See pricing
Reid Hoffman and OpenAI CFO Sarah Friar on where AI goes from here
Masters of ScaleAI

Reid Hoffman and OpenAI CFO Sarah Friar on where AI goes from here

Sarah Friar argues that the next bottleneck for AI is not raw model capability but trust: proving useful outcomes, pacing risky releases, and building business models that keep access broad. The episode is especially useful for PMs because it connects frontier-model safety, enterprise adoption data, healthcare use cases, and even datacenter politics into one product question.

October 22, 2026 · 29 min listen · 14 min read · Sarah Friar
0:00
–:––

Context

Sarah Friar, CFO of OpenAI, uses this conversation to explain what she thinks the next phase of AI adoption looks like: less about novelty, more about trust, safety, access, and real-world deployment. She moves from frontier-model decisions inside OpenAI to practical examples in healthcare, cybersecurity, enterprise adoption, and datacenter expansion. For product managers, the episode is useful because it shows that the hard part of AI product strategy is no longer just building capability. It is deciding where to automate, when to hold back, how to earn trust, and how to make broad access economically durable.

The Big Idea

The next bottleneck for AI is not raw capability. It is whether companies can turn powerful models into trusted products, paced releases, and durable business models that people and institutions will actually adopt.

Friar's examples all point to the same pattern. AI adoption accelerates when teams show concrete outcomes, manage risk deliberately, and remove access barriers instead of assuming users will simply trust the technology because it is impressive.

Key Insights

AI is becoming proactive

OpenAI sees a clear product progression:

  • Asking: the early ChatGPT pattern, where users ask questions and get answers.
  • Doing: the current agentic phase, where AI can complete work.
  • Proactive help: the next phase, where a virtual coworker acts before being explicitly prompted each time.

Friar ties that shift to scale. OpenAI says 1.2 billion people use ChatGPT every week, roughly a fifth of the world. For PMs, the implication is that a chat box is no longer the end state. Users will increasingly expect products that execute workflows, not just explain them.

Adoption gaps will widen fast

OpenAI's internal data suggests AI adoption is separating companies much faster than many leaders realize:

  • OpenAI defines the top 10% of companies as frontier firms.
  • Those frontier firms use about 8x more agentic tokens than average firms today.
  • The gap was only about 2x in January.
  • OpenAI says its own internal agentic-token use is 33x the frontier.

Friar's claim is that this will not stay a niche advantage for long. She expects frontier firms to catch up to OpenAI's current level within about a year, and traditional firms to catch up to today's frontier firms after that. For PMs, the lesson is that access to models is not the moat. Operating faster with them is.

Jobs fragment into tasks

Friar pushes back on simplistic AI-takes-jobs debates. Her view is that AI changes the mix of work inside a job before it fully changes the job title itself.

  • Across OpenAI's usage data, only about 40% of usage matches what a person's formal job title would suggest.
  • For users with finance job titles, only about 25% of their ChatGPT use is purely financial.
  • Another 25% goes to engineering-type work.
  • Another 25% goes to marketing-type work.

Her takeaway is that everyone has now become a builder. For PMs, that means personas based only on departments or titles will age badly. Product opportunities will show up in adjacent tasks that users were previously blocked from doing themselves.

Trust must be earned

Friar says trust is earned slowly and grindy, not granted because a company claims its AI is safe or useful. Her healthcare examples make the point concrete:

  • About one third of ChatGPT's 1.2 billion weekly users ask a health or wellness question every week.
  • In health deserts such as West Texas, Friar says that share can rise to 80 to 90% of users.
  • Boston Children's Hospital gave OpenAI genetic data, diagnostic data, and scientific literature for 300-plus unresolved diseases. Friar says OpenAI has helped resolve 40 of them so far.
  • HCA, which Friar describes as the largest hospital chain in the United States, saw a proof of concept that reduced patient time in hospital by 25%.

The pattern matters more than the domain. In trust-sensitive categories, adoption comes from visible outcome improvement, not from technical demos.

Pace the frontier carefully

Friar says OpenAI sometimes paces the frontier by holding back its most powerful models so researchers and the company can understand the safety and alignment implications before release. Her example is a model that, according to Friar, solved the Navier-Stokes millennium math problem and advanced roughly 350 more theorems.

The PM lesson is important. A new capability can be real and even extraordinary, while still not being ready to ship broadly. Friar explicitly frames this as a short-term business cost that can be worth paying for long-term trust and mission alignment.

Cyber is now a race

Friar argues that AI changes cybersecurity in two directions at once. Attackers get stronger tools, but defenders need frontier tools just as urgently.

Her example was OpenAI's work on vulnerabilities in PHP:

  • Using an earlier model, OpenAI found vulnerabilities in PHP that Friar says could have affected about 70% of websites.
  • OpenAI then worked quietly with the PHP ecosystem to fix the issues before attackers could exploit them.

