The moderator ran a show of hands. Who has launched an AI pilot? Nearly every hand in the room. Who has scaled one? Far fewer. Who hit unexpected problems doing it? The same hands stayed up (1:01).
That gap is what the panel was assembled to explain, and the interesting thing about the answers is where they do not point. Four executives running a healthcare manufacturer, a payments network, an energy producer and a consultancy — none of them located the obstacle in model capability, cost, or data infrastructure. All four located it in their own organisations.
Visa's eighteen wasted months
Ryan McInerney gives the most useful account of the session, and it is an account of failure.
Visa democratised access to frontier models across the company. The chief executive advocated for it from the top. The leadership team talked about it for roughly eighteen months. Nothing broke through (43:18).
What worked was different in kind. They put their top three hundred leaders in a room for two days and made them build — hands on keyboards, constructing agents, supervised and evaluated on the result (43:23). Once those three hundred had built something themselves, they had the confidence to direct their own teams, and that was the unlock.
His own summary is that he wishes he had known it earlier. The lesson generalises further than he states it: executive advocacy for a tool the executives cannot operate produces enthusiasm without capability, and the two are easy to confuse from the top of an organisation.
What Philips is actually measuring
Roy Jakobs offers the most concrete mechanism, and it is deliberately not a productivity number.
A nurse spends fifteen to twenty minutes of every hour on administrative work (3:24). Ambient listening in a patient room removes the note-taking: the encounter is captured, the clinician reviews and signs off, and ten to fifteen of those minutes come back.
The framing he insists on is that this is time returned to the practice rather than cost removed from it (3:53). His supporting figure is the one that makes the case: patients currently receive somewhere between three and seven minutes of a clinician's attention. Against that baseline, ten minutes recovered is not an efficiency gain — it is a different quality of encounter.
Julie Sweet extends the point with a pharmaceutical example where the compression is less interesting than the behavioural change. Regulatory content approval that took months moved much faster. But the observable shift was that people who had spent their time asking how to get something approved began asking who needed it and whether it helped (5:32) — because when revision is cheap, the question changes from clearance to usefulness.
Her survey figure supports the reframing: 78 per cent of senior executives now believe AI contributes more to growth than to productivity (6:12).
Aramco on why chips are not the constraint
Amin Nasser's contribution is the least quotable and the most operationally specific.
His flat rejection is of the assumption that acquiring compute produces value: you cannot create value by buying chips and installing them without the data quality and the trained people to use them (38:16). His formulation — you cannot scale without scaling the talent — is offered as a constraint rather than a sentiment.
He is also precise about where the value is not. Finance, translation, legal — the functions most organisations automate first — are not where the returns sit. The returns are in operations (39:40), which for Aramco means a capital programme of fifty to sixty billion dollars annually with a hundred billion under construction at any time (40:06). Applied there, a small percentage improvement dwarfs anything achievable in back-office functions.
The requirement he draws from this is governance rather than technology: an operating model with fast decisions about what to kill and what to scale, including during a pilot rather than after it.
The formulation worth keeping
Asked what they wish they had known earlier, Sweet answers in one line: human in the lead, not human in the loop (42:13).
The distinction is sharper than it first sounds. Human in the loop describes a person positioned as a checkpoint inside an automated process — reviewing, approving, catching errors. Human in the lead describes a person directing the work, with the system as the instrument. The first arrangement makes people slower versions of a validator. The second is what Visa's three hundred leaders became after two days of building.
Nasser's version is adjacent and organisational: it must be a business pull, not an AI or technology push (42:32). Where the business is not involved from day one, he says, you can buy the capability and even scale it, without ever capturing value across the enterprise.
What this suggests about the trillion and a half
The session opened with the figure: roughly 1.5 trillion dollars invested during 2025 (0:15).
Set that against four executives who agree the binding constraint is leadership capability and operating model, and the implication is uncomfortable for anyone allocating the next tranche. Almost none of that spending goes to the constraint. It goes to compute, models and platforms — layers that were already sufficient in every case described here.
Visa's eighteen months are the cleanest illustration. Nothing about the technology changed between month eighteen and month twenty. What changed was that three hundred people spent two days building something. That intervention costs almost nothing relative to the infrastructure line, and on this panel's evidence it is the one that decided the outcome.
