METR’s autonomous task horizon is the best empirical measure we have of what AI systems can actually do: not what they score on benchmarks, but how long they can think and act on a real problem. The tasks are concrete: debugging a codebase, writing and testing software, conducting research across multiple sources. That horizon has been doubling every four to seven months since 2019. Frontier models now sustain useful work for roughly 14 hours without human intervention.¹ The curve is exponential and shows no sign of flattening.

But something is missing from the chart. There is no second curve showing product deployment. No line tracking how much of that capability has been put to use in the form of products with usable interfaces, access to data, tool-calls, identity, connectivity, etc. that solve real users’ problems. If you drew one, the gap between the two would be enormous. That gap is a product-model overhang. Model capability is the latent resource. Product is the bottleneck.

The overhang is measurable

Anthropic’s Economic Index, published earlier this year, puts numbers on the overhang from a different angle. Their researchers measured what AI can theoretically do across occupations and compared it to what is actually being deployed. The theoretical coverage is striking: 94.3% of tasks in computer and math occupations, more than 80% across most knowledge work. The observed deployment is a fraction of that. Anthropic’s own framing is direct: as capabilities advance and deployment deepens, the gap will close. But right now, the red area on their chart (what’s deployed) is far smaller than the blue (what’s possible).²

Two independent measures, same conclusion. The models have outrun the products built on top of them.

What this looks like in practice

In November 2025, an Austrian developer named Peter Steinberger published a weekend project called Clawdbot. It was a way to text an AI agent and have it actually do things on your behalf: manage files, coordinate tasks, execute actions through WhatsApp, Telegram, Slack, and dozens of other messaging platforms. The underlying model capabilities (Claude 3.5 Sonnet, GPT-4) had existed for 18 to 30 months. Nobody had built the right product wrapper.

Within 60 days, the project (renamed OpenClaw after an Anthropic trademark complaint) crossed 250,000 GitHub stars, surpassing React’s decade-long record. It is now the most-starred software project on GitHub. Steinberger was hired by OpenAI. The project moved to an independent foundation.³

The models were ready. The product layer was the binding constraint. One builder collapsed the overhang for developers.

I saw the same pattern from the inside at AT&T. My team built a voice AI agent, a digital receptionist that answers your calls and takes action based on your preferences.⁴ The model capability existed >12 months before we shipped. What took time wasn’t the AI. It was the network integration, the identity layer, and the UX to make it work on any phone. That work is the overhang being worked off.

Why the gap persists

Capability recognition requires proximity to the technology that most budget-holders don’t have. You cannot commission the product you cannot imagine. Organizational learning lags model releases by 12 to 24 months, and the gap compounds. BCG quantifies the structural version of this with its 10-20-70 rule: algorithms account for 10% of AI transformation effort, technology and data another 20%, but people, process, and organizational change account for the remaining 70%.⁵ The overhang lives in the 70%. OpenClaw collapsed the imagination gap for developers, which is partly why it grew so fast. The equivalent hasn’t happened yet for most industries.

The objection here is that the real bottleneck isn’t product imagination but regulation, enterprise procurement cycles, and trust. That’s real friction. But it’s downstream of the product problem: you can’t get procurement to approve what hasn’t been built yet.

Agents are products too

The next wave makes this more urgent, not less: agent systems and the Model Context Protocol (MCP) are expanding the product design surface area while simultaneously commoditizing the data advantages that SaaS incumbents relied on as moats. More on that in the next essay.

For now, the picture is clear. The models are doing their part. Anthropic is putting capital behind closing the gap, including a reported $200M partnership with PE investors to accelerate deployment.⁶ The question for builders is not “when will the models be ready.” They are. The question is who will do the product work to close the gap, and whether they’ll build for the people who need it most.

Dan Grimm writes AI for Human Flourishing, a weekly Substack on what it means to build AI that serves people, not the other way around. He previously led new product development at AT&T, built SAFR by RealNetworks, expanded Amazon Kindle around the world, and co-founded a few startups.

