Somewhere in the past year, the question about AI changed. It used to be "what can it do?" Now it is "what did we get for it?".
Gartner's 2026 Hype Cycle for Generative AI puts the technology in the Trough of Disillusionment. The pilots-everywhere energy of 2024 and 2025 has met its first serious budget-review cycle, and every week brings another skeptical headline: abandoned enterprise projects, bubble warnings, studies suggesting the productivity gains were imaginary.
Read enough of these and they share a denominator, and it is not "the models aren't smart enough." It is that nobody can actually prove the return - not the enterprises, not the investors funding the buildout, not the developers using the tools every day. And they can't prove it for a dull reason: almost nobody decided, before reaching for AI, what it was supposed to achieve. That is the argument here, up front: the fix for the ROI crisis is not smarter AI. It is deciding what you want from the tool before you adopt it.
Enterprises can't show the returns
The number everyone quotes is from MIT's Project NANDA. Its 2025 report, The GenAI Divide: State of AI in Business, found that despite $30-40 billion of enterprise investment, 95% of organizations were getting zero return - most integrated pilots stuck with no measurable impact on profit. It is not an outlier: S&P Global found the share of companies abandoning most AI initiatives before production jumped to 42% in 2025, from 17% a year earlier, and in PwC's 29th Global CEO Survey, 56% of 4,454 CEOs said AI had produced no significant financial benefit in the past year. Only about one in eight reported both higher revenue and lower costs.
In almost none of these failed pilots can anyone say why they failed, because almost nobody wrote down, before starting, what success would look like. When the criteria are set only after the fact, success cannot be declared even when the technology works. The pilot does not fail against a benchmark; it simply ends without a verdict.
An irony worth sitting with: the most-quoted proof that AI doesn't pay off is itself hard to verify. The NANDA report rests on 52 interviews and 153 survey responses, with the authors' own caveat that the figures are "directionally accurate based on individual interviews rather than official company reporting." Even the evidence for the measurement crisis has a measurement problem.
The buildout runs on circular claims
The spending on the supply side has gone vertical. The four big cloud giants - Microsoft, Google, Amazon, Meta - have guided to combined 2026 capex that analysts put at around $700 billion, up roughly 77% in a year. What is the evidence it will pay off? Increasingly, demand from inside the same small circle doing the spending. Man Group's analysis of the AI bubble describes megacaps that "act simultaneously as suppliers, customers, investors, and validators" of one another: chip makers invest in AI labs, which commit to buying compute from the cloud giants, which buy the chips. Revenue looks spectacular because each node pays the next, and "the demand signal becomes circular and divorced from the market."
This is no longer a fringe worry. The Bank for International Settlements devoted part of its 2026 Annual Economic Report to it, warning that the cloud giants are spending faster than they earn, that the borrowing behind it is "typically poorly disclosed," and that a sudden pullback could turn "the capex boom into a protracted investment bust." Its comparison is the 1840s railway mania. Central banks do not reach for that for fun.
Strip away the market drama and the issue is familiar: the claims justifying the biggest capital-spending spree in tech history are supplied by the parties who benefit from them, and outsiders have no independent way to check.
Even the developers can't tell
You would expect the ground truth to be clearest closest to the work. It isn't. In mid-2025, METR ran a randomized trial with experienced open-source developers on their own repositories: 16 developers, 246 real tasks, AI allowed or disallowed per task. They forecast AI would speed them up 24%; afterward they estimated about 20%. The measured result: AI made them 19% slower. Not slightly overestimated. Directionally wrong, with confidence.
The follow-up is stranger. When METR re-ran it in early 2026, the data came back unreliable, partly because 30-50% of developers refused to do some tasks without AI. You cannot run a controlled comparison when participants won't accept the control condition. Self-reported productivity, it turns out, is not a measurement. It is a feeling, and a lot of the AI economy's accounting is built on it.