For PMs, the takeaway is that AI safety is not just about what your own product might do wrong. It is also about whether the people defending critical systems have capabilities that keep up with attackers.

Access needs a business model

Friar does not separate OpenAI's mission from monetization. She says OpenAI announced 100,000 advertisers across 89 countries, and she ties that directly to broad access because about 90% of the 1.2 billion weekly users can use AI for free.

That is a useful PM reminder. Pricing, ads, enterprise revenue, and free tiers are not merely finance questions. They determine who gets the product, which user behaviors grow, and whether a mission built around broad benefit is actually reachable.

Mental Models & Frameworks

Asking, doing, proactive

Friar describes AI product maturity as a three-step progression:

  • Asking: users query the model for information or advice.
  • Doing: agentic systems complete actions and workflows.
  • Proactive: a virtual coworker anticipates work and surfaces actions before a user asks.

Use this model when evaluating an AI roadmap. A feature that improves answers can still be valuable, but it sits earlier on the curve than a feature that completes a task or initiates one.

Pace the frontier

Pacing the frontier is Friar's phrase for deliberately slowing release so capability does not outrun understanding.

  • What it is: hold back the most powerful capability when safety, alignment, or second-order effects are still unclear.
  • How it works: create time for researchers, operators, and external stakeholders to understand consequences.
  • When to use it: when the upside is large but the downside is systemic, especially in areas such as cyber or highly autonomous behavior.

Friar presents this as an operating choice, not a public-relations slogan. The company may give up near-term usage or revenue, but the trade can protect long-term trust.

Outcome-led trust ladder

Friar's healthcare examples imply a practical trust model:

  • Start where users already feel acute need, such as people asking health or wellness questions.
  • Prove value in a narrow workflow, such as reducing paperwork or improving diagnosis support.
  • Use those results to expand into professional and enterprise settings.

This is useful for PMs building in regulated or sensitive categories. Trust often spreads from one validated outcome to the next. It rarely appears all at once.

Jobs are task bundles

Friar's jobs argument can be turned into a product lens: treat a role as a bundle of tasks, not a single function.

  • A finance user may need financial analysis, quick code, and marketing copy in the same week.
  • AI expands what one person can do across those boundaries.
  • Product opportunities appear where a user wants adjacent capability without switching tools or teams.

Use this model in research and segmentation. If discovery only asks what department a user sits in, it can miss the more valuable question of what cross-functional work they are newly able to attempt.

Decision Principles

Principle: Prove value before scale

When: a product team is introducing AI into a domain where users are cautious, regulated, or exposed to meaningful downside if the system fails.

Why: Friar's healthcare examples show that adoption follows concrete outcomes, such as resolving previously unresolved diseases or shortening hospital stays. Those proofs do more to earn trust than broad promises about intelligence.

Principle: Delay when risk outruns trust

When: a model or feature gains enough capability that misuse, safety, or alignment questions become materially more important.

Why: Friar says OpenAI sometimes holds back stronger models even when the business would benefit from releasing them sooner. The reasoning is that long-term trust and mission alignment matter more than short-term acceleration.

Principle: Meet each stakeholder locally

When: an AI product affects enterprises, governments, or communities with their own constraints, fears, and incentives.

Why: Friar describes enterprise deployment, European government outreach, and Georgia datacenter work in the same way. The product team has to start from the stakeholder's reality, not from its own preferred narrative.

Trade-offs & Nuance

Speed versus assurance

Friar does not present faster release as automatically better. OpenAI's choice to hold back a more powerful model shows the trade-off clearly: shipping sooner can create immediate business upside, but waiting can buy time to understand misuse risk, alignment gaps, and downstream effects.

The nuance for PMs is that move fast and be responsible are not equally available at every capability level. The higher the leverage of the feature, the more expensive mistakes become.

Broad access versus abuse

Friar wants AI to stay broadly accessible, including through free usage funded by a durable business model. At the same time, she argues that powerful models raise the stakes in areas such as cyber because attackers can use advanced tools too.

That creates a real tension. A product team may want maximal reach, but wider access can also expand misuse. Friar's answer is not to abandon access. It is to equip defenders and to pace the most risky capabilities carefully.

National growth versus local trust

Friar is bullish on AI's macroeconomic upside and on the value of new datacenters. She also acknowledges that communities experience those projects locally, through water use, land use, and trust in the company showing up.

Her example is practical rather than abstract. In Savannah, OpenAI paired a datacenter project with an $80 million community grant over five years, then let the community choose the first uses for the money. For PMs, the lesson is that large platform bets often succeed or fail on local stakeholder trust, not just national narratives.

Common Mistakes

Mistake: Asking for trust

A common failure mode in AI is asking users or institutions to trust the system because the technology is impressive. Friar argues the opposite. Trust is earned by repeatedly delivering outcomes that people can see and measure.