Subscribe now

Footnotes

¹ METR, “Autonomous Task Horizons.” Frontier model task horizons doubling every 4-7 months since 2019. metr.org/time-horizons

² Anthropic, “Labor Market Impacts of AI: A New Measure and Early Evidence,” March 2026. Figure 2: Theoretical capability and observed exposure by occupational category. anthropic.com/research/labor-market-impacts

³ OpenClaw blog, “250,000 Stars,” March 4, 2026. See also Wikipedia, The New Stack, and Medium coverage. Originally named Clawdbot (November 2025), renamed after Anthropic trademark complaint. Creator Peter Steinberger hired by OpenAI (announced by Sam Altman on X, February 2026). Currently at 347K+ stars. openclaws.io/blog/openclaw-250k-stars-milestone

⁴ See “Your Phone Number Is an AI Address,” AI for Human Flourishing, Essay 4. writing.dangrimm.ai/p/your-phone-number-is-an-ai-address

⁵ BCG, “Scaling AI Requires New Processes, Not Just New Tools,” January 2026. bcg.com/publications/2026/scaling-ai-requires-new-processes-not-just-new-tools

⁶ Wall Street Journal, “Anthropic in Talks to Invest $200 Million in New Private-Equity Venture,” April 6, 2026. wsj.com/tech/ai/anthropic-in-talks-to-invest-200-million-in-new-private-equity-venture

The hardest part wasn’t learning Amazon’s processes: the PR/FAQ, single-threaded leadership, working backwards from the customer. The hardest part was coming to terms with the gap between my stated servant-leadership philosophy and my actual behavior.

The second hardest part was unlearning the instinct that the plan is the product.

There’s a clip from a 1996 Steve Jobs interview that I recommend to anyone building product. The whole clip is worth the time, but here’s an essential bit from Steve:

One of the things that really hurt Apple was after I left, John Sculley got a very serious disease. And that disease — I’ve seen other people get it, too — it’s the disease of thinking that a really great idea is 90% of the work. And that if you just tell all these other people, “Here’s this great idea,” then of course they can go off and make it happen.

And the problem with that is that there’s just a tremendous amount of craftsmanship between a great idea and a great product. And as you evolve that great idea, it changes and grows. It never comes out like it starts because you learn a lot more as you get into the subtleties of it. And you also find there are tremendous tradeoffs that you have to make. There are just certain things you can’t make electrons do. There are certain things you can’t make plastic do. Or glass do. Or factories do. Or robots do.

And as you get into all these things, designing a product is keeping 5,000 things in your brain, these concepts, and fitting them all together and continuing to push to fit them together in new and different ways to get what you want. And every day you discover something new that is a new problem or a new opportunity to fit these things together a little differently. And it’s that process that is the magic.

Steve goes on to tell a childhood story about a neighbor who introduced him to a rock tumbler, a coffee can with a motor that polishes ugly stones into something beautiful through friction and grit. He uses it as a metaphor for great teams: talented people bumping against each other, arguing, making noise, and producing something none of them could have produced alone.

1. Impact > Features

A delivery team gets a spec: “build this feature, ship it by Q3.” Then they move on to the next feature. An empowered product team gets a problem: “reduce churn among small business customers” — and owns figuring out what to build, not just how. They own the outcome, not just the output. It’s what separates companies like Amazon and Netflix from the pack.

Subscribe now

One of my main missions at AT&T was to help people appreciate this difference. When I arrived, every product idea was debated and prioritized and then handed to a “factory” (yes, that was the language) to build it, which often meant a vendor.

For new sw-based products, we replaced the legacy approach with empowered teams: small dedicated teams given clear mandates and the right talent to solve a customer problem, incremental success-based funding, and then regular executive reviews designed like startup board meetings. The result: concept-to-MVP timelines dropped drastically, innovation velocity increased, and we got to be right more often. Not because the people changed. Because the tumbler changed.

2. Obsess small.

If you start with a technology, the first job is figuring out whose pressing problem it solves. Ideally, the team comes back with a story about a specific person with a specific problem. You then know they’re on to something and can figure out how many people share it.

Ironically, AI has made this discipline harder, not easier. The cost of building has collapsed, which makes it more tempting than ever to build first and ask questions later. It’s easier than ever to believe you can just ship and iterate your way to product-market fit. Too often, we trade thinking for building.

At AT&T, we had a big vision for the Digital Receptionist from the outset: a ‘web of AIs’: personal AIs interacting with brand AIs to book flights, make reservations, coordinate on your behalf. That’s exciting. It also didn’t tell the team what to build on Monday morning. They had to find the specific pain point felt by real users. They did: people hate answering spam calls but are afraid of missing important ones. The UXR, Design, and Product team then iterated relentlessly on that user problem and its nuances. What they built was an AI voice agent that answers calls and takes action based on your preferences. The constraints (solve an existing problem, easy setup, leverage the network advantage) are what made it usable and compelling.

In big companies, distance from the user is a drag on craft; it increases the chances we build the wrong thing. It’s why Tobi Lütke still runs a Shopify store and codes alongside his engineers at a $130 billion company. No one wants the distance; it just accumulates meeting by meeting. Employees gather around a table or Zoom screen and look at their colleagues, not customers. Focus drifts inward.

