The Simons Foundation just put $15 million on the table with Anthropic and OpenAI, and the stated goal is not better models. It is standards. As foundation president David Spergel put it, the mathematics community “should be driving discussions of how these tools are evaluated, adopted and used, not any single company or institution.” Mathematicians saw where this was heading and decided to write the verification rules themselves, before a vendor wrote them first.

You face the same choice on the factory floor. When an AI model flags a defect, approves a batch, or drafts a CAPA, somebody has to define what counts as verified. If you leave that definition to your software supplier, you inherit their logic and their blind spots. Here is how to set AI adoption standards your auditors, your operators, and your customers will actually accept.

When AI Produces an Answer Nobody Can Verify

Mathematics has a head start on this problem because it built the tooling first. The Simons Foundation funded the Lean proof assistant specifically so a machine-generated proof could be checked line by line rather than trusted on reputation. Even with that, the foundation is now paying for workshops that “help researchers evaluate recent AI-assisted results.” The verification layer exists and the field still admits it cannot keep pace.

Quality has no Lean. An AI vision system rejects a part, a model scores a supplier as high-risk, a tool drafts a root-cause narrative, and the inspector reviewing it has no mechanism to test the reasoning. They can agree or override. That is not verification, that is a coin flip with extra steps.

The people closest to the work are rarely the ones setting AI adoption standards. That is backwards.

What the Simons Foundation, Anthropic and OpenAI Actually Committed

Where the money goes: workshops, evaluation and shared standards

The structure matters more than the headline figure. The Simons Foundation manages and awards the funding itself, including through an open call for proposals. Anthropic and OpenAI put money in. They do not decide where it lands.

What it buys is deliberately unglamorous: workshops and initiatives that let researchers evaluate recent AI-assisted results, learn and use new tools, share perspectives and shape collective standards for AI use. No new model. No benchmark leaderboard. The spend goes to the human process of agreeing what counts as checked work, which is the part most organisations skip entirely.

The compute-and-tools side of the deal, and why it reaches past mathematics

The second half of the commitment is in-kind. Anthropic and OpenAI supply AI tools and computing resources to researchers the foundation supports, and that access extends beyond mathematics into basic science. Vendors give capability, the community retains the authority to judge it.

That split is the transferable lesson for anyone setting AI adoption standards in a plant. Your vendors will happily supply the model, the inference capacity, and a validation protocol written by their own team. Accept the first two. Write the third yourself, with your quality engineers, line supervisors and auditors in the room.

The awards and resources will be community-driven and leverage the foundation’s decades-long dedication to supporting and championing mathematics.

Community-driven is the operative phrase. Spergel is explicit that the standards should not come from “any single company or institution.” Replace “community” with “your operations and quality function” and you have a working governance model that survives an audit.

Why Spergel Insists the Community, Not a Vendor, Sets the Rules

Spergel’s point is about authorship. Whoever writes the acceptance criteria decides what counts as a correct answer, and a vendor writing those criteria will set them exactly where their model performs. That is not malice. It is the natural outcome of asking the seller to define the standard.

On the plant floor this shows up in the details nobody reads before signing. The vendor defines the confidence threshold for a defect call. The vendor defines which failure modes go in the validation set. The vendor defines what counts as drift and how often it gets checked. Three months later, your quality system is running on somebody else’s definition of good enough, and your auditor will ask you to defend it.

Practitioner-led rollout inverts the order. Your quality engineers write the acceptance criteria before the demo, based on the defects that actually escape to customers, not the ones that photograph well. You supply the validation images, including the ambiguous ones your inspectors argue about. You set the review threshold and the escalation path. The vendor then meets that bar or loses the deal.

Notice how the Simons Foundation structured this. Anthropic and OpenAI provide funding, tools and computing resources. The foundation manages the awards and the community shapes the standards for AI use.

That separation is the whole design. Same split works in manufacturing: buy the model from the vendor, keep the AI adoption standards inside your own quality organisation. The operators, inspectors and process engineers who know where your process actually fails are the only people qualified to say when the output is trustworthy.

