Sixteen Nobel Laureates Say Act Now. Product Teams Already Are

Every AI feature they ship already answers, in miniature, the question the statement poses at the level of the economy: does this complement people, or replace them? And that answer lands on someone's budget and someone's risk profile long before any policy does — usually whoever owns the roadmap and the AI spend. AMS AI, an AI-native procurement platform for the $1.5T healthcare supply chain, shows what "complement" looks like as a design decision, not a policy goal.

The gap between advice and action

On 13 July 2026, Stanford's Digital Economy Lab released "We Must Act Now: A Statement on AI's Transformation of the Economy" — four sentences, signed by sixteen Nobel laureates and over 200 economists and AI researchers, including people from Anthropic, Google, and OpenAI.

The statement itself is short enough to read twice:

Ajay Agrawal put the stakes plainly: institutions can't keep relying on "institutional scaffolding that was optimised for a pre-high-fidelity-prediction world." Whether AI concentrates wealth or lifts living standards broadly, he argues, depends on choices being made now — not once the transformation has already arrived.

That's a claim about governments, regulators, universities. But the same logic runs downward. The decisions a team embeds in an AI product — what it does, and what it deliberately doesn't do — are governance decisions too. Someone owns the risk the moment that product ships, and that someone is usually whoever controls the roadmap and the AI budget. Just a much smaller and much faster institution than the ones the statement is addressed to.

Institutions move in years. Products ship in weeks.

Anton Korinek, a Stanford economist currently on leave at Anthropic, framed the timeline problem sharply:

A few years is a long runway for a product roadmap. It's a short one for tax policy, labour law, or a national retraining programme. The statement's authors know this — it's why the letter reads as urgent rather than descriptive. But the asymmetry cuts the other way too. While policymakers debate what guardrails should exist, product teams are shipping the thing the guardrails are meant to govern. And the decisions baked into those products now don't stay neutral: they harden into a compliance liability later, when regulation finally arrives and applies — often retroactively — to what teams already put in the field. Every sprint is quietly setting the terms a future guardrail will judge.

That's not a criticism of the statement. It's a gap the statement itself names but can't close. Someone has to close it in the meantime. It's usually whoever is designing the product.

The real choice: complement or imitate

Erik Brynjolfsson's line from the Stanford release is the one worth sitting with: the goal is "to guide AI to complement humans rather than simply imitate them." At the level of an economy, that's a policy direction. At the level of a product, it's a design decision — made feature by feature, not once but constantly.

Imitation is the easier path. An AI that mimics what a person would do — writes the email a person would write, picks the vendor a person would pick — feels impressive in a demo. It's also the version most likely to be judged purely on whether it's "as good as" a human, and to displace the person doing that job today.

Complement is harder to build and easier to defend. And "defend" is concrete: easier to defend to the board, where accountability and ROI are clear; to the regulator, because human oversight is built in rather than bolted on; and to the customer, who can trust that a person still owns the decision. That difference also shows up in the portfolio. Complement features tend to carry a more predictable ROI and a lower write-off risk, while imitation projects more often stall on "not quite as good as a human" and burn budget without ever reaching production. So complement versus imitate isn't only a design call — it's a portfolio-prioritisation call. It means the AI extends what a person can see, decide, or catch early — without quietly taking the decision away from them. That distinction doesn't show up in an economic model. It shows up in the feature list.

Three questions that tell you which side you're on

"Complement, not imitate" is easy to agree with and hard to apply. Here's a test that works on any AI feature already in your roadmap:

1. Who makes the final call? If the AI decides and the person finds out afterwards, that's imitation. If the AI surfaces what the person needs to see before they decide, that's complement.
2. What happens when it's wrong? Complement systems make an error visible and cheap to catch. Imitation systems bury the error inside an answer that looks right enough that nobody checks it.
3. Does the feature replace a search or a decision? AI that replaces the search — filtering, flagging, ranking, warning early — sits on the complement side. AI that replaces the decision itself — approving the claim, picking the vendor — sits on the imitation side, even when it's technically more accurate than the person it's replacing.

The same test travels beyond your own roadmap. It works as a filter for prioritising an AI portfolio, and for scoring vendor and third-party AI features before you buy them — not just for the features you build in-house. None of these questions require an economist. They require someone willing to look honestly at what a feature actually does once it ships.

What that choice looks like in practice

AMS AI, a platform we designed for the $1.5 trillion healthcare procurement market, is worth reading not as a product pitch but as an illustration of what "complement" means when it stops being a slogan and becomes a product decision.

Procurement teams already know roughly what they need. What they lack is time and visibility — into better-priced alternatives, into looming shortages, into whether their own spending is out of line with peer hospitals. The team behind AMS AI didn't automate procurement decisions. It automated the discovery of context those decisions depend on: shortage alerts up to 14 days in advance, benchmarking against peer hospitals, sourcing suggestions surfaced inside the actual buying workflow instead of a separate search step. The procurement officer still makes the call — the system just makes sure they see what they need before they make it.

That design choice is the point. The metrics follow from it, and they're what tell you the enabler approach actually pays: every capability was tested against one question — does this reduce time, cost, or risk? — and if the answer wasn't clear, the feature didn't ship. That filter is why the product targets 20-plus hours saved per person weekly and an 85 percent cut in logistics time, rather than a fully autonomous buyer. The savings are the evidence; keeping the human in the decision is the design.

This isn't a one-off. The same pattern — automate the discovery, not the decision — recurs across our work.

Run AMS through the three questions above and the pattern holds: the procurement officer makes the final call, not the system. A missed alert is a visible gap to fix, not a hidden wrong answer. And the feature set replaces the search for options — not the decision to buy. Nobody at Stanford would call that AI policy. It is, in miniature, exactly the choice the statement is asking institutions to get right.

The institutions we're waiting for are us

The Stanford statement is addressed to economists, policymakers, and technology leaders — asking them to build the guardrails before the transformation outruns them. That's the right ask, and it's not a fast one. Guardrails at the level of an economy take hearings, legislation, international coordination.

Products don't wait for that. They ship this quarter. Every team building with AI right now is already answering the question the statement is trying to get institutions to ask — whether they've framed it that way or not. The Nobel laureates are right that someone needs to act now. For product teams — and for the CTOs, CIOs, and every leader accountable for what those products are allowed to decide — "now" was already the deadline.

Key Takeaways

• 𐩒Sixteen Nobel laureates and 200+ economists signed a statement warning AI could transform the economy faster than the Industrial Revolution, urging institutions to build guardrails now.
• 𐩒Institutions have years to adapt; products ship in weeks — so the decisions teams bake into AI today harden into compliance exposure tomorrow, when regulation catches up to what already shipped.
• 𐩒"Complement versus imitate" isn't only a design choice — it's a defensibility choice: complement features are easier to defend to the board, the regulator, and the customer, and tend to carry more predictable ROI and lower write-off risk.
• 𐩒Three questions test which side a feature is on — who makes the final call, what happens when it's wrong, and whether it replaces a search or a decision — and the same test works as a filter for an AI portfolio and for evaluating vendors.
• 𐩒AMS AI treats AI as an enabler, not the decision-maker: every feature was filtered through "does this reduce time, cost, or risk," with savings (20+ hours per person weekly) as evidence the approach works, not as the goal.
• 𐩒The guardrails the statement calls for at an economic level already exist, in miniature, in every product decision about what AI is allowed to do without a human — which makes "now" the deadline for everyone accountable for AI decisions, not just product teams.