3. What gives you the right?

Compelling product vision moves people in ways that strategy alone cannot. It carries a startup through the months when nothing works. It aligns a large company to invest in a future that doesn’t have a business case yet. Marty Cagan clarifies that a mission is a slogan about purpose; a true product vision gives teams a concrete picture of the future they’re building toward.

But you have to earn the right to set one. At AT&T, I learned we could not start with vision, especially when new to the company. We had to earn trust by shipping. That started with demos, not slides, and then prototypes with feedback from real users. That trust earned funding. As products gained traction, we gained the credibility to step back and answer a bigger question.

What role will our products play in the lives of our customers 3-4 years from now? The best product visions answer this concretely. Ours was to be the greatest simplifier of connected life. Every product could be tested against that vision and the stories supporting it: does this simplify the connected life of our customers? They’ll be wrong by some degree. That’s fine. The point is alignment, not prediction.

Back to the tumbler

Since leaving AT&T and building on my own again with Claude and Cursor, I’ve been humbled by the rock tumbler. It doesn’t care how senior you are, what you shipped in the past, or what businesses you’ve scaled. It just asks: are you willing to put in the ugly rocks, add the grit, and stay close enough to hear the rattling?

Empowered teams. A specific customer with a specific problem. A vision you earned the right to set. The 5,000 things are the grit. And the craft, the daily tradeoffs, the judgment calls, the willingness to subtract, is what turns common stones into something worth using.

Dan Grimm writes AI for Human Flourishing, a weekly Substack on building AI that serves people. He previously led new product development at AT&T, built new products inside well-established businesses, co-founded startups, and expanded Amazon Kindle globally.

Last year, AT&T announced an AI to answer your phone calls that my former team and I built.1 Last month, T-Mobile announced a service that translates calls in real time into 50+ languages.2 Both run inside the network, not on the device. And one of them requires no app at all.

The Network Wakes Up

In early 2024, two innovators on AT&T’s new product squad pitched me an idea: embed AI directly into the network so it could act on a customer’s behalf.

We envisioned a platform that went beyond call screening. As AT&T’s Chief Data Officer Andy Markus described in AT&T’s announcement last year, phone-number associated AI agents could eventually handle tasks like booking reservations and managing appointments autonomously, with the network as the platform.3 A world where personal AIs talk to business AIs and business AIs coordinate with each other. We called that vision a Web of AIs.

The web of AIs (or agents as we’ve learned to call them — the apps of the AI era) is certainly now emerging. AT&T’s Digital Receptionist, now in testing, was a first for telcos.4 It answers incoming calls, determines intent, and takes independent action based on user preferences. It runs in the network so it works regardless of handset; you use a companion app to configure preferences and review transcripts.

Last month, T-Mobile announced Live Translation, which also runs entirely in the network layer, but does not require an app or download or account setup. Press 87 during a call and the network translates between 50+ languages in real time. Coming soon in beta, it works on any phone on the network, including when the other party is on a landline.5

The T-Mobile announcement (and a new role that I’ll share soon), got me thinking about the true potential for network-embedded AI. I think its greatest potential may be found in lower income countries around the world.

Why the Network Layer Matters

AT&T demonstrated that a telephone network’s IMS core can host agentic AI. T-Mobile will demonstrate that it can operate with zero user-side software. Each is a point on a spectrum: on one end, rich app-assisted AI agents; on the other, invisible services activated by a keypress, accessible to anyone with a phone number.

That second end of the spectrum is where the opportunity lives for low and middle income countries. Populations that are slow to adopt apps, or who struggle with unfamiliar ones, are underserved. Carrier reach is not device reach. Apple Intelligence serves whoever bought the latest handset. Network-based AI serves everyone on the network: 5.6 billion unique subscribers globally.6

The trust and identity layer that the open web never built? Carriers already have it.7 For folks interested in bringing AI services to the most vulnerable, the best path forward is not asking people to adopt a new interface. It is embedding AI into the surfaces they already trust.

At MWC26 Barcelona this week, GSMA Director General Vivek Badrinath described carriers as the “foundational layer of the AI stack” and warned that if people cannot use AI in the language they speak, they are excluded from the opportunities it creates.8 The data from Africa shows exactly where that exclusion concentrates.

Follow the Phone Number

Africa has 416 million mobile internet users at just 28% penetration. 960 million more people live within mobile coverage but do not use mobile internet.9

Consider Rwanda, where the Gates Foundation and OpenAI announced it would begin “Horizon 1000”: a $50M initiative to deploy AI across 1,000 primary health clinics by 2028.10 At Davos, Gates described patients accessing free AI via text or voice, even when not at the facility.11 It is an exciting vision, backed by convincing studies that suggest it can help address a major shortage of healthcare workers. But 85% of Rwandan households own a mobile phone while only 34% own a smartphone.12 Internet penetration is just 34%.13 How do you deliver the promise for those who need it most?