关键数据
- $1.5 trillion
- invested in AI during 2025, per the session's framing 0:15
- 15-20 minutes per hour
- nurse time spent on administrative work, of which ambient capture returns 10 to 15 3:24
- 78%
- of senior executives who believe AI contributes more to growth than to productivity 6:12
- 18 months
- of executive advocacy at Visa that produced no breakthrough 43:18
- $50-60bn a year
- Aramco's capital programme, where Nasser locates the real returns from AI 40:06
演讲章节
关键要点
- 01
Visa democratised model access and advocated from the top for eighteen months with no breakthrough, until three hundred senior leaders spent two days building agents under supervision. 43:23
- 02
Jakobs measures returned clinician time against the three to seven minutes of attention a patient currently receives, not against cost — which changes what the intervention is for. 3:53
- 03
A nurse spends fifteen to twenty minutes of each hour on administration, of which ambient capture in the room returns ten to fifteen. 3:24
- 04
After compliance content became cheap to revise, people stopped asking how to get approval and started asking who needed the material and whether it helped. 5:32
- 05
Accenture's survey puts 78 per cent of senior executives behind the view that AI contributes more to growth than to productivity. 6:12
- 06
Nasser rejects the assumption that buying chips creates value: without data quality and trained people, the compute produces nothing. 38:16
- 07
He locates returns in operations rather than in finance, translation or legal — the functions most organisations automate first. 39:40
- 08
Sweet's formulation for what she wishes she had known: human in the lead, not human in the loop. 42:13
- 09
Nasser's organisational condition is that adoption must be a business pull rather than a technology push, with the business involved from day one. 42:32
提及的实体
相关演讲
Huang brings a diagram to Davos: AI as a five-layer cake running energy, chips, cloud, models, applications — with economic benefit landing at the top and every layer below it a precondition. His argument for why this is a genuine platform shift rather than a product cycle is the strongest part, and it does not rest on his commercial position: software was pre-recorded and worked on structured data, whereas a machine that reasons about unstructured input and inferred intent makes previously impossible applications possible. What the framing accomplishes is worth noticing separately. By presenting the layers as a chain rather than a portfolio, it converts infrastructure spending from a bet into a prerequisite, and the question of proportion between layer-two spending and layer-five value stops being askable. Read against the GTC keynote two months later, the same business gets two framings: one a case for choosing his product, the other a case for the category existing at the scale he needs.
A year after their first joint appearance, the heads of Anthropic and Google DeepMind returned to a shared stage and disagreed mainly about speed. Amodei held to a horizon of one to two years for systems that outperform humans across most cognitive work, resting the claim on a self-improvement loop that runs through code; Hassabis kept to the end of the decade, arguing that verifiable domains like coding and mathematics automate far earlier than natural science, and that the capacity to pose a new question rather than answer an existing one is still missing. The exchange is most useful where they converge: both accept the loop is the variable that decides everything, both are sceptical of doomerism without dismissing the risk, and both want more time than the competitive dynamic allows. Amodei's chip-export argument and Hassabis's call for minimum international safety standards are the two concrete policy asks.
Harari spends most of his address establishing terms before asking the question he came to ask. His preliminary work is to dismantle the word tool: a knife's use is decided by whoever holds it, whereas what is arriving decides for itself, and can also invent new kinds of knives. From there he argues that anything constituted by words — law, books, text-centred religion — is exposed, while drawing a firm line at feeling, where he says there is no evidence at all. The structural claim is that the ancient tension between letter and spirit has always run inside humanity and is about to be externalised between humans and the new masters of words. Only then does he arrive at personhood, and his handling is precise: corporations, New Zealand rivers and Indian deities hold legal personhood safely because the decisions are made by humans behind the container. An entity that decides for itself ends that arrangement. He does not answer the question; he tells the room it is coming.
The most useful sentence is a diagnosis of how clients are getting it wrong: they see a cost opportunity rather than a revenue one. That framing decides everything downstream, because an organisation treating AI as cost reduction measures success by what disappears and arrives at a smaller version of what it already was. The market number offered — seventeen per cent index growth in 2025 that nobody foresaw — is deployed against bubble anxiety and argues in both directions, since unforecast gains are poor evidence for confidence in any current consensus. The panel then locates the real source of instability outside technology entirely, in three geopolitical situations generating enough uncertainty to dominate planning, which is a corrective worth taking seriously at a technology conference. Their closing question about thriving across multiple futures is correct and sits uneasily beside the cost framing they diagnosed at the start.
The distinction in this session that deserves to travel is a two-word change: the model is moving from an attention economy to an attachment economy. That changes what is measured and what regulation would have to address, because attention competes for time while attachment competes for relationship, and the two produce different products from identical technology. Attention is finite in a way people notice; attachment produces reliance that feels like preference, which makes it harder to regulate for the same reason it is harder to notice. The supporting argument is about incentives rather than intent: earlier engagement produces more data and longer relationships. The regulatory proposal — measuring well-being outcomes rather than asking for safety by design — identifies the right target without solving the measurement problem that made regulators settle for a floor in the first place.
The claim worth arguing about here is that the native currency for AI agents will be crypto, because agents will not carry cards and blockchain is the interface most native to them. It comes from someone with an obvious interest in it being true, which is a reason to examine it rather than dismiss it. The strong part is structural: card networks assume a cardholder who can be contacted and can attest to a transaction, and an autonomous process breaks each of those assumptions. What does not follow is the conclusion, because nothing prevents existing networks issuing delegated credentials with limits and revocation. The panel's most direct voice calls these areas highly speculative with hard use cases, and both positions can hold, since stablecoins and speculative assets are separable in a way the panel treats as one thing. The quieter claim about tokenised government instruments is the more consequential one.