The Lean Lesson: Verification Infrastructure Beats Faster Output

Lean did not make mathematicians faster. It made their claims checkable. That is why the Simons Foundation calls it a tool that “revolutionized how mathematicians work,” and why a proof assistant, not a model, is the piece of infrastructure the field now builds on. Checkability came first. Speed came after.

Most plants do it backwards. The AI use case ships, output volume jumps, and the verification layer gets scheduled for a later phase that never arrives. Then a customer complaint lands and nobody can reconstruct how a decision was made.

From trust to verification: what checkable output looks like in a quality system

Checkable means an auditor can take any single AI-influenced decision and walk it backwards without calling the vendor. Model version, input data, confidence score, threshold in force that day, who reviewed it, what they changed.

If that record lives in a vendor dashboard you cannot export, you do not have an audit trail. You have a screenshot. Log decisions into your own QMS, in your own schema, from day one.

Building your own acceptance criteria before you scale a use case

Write the pass condition before the pilot starts, not after you see results. Pick the failure modes that actually hurt you, build a held-back test set from your own historical nonconformances, and define the miss rate you will accept in writing.

Then keep that set permanently. Re-run it every time the model updates, and treat a drop below your threshold as a stop condition, not a discussion point. Plants that set these AI adoption standards early scale one use case at a time and defend every one of them. Plants that skip it scale volume they cannot explain.

Borrowing the Playbook: A Practitioner-Led AI Standard for Your Operation

The foundation’s model is copyable at plant scale. Awards are “community-driven,” meaning the people doing the work decide what good looks like. Your equivalent is the inspectors, process engineers and quality technicians who already know which defects get missed and which calls are borderline. They write the acceptance criteria, not procurement and not the supplier’s solutions engineer.

A 90-day sequence: evaluate, document, then scale

Start with evaluation, not deployment. Pull 200 to 300 historical cases with known-good dispositions, run the tool against them blind, and have your own people score the disagreements. You learn more from twenty contested calls than from a vendor demo.

Then document. Write a one-page internal standard that names three zones: where AI output is accepted outright, where it triggers human review, and where it is rejected regardless of confidence score. Publish it, date it, and version it like any controlled document. Keep tool choice reversible by holding your evaluation set and your criteria independent of any single platform, so swapping vendors costs weeks instead of restarting the program.

The ROI case: audit readiness and reclaimed review hours

Rework loops shrink first. Clear accept/review/reject zones stop the ping-pong between quality and production over who owns a disputed call, and borderline parts get routed once instead of three times.

Audit readiness is the harder number to ignore. When a customer or notified body asks how an AI-assisted disposition was made, you hand over a dated standard and an evaluation record rather than scheduling a week of reconstruction. Adoption also survives staff changes, because the criteria live in a document instead of in the head of whoever championed the pilot.

The Next 18 Months: Professions That Write Their Own AI Rules Win

Mathematics got there first because it had a funder willing to pay for the boring part. Manufacturing has no equivalent. No foundation is going to bankroll workshops where your inspectors argue about acceptance thresholds for vision models. That work gets done inside your plant or it does not get done.

Auditors are already asking. Notified bodies, automotive customers and pharma clients are starting to ask how AI-influenced decisions are validated, who owns the override, and what evidence exists when the model gets it wrong. An answer written by your supplier’s compliance team is not an answer. It is a document you will be asked to defend and cannot.

Two outcomes are forming. Organisations that documented their own verification criteria will hand auditors a file and move on. Organisations that waited will adopt whatever standard a regulator or a dominant vendor settles on, then retrofit two years of production decisions to match it. Retrofitting is always more expensive than writing it the first time.

The first move this quarter costs almost nothing. Pick the one AI-assisted decision in your operation with the highest consequence, usually a release or disposition call. Write down, on a single page, what evidence would convince a customer that decision was correct. Then check whether you can produce that evidence today.

Most teams cannot, and that gap is the whole project. Spergel framed the ambition for mathematics as treating the field “as a collective understanding driven by human curiosity” rather than a pile of outputs. Quality works the same way. The output is a judgement your organisation stands behind, and AI adoption standards exist so that judgement survives scrutiny.

Source: simonsfoundation.org