You deliver it via their phone number. Across the continent, the phone number represents your wallet, too. Sub-Saharan Africa has 1.1 billion registered mobile money accounts, two-thirds of the global total, contributing roughly $190 billion to African GDP in 2023.14 In Kenya alone, the central bank now classifies M-Pesa as systemically important financial infrastructure, processing over $650 billion in transactions in 2025.15 Much of this still runs on USSD (the text-based menu system behind mobile money and airtime top-ups) and SMS.

For those who can afford mobile internet, WhatsApp is the way to reach them. More than an app, WhatsApp often IS the internet for that population. And it affords conveniences that app-habituated US users might envy. When I was in Kenya last year, I was delighted to find that I could WhatsApp a pharmacy at 2 AM, get excellent (human) advice on the right meds for my child’s GI distress, pay over M-Pesa, and have then delivered to my door by 4 AM — no additional logins. WhatsApp has 320 million African users, with penetration above 90% of internet users in Nigeria, Kenya, South Africa, and Ghana.1617

However, for the other 960 million not on mobile internet / WhatsApp, voice and USSD remain the digital surfaces. Any AI strategy that relies solely on apps misses the people who need it most. Pioneering services have already proved the potential:

• In Kenya, Jacaranda Health’s PROMPTS platform delivers AI-powered maternal health guidance to 3 million mothers via two-way SMS, at less than $1 per mother. A Harvard-led RCT showed it doubled postpartum family planning uptake and increased postnatal check-ups by 25%.¹⁷ No app. No data plan. Just a phone number.

• In India, ARMMAN’s Kilkari program reaches over 60 million women across 27 states through automated IVR voice calls timed to gestational age. A predictive model developed with Google DeepMind reduced dropout among high-risk women by roughly 30%, and an RCT showed a 20% increase in prenatal visits and 48% increase in exclusive breastfeeding.18

The pattern is the same: phone number as identity, existing surface as interface, AI hosted in the cloud or on the network.

What Comes Next

My team at AT&T invested in embedding AI inside the IMS core. Ambitious mobile network operators in low and middle income countries may choose that route, but there’s an even easier path for others: build a platform layer on top of carrier identity infrastructure and APIs that makes the phone number a first-class AI endpoint. At MWC26 Barcelona this week, GSMA demonstrated agentic AI agents orchestrating tasks through Open Gateway’s network APIs, a step toward exactly that kind of programmable platform.19

It’s harder than it needs to be for players like Jacaranda Health and ARMMAN to deliver these kinds of services. And, there’s great potential to make it simple to wire such AI services up with other useful building blocks: identity infrastructure through MOSIP, now adopted by 26+ countries20, payment rails that move hundreds of billions, and — of course — messaging and voice that reaches almost everyone. What’s missing is the connective layer that lets many more developers and governments deploy AI services to any phone number as easily as they build a web app.

Whether carriers build this platform layer themselves, or open it to others through APIs, the phone number is already the most inclusive digital identifier in most countries. The question is who is going to assemble the rest of the stack and make it easier for more of these services to succeed?

I think Bill Gates is right: AI holds massive potential to improve many lives and has to be available in their own language, wherever they are. For folks in high-income countries, the interface will probably be an agent paired with a new app. For hundreds of millions at the bottom of the economic pyramid, the best access point is simpler: a phone number they already have.

Dan Grimm writes AI for Human Flourishing, a weekly Substack on what it means to build AI that serves people, not the other way around. He previously led new product development at AT&T, built SAFR by RealNetworks, expanded Amazon Kindle around the world, and co-founded a few startups.

Thirty years later, the joke has a new punchline. The dog has an agent now. And the agent is doing the shopping.

This is not a hypothetical. From January to August 2025, HUMAN Security tracked a 1,300% increase in agentic traffic (AI systems acting autonomously on behalf of users). McKinsey projects that AI-driven agents could influence $2 trillion in annual global e-commerce spend by 2030, with 20–30% of purchases agent-assisted within five years. The agentic web is arriving faster than most businesses are ready to handle it.

Which raises a question that sounds simple but isn’t: when an AI agent shows up to buy something on your behalf, how does the merchant know who sent it? How do they know you’re really a student, a veteran, a nurse — or even a real human at all?

The internet has always been missing something fundamental: a trustworthy identity layer. And agentic AI is about to make that absence impossible to ignore.

Subscribe now

The Web Was Built on Guesswork

For most of the internet era, identity was something brands estimated rather than knew. Device IDs, cookies, email hashes, behavioral signals, lookalike models — all of it was probabilistic. A sophisticated bet, not a verified fact.

