah has let an AI write first-time prescriptions, but only by choosing from a short list of acne creams a doctor has already approved. The further into a regulated process you go, the more the useful system turns out to be ordinary software with a language model on the front. That's useful, but it's a long way from transforming the whole economy.

We've seen this first-hand when we built a customer due diligence system for banks. The way that banks check who their customers are uses a fixed set of rules — if the customer is a business in a higher-risk sector like gambling, they may have more stringent rules applied. Most of this work is both easier and less risky to build with code than by asking a language model what to do next. We were told frequently by domain experts that the process needed to be deterministic. We used AI for things like building a profile of the other party in a transaction. But the core of the process, deciding who does what and what has to happen next, stayed with traditional rules-based software.

We can't build it fast enough

Underneath all the predictions of an AI-powered world is an assumption that compute will keep getting cheaper because there'll keep being more of it. However, the reality is that there won't be enough, at least not on the timetable being promised. You can ship software overnight, but a data centre needs power, transformers, turbines and planning permission, and each of those takes years.

In the UK, about 140 proposed data centres want roughly 50GW between them. The whole of Great Britain's grid currently peaks at about 45GW. To meet the demand we need to double the grid capacity and that's an enormous undertaking. In Ireland, data centres already use 23% of the country's electricity. In the US, the government has used a 1935 emergency law three times this year to push data centres onto diesel generators during heatwaves so homes kept their power.

The kit is sold out as well, and so is the capacity. Andy Jassy, Amazon's CEO, says AWS is sold out for this year and 2027 is largely gone as well. Gas turbine production is booked up to 2030. Nearly half the US data centres planned for 2026 have been cancelled or delayed, mostly for want of transformers and switchgear. Micron says memory will be even tighter in 2027 and 2028 than it has been this year. Then there's local opposition: New York has a moratorium on new large data centres, and Texas, which was meant to be the US's easy option, paused grid approvals in August. Last weekend Amazon dropped a $427m data centre in Ashburn, Virginia, home to the world's densest cluster of data centres, after opposition from residents and local officials.

The workarounds say a lot. Meta is putting GPUs in tents and running them off its own gas plants.

That leaves two possibilities, and neither suits the bulls. If the demand forecasts are fantasy, a lot of this investment is wasted. If they're real and can't be met, compute stays scarce and expensive, and the cost problem gets worse.

The money only works if all of this goes right

Put those together and you get to the money. The Financial Times did the sums in May. Hyperscaler spending is growing about 20% a year and their revenue about 15%. That gap is painful for investors and they are starting to notice: last week OpenAI cut the valuation it's asking for in its latest funding round by $150bn after they pushed back.

Some of the money also goes round twice: Nvidia and other suppliers invest in their own customers, who then use the cash to buy more chips. Meta launched a $95bn share sale to pay for its AI buildout, with its own chip suppliers, Nvidia, AMD and Broadcom, offering to buy up to $5bn of it. Its shares fell 7%. TS Lombard calls this "Balance Sheet as a Service". Torsten Slok sees the same thing in token spending. Data from the card provider Ramp shows it's concentrated in a small number of venture-backed companies, while mainstream businesses don't think they need the latest models and won't pay for them. None of it tells you whether anyone outside the AI industry will pay an economic price for what it produces, and that's the only revenue that can make the sums work.

Even if the demand is real, there’s no guarantee the economics will work. Software keeps getting more efficient and consumer hardware keeps getting faster, so AI is already moving onto the device itself: both iOS and Android now ship with on-device AI. Open-weight models that run on a laptop have gone from a joke to something developers routinely run to cut token costs. If capable AI ends up effectively free, the economic case for a disruptive AI industry may start to collapse.

Why it adds up

Taken one at a time, each of these problems has an answer. Costs are falling. Models are getting more reliable. Regulators will adapt. Power stations will get built. The bulls are right that every one of them is being worked on.

The sceptic's point is that you don't get to take these points one at a time. If any one of them fails, the whole chain breaks.

For AI to earn back what's been spent on it, it has to do real work, not just answer questions. Real work means agents. Agents get through enormous numbers of tokens, so the economics depend on compute getting a lot cheaper. Compute only gets cheaper if supply keeps growing fast, and supply is stuck behind grid connections, turbine orders and planning committees. Meanwhile today's prices are being held down by investors who provide the funds that allow the AI labs to operate at a loss. But they expect to be paid back. So over the next few years, costs are more likely to rise than fall.

At the same time, agents aren't reliable enough to be left alone on long tasks. That means paying people to check their work, which eats into the savings. And they're least reliable on the work where most of the enterprise money is: judgement calls with no right answer to check against, and decisions a regulator will ask about later.

So the industry needs costs to fall while supply is short and the subsidies are running out. It needs reliability to turn up first in the sectors that are least forgiving of mistakes. It needs the labs to hold onto their margins while cheaper open models close the gap. And all of this to happen before investors lose patience with returns that, on the FT's numbers, don't add up. Any one of those could work out. Needing all of them to work out is a lot to ask.

If the sceptics are right, what happens next isn't a collapse. AI keeps improving and stays useful, especially for writing code. But prices go up as the subsidies end. Companies stop talking about transformation and settle for a handful of narrow uses where the numbers work. Some of the planned data centres never get built, and some of those that do turn out not to be needed. The investors who paid for the buildout take the losses, and everyone else ends up with a useful technology that costs more and arrives more slowly than we were told.

We've seen this before. Britain's railway mania in the 1840s left the country with a rail network and a lot of ruined investors. The dotcom bust left us with the fibre that the modern internet runs on, laid by companies that went bust. Both times the technology was real. Both times the money was mostly lost.

Next week I'll argue the other side of this, that AI really will have a big impact, as hard as I can. Read both and then make up your own mind.