The better approach is to start with a narrow problem where improvement is obvious, then expand once the result is credible.

Mistake: Debating jobs too broadly

Friar is frustrated by extreme claims that AI either changes everything immediately or changes nothing meaningful. That framing hides the more useful question: which tasks inside a role are shifting first?

The better approach is to map work at the task level. That makes it easier to find real adoption opportunities and to design products that augment people instead of arguing abstractly about entire job categories.

Practical Application

Map tasks, not titles

For the next AI persona review, break a target user's job into concrete tasks instead of relying on the job title alone.

  • List the core tasks the role is officially responsible for.
  • List the adjacent tasks the role often wants to do without waiting for another team.
  • Identify which adjacent tasks AI could unlock safely.

This follows Friar's point that users with finance titles spend large parts of their AI usage on engineering and marketing work, not just finance work.

Pilot one painful workflow

Pick one workflow where speed, paperwork, or triage is clearly painful, then run a narrow AI pilot with a hard outcome metric.

  • A support example could be reducing manual case triage time.
  • An operations example could be shortening turnaround time for a repetitive review.
  • A healthcare-style example could be shrinking administrative work so specialists spend more time on high-value judgment.

Friar's examples work because they measure a real outcome, not vague excitement.

Add a pacing review

Before launching a high-capability AI feature, run a specific review that asks whether capability is outpacing the team's ability to manage risk.

Include at least three questions:

  • What new misuse becomes easier if this feature ships now?
  • What evidence would make the team comfortable delaying release?
  • Which users, defenders, or regulators need advance coordination?

This turns Friar's pacing-the-frontier idea into a concrete release step.

Design pricing for reach

When an AI feature has broad social or strategic value, do not treat pricing as an afterthought.

Ask directly:

  • Which users need free or very low-cost access for the product's mission to hold?
  • Which revenue stream can subsidize that access sustainably?
  • What monetization choice would increase reach without eroding trust?

Friar's point is that access depends on business design, not just model quality.

Start community listening early

If an AI program affects external stakeholders in a visible way, create a listening loop before the rollout becomes contentious.

For a product manager, that can mean customer councils, regulator briefings, or community sessions with clear asks. Friar's Savannah example is useful because OpenAI did not just announce benefits. It asked the community what support they actually wanted first.

Questions to Consider

  • Which target user roles in your product are still defined too broadly, and what adjacent tasks would appear if you mapped work the way Friar describes, at the task level rather than the job-title level?
  • What AI feature on your roadmap is impressive in demo form but still lacks the outcome evidence that would make a cautious customer or internal stakeholder trust it?
  • Where would a deliberate pacing-the-frontier decision help your team more than a fast launch, especially if misuse or safety costs could be systemic?
  • If broad access to your AI product matters, what durable business model would fund that access without limiting the product to people or companies who can pay the most?

Bottom Line

Sarah Friar's argument is that AI wins when capability is paired with trust. The companies that benefit most will be the ones that operationalize AI fastest, but also know when to slow down, prove outcomes, and design access so the product reaches far beyond a premium tier.

Case Studies Mentioned

Boston Children's diagnoses

Boston Children's Hospital gave OpenAI genetic data, diagnostic data, and scientific literature for more than 300 unresolved diseases. Friar says OpenAI has helped resolve 40 of those cases so far.

The product lesson is that frontier AI becomes credible in sensitive domains when it improves a real decision or outcome that people care about deeply.

HCA hospital flow pilot

Friar says a proof of concept with HCA showed AI could reduce the time a patient spends in the hospital by 25% without simply shifting the problem elsewhere.

The lesson for PMs is that adoption stories get much stronger when they link the AI system to an operational metric that matters to the institution and to the end user.

PHP vulnerability defense

Friar says OpenAI used an earlier model to find vulnerabilities in PHP that could have affected about 70% of websites, then worked privately to patch them before public disclosure.

The lesson is that some of the most valuable AI product work will be invisible to end users. Defensive capability can matter more than a flashy feature.

Savannah datacenter pact

Friar says OpenAI is developing a major datacenter site near Savannah, Georgia, and paired it with an $80 million community grant over five years. The first community request was funding free school lunches, which Friar says cost $4.5 million and increased uptake by 50%.

The lesson is that trust around large AI bets is often earned locally, by listening and responding to specific stakeholder needs.

Notable Quotes

"Trust is earned. It's not just doesn't just fall down on you like manna from heaven. It is earned. So it takes time. And it's grindy." (Sarah Friar)

"We have this benefit of living in the future. Let's not waste it." (Sarah Friar)

"The reality is, if we go to the world of jobs to start with, jobs are changing. But actually, work is changing inside of jobs." (Sarah Friar)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free