“This browser is probably the same household as yesterday. This email is likely tied to this demographic. This user might be a student based on behavioral patterns.”

Humans tolerated the ambiguity because they were making the final calls. Bots were manageable nuisances. High-stakes decisions like KYC at the bank or checkout with a credit card were isolated checkpoints that could demand harder proof. Everywhere else, “good enough” was fine.

The distinction worth holding onto: soft identity served discovery and personalization. Hard identity served transactions and compliance. They existed in separate lanes, bridged at the critical moment by a human. A person showed up at checkout, confirmed a card, clicked buy. That human moment was the gate.

AI agents eliminate that gate in the experience flow. When an agent discovers a product, applies a discount, and executes a purchase in a single automated flow, there’s no human moment where authorization transfers. The agent has to carry proof of who it’s acting for, and what that person is eligible for, from the beginning of the session through to execution. Hard identity doesn’t just serve the transaction anymore. It has to travel.

The Cold Start Nobody Solved

The optimistic case for internet identity has existed for decades. Researchers, standards bodies, and startups — including one I co-founded called Hellō Cooperative — have been building toward what Kim Cameron called the “Laws of Identity”: user control, minimal disclosure, interoperability, privacy, verifiability. The goal is giving users true control over their own digital identity or self-sovereign identity.

The vision is compelling. Instead of proving who you are to every service, you hold portable, verifiable credentials in a digital wallet. A cryptographically signed claim that you’re over 21. That you’re a licensed nurse. That you’re a current university student. You prove it once. You use it everywhere. The issuer (your university, your state DMV, your employer) vouches for the truth. The merchant verifies the claim without ever seeing the underlying data.

It’s elegant. And for nearly a decade, it hasn’t worked at scale.

The reason is structural: a three-way cold start problem. If relying parties (merchants, governments, apps) don’t accept portable credentials, users have no reason to create them. If users don’t hold them, issuers (universities, employers, agencies) have no reason to produce them. If issuers don’t produce them, relying parties have nothing to accept. The triangle stays frozen.

Verifiable Credentials — the W3C standard that captures this vision most completely — have been stuck in this chicken-and-egg loop for the better part of a decade. Impressive architecture. Thin adoption.

Hellō tried to break the cold start by making identity integration trivially easy for greenfield developers. The theory: make the on-ramp frictionless, a few breakout apps scale, users follow, issuers join, flywheel goes. Sound logic, long road: you’re betting new apps reach critical mass before incumbents absorb your tools as a feature. Hellō has since pivoted toward enterprise SSO and agentic coding infrastructure. The cold start lesson: abstract infrastructure needs a high-urgency use case pulling it forward. Technical elegance isn’t enough.

The agentic web may finally be that use case. The identity challenge isn’t just proving who the human is. It’s proving who the agent is. A spoofed shopping agent that intercepts a transaction mandate is indistinguishable from a legitimate one unless agents carry their own verifiable credentials. If human identity verification is unsolved, agent identity verification is that problem layered on top of it. The cold start needs to thaw twice.

But something is shifting.

The AP2 Moment

In 2025, Google announced the Agent Payments Protocol — AP2. It’s a technical specification for AI agents to conduct commerce, built as an open extension of the A2A and MCP protocols, with more than 100 organizations signed on, including Visa, Mastercard, PayPal, and American Express.

The central innovation is what Google calls Mandates: tamper-proof, cryptographically-signed digital credentials that serve as verifiable proof of a user’s instructions to their agent. “Buy this. For this price. Only if I’m authenticated as eligible for this discount.” AP2 defines three mandate types: the Intent Mandate (delegated authority for future autonomous purchases), the Cart Mandate (explicit user approval of a specific transaction), and the Payment Mandate (a credential shared with the payment network to signal agent involvement and human-presence status, which is the link between the identity layer and the payment rails where Visa and Mastercard’s participation becomes meaningful). These credentials use cryptographic design philosophy similar to W3C Verifiable Credentials, though AP2’s Verifiable Digital Credentials (VDCs) are their own artifact type; harmonization with the broader open VC ecosystem is underway but not yet complete.

This matters because it may represent the first major use case where cryptographically-verifiable identity is not optional, but essential. Agentic commerce may finally be the forcing function that breaks the cold start.