Three symptoms, one disease
| Who's claiming | What's claimed | Who supplies the evidence | Can anyone check it? |
|---|---|---|---|
| Vendors and pilots | "The deployment is working" | Vendor dashboards, usage stats | No |
| Cloud giants and labs | "The demand justifies the capex" | The same circle of companies | Barely |
| Developers | "I'm about 20% faster" | Self-perception | Measured: no |
All three are the same problem in different clothes: a gap between what is claimed and what anyone can check. Every layer of evidence - vendor dashboards, self-reported speedups, circular purchase commitments - is supplied by a party with a stake in the answer, and none of it can be checked independently. That is not a technology failing; the models keep getting better. It is an accounting failing. The industry built extraordinary machinery for producing intelligence and almost none for saying, in advance, what would count as a win.
Then the agents started spending money
AI systems are moving from talking to transacting. Over the past year, real payment rails for autonomous agents have shipped: Coinbase's x402 revived the dormant HTTP 402 "Payment Required" code into a working standard - an agent requests a resource, the server quotes a price, the agent pays in stablecoins, no accounts or human in the loop - and Google answered with AP2, the Agent Payments Protocol, launched with sixty-plus partners including Mastercard, PayPal, and Amex.
A measurement problem does not shrink when the system acts autonomously. It explodes. No human eyeballs each step, no expense report, no invoice a controller can question. And we know how this goes, because it already happened: Chainalysis counted over 100 million x402 transactions on Base within three quarters of launch, but when Artemis went through the volume it called the boom "still mostly a mirage" - roughly half the transactions were self-dealing or wash trading. A meme token called PING generated over 150,000 "agentic payments" in its first month; on one representative day, about 131,000 transactions amounted to roughly $28,000 of real volume. "Trust the dashboard" was a weak answer for enterprise pilots. For autonomous agents it is no answer at all.
The boring fix: decide the outcome first
Line up the three failures and the common thread is not the technology. It is that AI kept getting adopted because it was AI, not because someone had defined a problem it was meant to solve. The fix is unglamorous and older than AI: decide the outcome first, in numbers, and write it down before you start. What should move, by how much, by when? Pick a target a hostile reviewer would accept, and record it somewhere you cannot quietly revise once the results are in.
Only then does the tool question matter. Measure the work against that target, not in place of it - an agent that made 10,000 API calls has only proven it made 10,000 API calls, not that the calls were worth making. The teams getting real returns are rarely running smarter models than everyone else. They decided what "done" meant before they wrote the check, so they can actually tell whether they got there.
None of this needs new infrastructure. When the work is autonomous - agents spending on their own, with no human watching each step - keeping the record where it cannot be quietly edited helps, and an open ledger is one honest way to do that. But that is plumbing, not the point. A tamper-proof record of the wrong goal is still a perfectly documented failure. The record removes the excuse; it does not do the homework.
The disillusionment is the good news
The trough gets read as a verdict on the technology. It reads better as the hype burning off to show where the tools actually work. The same stretch that produced METR's 19%-slower result also produced hard evidence of the opposite: GitHub's controlled trial had developers finish a real task 55% faster with an AI assistant, and McKinsey found the time for a new business to reach revenue fell from 38 months in 2023 to 31 in 2025. The gains are real, and they land with the teams that decided what they wanted before they reached for the tool.
We are one of those teams. We lean on AI heavily to build our own software - Boar Finance went from idea to production with a small team on short timelines - and we ship AI products of our own, including the MCP endpoints that let agents read blockchains behind Boar MCP. Both serve real users today, not a pilot deck. The speedup is not a forecast for us; it is how we already work.
So that is the note to end on. The Trough of Disillusionment is not where AI dies; it is where the noise clears and the teams building real things on top of it pull ahead. The models are good and getting better, the tooling keeps maturing, and the people shipping real products with it - us included - are moving faster than we ever have.
AKENA is a blockchain engineering studio. We build AI and blockchain infrastructure - agent-facing RPC and MCP endpoints, on-chain data pipelines, and the products that run on top of them. If you are building with AI and want to move faster, we should talk.