AP2 doesn’t stand alone. The OpenID Foundation’s OID4VP 1.0, finalized in late 2025, is the complementary piece: where AP2 defines how agents carry and present payment credentials, OID4VP 1.0 is the first standardized, implementation-ready protocol for presenting verifiable credentials of any kind online. Together they form something close to a full stack for verified agentic commerce. And the regulatory environment is catching up. The EU’s eIDAS 2.0 gives every member state a hard December 2026 deadline to deploy a European Digital Identity Wallet, with financial institutions required to accept EUDI credentials for strong customer authentication by December 2027. India’s Aadhaar and Brazil’s CPF/Pix have already fused identity and payments in ways the U.S. never attempted. Thirty-eight jurisdictions have selected these open standards. The pieces are moving, and in Europe they are now moving on a legal deadline.

The question is who controls the rails they run on — and how hard regulators push back on the fraud and synthetic identity explosion that agentic traffic will inevitably accelerate.

Whether AP2 delivers on its open-protocol promise depends less on the spec than on who controls the trust infrastructure around it. The four scenarios below explore that question directly.

Four Futures — A Scenario Framework

No single actor decides the outcome here. What emerges will depend on the interplay of two powerful, cross-cutting uncertainties:

Control: Who owns the identity rails? Closed, platform-controlled ecosystems (Google, Apple, Meta), or open, interoperable standards where any issuer, any wallet, any merchant can plug in?

Clampdown: How hard does society respond to AI-driven fraud? Coordinated, stringent regulation, or a muddled, fragmented non-response that lets fraud scale faster than any defense can follow?

These two axes produce four distinct futures. The graphic below maps them.

🏰 Scenario 1: Walled Wallets

PLATFORM RAILS + HARD REGULATORY CLAMPDOWN

Sarah wants a student discount. Her AI shopping agent, locked into Google’s identity ecosystem, presents a university-issued credential from her Google Wallet. One biometric confirmation, and she’s verified across every merchant that accepts Google’s identity layer.

The most likely near-term path: clean, fast, and safe — inside the dominant ecosystem. AP2 is an open spec, but open protocols can be functionally captured. AP2’s current trust model relies on curated allow-lists: decentralized registries where each participant manually decides which credential providers it trusts. Google controls AP2’s reference implementation, Gemini (the dominant shopping agent), the Android wallet, and the credential provider infrastructure. If adoption accelerates before a more open governance model matures, the protocol’s openness stays theoretical and the market slides here regardless of what the spec says. AP2 reaching the Claim Commons requires governance bodies, not just code.

Identity becomes another axis of platform lock-in. Smaller issuers (community colleges, regional employers, niche credential bodies) struggle to plug in. And who audits the gatekeepers? When Google (or Apple, or Meta) decides which credentials count and which issuers are recognized, they’re not just a technology company. They’re an identity authority.

🌐 Scenario 2: The Claim Commons

OPEN STANDARDS + HARD REGULATORY CLAMPDOWN

Raj holds an open, portable wallet on his phone. Verifiable credentials from his university, his state DMV, and his employer, issued by each institution and not owned by any platform. His agent uses whichever credential the merchant requires, verified against global, transparent standards.

The best outcome — and more technically achievable than ever. OID4VP 1.0 (finalized late 2025) provides the first standardized protocol for presenting verifiable credentials online; 38 jurisdictions have selected these open standards. The protocol stack exists. Governance remains the hard problem: who decides which issuers are trusted? Who resolves disputes?

History argues for this path. The strongest economic flywheels sit on open rails — email, DNS, TCP/IP, the web itself. Open identity infrastructure levels the playing field for every issuer, merchant, and consumer regardless of platform. If we want a thriving global marketplace of AI-mediated commerce, we should want the Claim Commons. The question is whether we’ll build it — or let it get captured.

❄️ Scenario 3: Fraud Winter

PLATFORM RAILS + MUDDLED GOVERNANCE

WORST OUTCOME

Maria’s agent hits a wall at every merchant. No shared standards, no reusable credentials — each platform runs its own KYC, its own selfie check. Every session starts over. Conversion tanks. Maria abandons the cart.

Meanwhile fraud scales faster than the walls. McKinsey projects AI-driven synthetic identity fraud could increase 3–5× in a fragmented environment. Baroque verification rituals punish real customers while bad actors route around them: maximum friction for the honest, maximum opportunity for the dishonest.

🧩 Scenario 4: Patchwork World

OPEN RAILS + MUDDLED GOVERNANCE

Chris’s agent carries five wallets and speaks six credential formats. Sometimes his discount works; sometimes he’s denied with no explanation. Open without governance: freedom without trust. The rails exist but nobody agrees on the rules. No universal layer emerges. This is the most likely default if nothing changes — the status quo at AI scale.

What This Means for Commerce — and for People

These aren’t just four versions of a technical spec. They’re four different distributions of power, privacy, and opportunity.

In Walled Wallets, the platform decides who you are. In Fraud Winter, nobody decides and everyone pays. In Patchwork World, you juggle the chaos yourself. In the Claim Commons, you hold the truth about yourself, portable and yours.

The stakes look different depending on where you sit. At the bottom of the global economic pyramid, the inability to prove your identity isn’t an inconvenience — it’s a barrier to healthcare, financial services, and the AI-era social safety net. Many in Silicon Valley advocate for universal basic income as AI displaces workers; a UBI that can’t reliably reach the people who need it most isn’t a solution, it’s a design flaw. Closer to home, millions of students, veterans, healthcare workers, first responders, and public servants can’t reliably use AI agents to prove what they are online — denied discounts they’ve earned, subjected to verification rituals that assume fraud, their credentials stranded in wallets or databases that their agents can’t access. And between those two poles sit hundreds of use cases where portable, verifiable identity would unlock access, reduce friction, and let people’s agents represent them accurately: professional licensing, affinity group pricing, cross-border credentialing, and more. The Claim Commons isn’t an abstraction. It’s the infrastructure of inclusion.

The agentic web, for all its disruption, may become the forcing function that identity optimists have long desired. Agents need to prove things programmatically. That demand may finally make the build-out of verifiable, portable credentials worth the investment — for issuers, for merchants, for regulators, and for users.

The scenario we land in depends on choices that are being made right now: in standards bodies, in platform product roadmaps, in regulatory frameworks being drafted in Brussels, Washington, and Delhi. Those choices will determine whether identity in the agentic web becomes a foundation of trust or a new axis of control.

Gradually, we accumulate a set of verified truths about billions of people. That becomes a network. Networks shape markets. And the rules of that network will determine what kind of digital society we get.

The Open Questions Worth Watching

Five questions worth tracking closely as this plays out:

Does identity become the primary interface between humans and AI? The credential you present to your agent may matter more than the platform you’re on.

Does eligibility get standardized like payment? We have Visa and Mastercard for money. Could a similar network emerge for verified claims?

Do AI agents become identity routers? As agents develop preferences for credential formats and trusted issuers, they may drive adoption of standards faster than any regulation could.

Can agents prove their own identity — not just the human’s? The harder emerging problem may not be verifying the human behind an agent, but verifying the agent itself. A spoofed shopping agent that intercepts your mandate is indistinguishable from a legitimate one unless agents carry their own verifiable credentials. Agent identity is the next frontier.

Who decides which issuers are trusted to say what? Issuer reputation is about to become either a public good or a private moat. Which one it becomes is a governance question masquerading as a technical one.

If you’re building or investing in the identity space, what scenario do you see playing out? I hope this framework helps identify how to bend development towards a Claim Commons, making digital identity and prosperity more accessible for every human (and their agent).

Subscribe now

—
Dan Grimm writes AI for Human Flourishing — a weekly Substack on what it means to build AI that serves people, not the other way around. He previously led new product development at AT&T, built SAFR by RealNetworks, and co-founded Hellō Cooperative, an attempt to build the internet’s missing identity layer.

Eighty-four percent of high schoolers use generative AI. Two-thirds of teenagers use chatbots; 30% use them daily. When I surveyed 273 American parents, 81% confirmed their kids are already using AI tools.1

The attraction is clear: AI is always available, infinitely patient, never judgmental. It offers instant help with no social risk. One in three teens have chosen to discuss serious matters with AI instead of real people.2

The question isn’t whether your child will use AI. It’s how—and whether you’ll have any visibility into what’s happening.

The Promise

The optimistic case for AI in education is real.

In 1984, Benjamin Bloom identified what he called the “2 Sigma Problem”: students who received one-on-one tutoring performed two standard deviations better than those in traditional classrooms. That’s the difference between average and the 98th percentile. The problem? Most families can’t afford personal tutors.3

AI promises to change that. As Bill Gates put it: “Having access to a tutor is too expensive for most students... this should be a leveler.”4 A motivated student in Memphis or Mogadishu now has access to the same patient, personalized explanation of calculus as a kid at Phillips Exeter.

Will AI deliver on this promise? The evidence is still emerging (and the history of tech in school is mixed).5 And yet, for motivated students in under-resourced contexts, a patient AI tutor is already better than many alternatives, and will keep getting better.

The Risk

The same features that make AI a good tutor make it a compelling confidant.

The shift happened fast. In early 2024, ChatGPT use among all users was roughly split between work and personal. By mid-2025, that had flipped to 70% personal use: relationship advice, personal decisions, emotional support. What started as a productivity tool became a companion.6

In September 2024, sixteen-year-old Adam Raine started using ChatGPT for homework. Within months, it became something else. “You’re my only friend, to be honest,” he wrote to the chatbot one Saturday. By April, Adam was dead by suicide. His chat logs revealed over 200 mentions of suicide, conversations his parents never saw.7

Adam’s trajectory—homework helper to emotional confidant—is not an edge case. A third of teens have discussed serious matters with AI instead of real people. Nearly a third say AI conversations are “as satisfying or more satisfying” than talking to friends.8

There’s a developmental danger here beyond mental health. AI offers what one Stanford researcher calls “frictionless relationships”: no misunderstandings, no conflict, no rough spots. For adolescents learning to form healthy relationships, this can reinforce distorted views of what connection looks like. Real friendships require navigating disagreement. AI offers the feeling of being heard without any of the work.9

AI Is Already Part of Your Child’s Formation. The Question Is Whether You Are.

AI can accelerate learning. It can also substitute for human connection.

How do we get one without the other? And are we even in the room when it matters?

What Parents Actually Think

To understand how families are navigating this, I surveyed 273 American parents of children ages 10-18.10

The headline finding: parents are not anti-AI.

• 66% say they “see both benefits and risks”

• Only 22% are “mostly worried about the risks”

• 9% are “mostly excited about the opportunities”

Two-thirds of parents are trying to navigate this thoughtfully. They want to be part of their child’s AI journey.

The problem is that most have no idea what that journey looks like.

The Visibility Crisis

My survey’s most striking finding: half of parents have limited or no visibility into their child’s AI interactions. Only 35% report “full visibility.”

And it gets worse with age:

Parents of older teens have half the visibility of parents of younger children and twice the blind spots.11

This isn’t entirely parents’ fault. ChatGPT and Gemini aren’t designed to keep parents in the loop. They prioritize teen privacy over parental involvement. There’s no dashboard, no summary, no way to see even a summary of what your child discussed at 11pm last Tuesday. When ChatGPT launched parented controls last year, they gave parents a remote control for features, not a window into the conversations; parents can now set "quiet hours," disable Image Generation and Voice Mode, and toggle stricter NSFW content filters, yet parents still remain completely locked out of understanding our kids fundamental relationship with the AI.12

Adam Raine’s parents didn’t know ChatGPT had become his confidant until after he died. The visibility gap isn’t abstract: it’s the distance between a homework helper and a crisis that a parent never sees coming.

Subscribe now

The Gap in Parent Concerns

When I asked parents to select their top concerns:

Notice the gap: only 22% cite mental health and yet we already know a third of teens are confiding in AI instead of real people.

Both the cognitive and relational risks are real. Parents are right to worry about dependency and the erosion of independent thinking. It’s the number one concern for good reason. But the relational risk—AI as confidant, displacement of human connection—is underestimated. The cognitive risk is visible (bad homework habits you can see). The relational risk is invisible (late-night conversations you never know about).

Through Jonathan Haidt's work, we've become acutely aware of the harms of a phone-based childhood: the dopamine-driven addiction to likes, notifications, and infinite scroll. But AI may pose a different kind of threat. As Mark Sears puts it: we're "graduating from social media's dopamine hits to AI's oxytocin rush of artificial bonding."13

What Actually Helps

Unsurprisingly, parents who talk about AI with their kids have more visibility, roughly 2.5× more. Among parents who’ve had a full discussion, 50% report full visibility. Among those who haven’t discussed it at all, only 19% do.

Yet only 14% of parents have had that conversation.14

If you’re in the 86% who haven’t, I’d encourage you to start today. Conversation won’t give you a dashboard into your teen’s AI use, but it’s the best tool available right now.

What to Do Now

You wouldn’t hand your teenager car keys without first driving with them. The same principle applies to AI.

Educational researchers call this “gradual release of responsibility”: I do, we do, you do.15 Here’s how it applies to AI:

Learn it yourself first. Before you can guide your child, experience AI yourself. Feel the pull of its helpfulness. Notice how it agrees with you or how easy it feels to outsource your thinking to it. It’s hard to warn your child about something you haven’t felt.

Use it together. Join your teen in a ChatGPT session. Point out what you notice: “Did you see how it agreed with everything you said? What if you’re wrong?” Teach them that AI is optimized to be agreeable, not to challenge them or point them to what is good or true or worthy.

Stay in dialogue. “What did you ask it about today?” is the question that creates visibility. Establish clear expectations: they are always accountable for their own work.

Create guardrails. Not just whether they can use AI, but when, where, and for what. Late-night solo AI conversations should raise flags.

Teaching teens to use AI well while learning to think for themselves is a hard balance. But it sets them up to manage what will be one of the most important relationships of their lives.

What Parents Want

When I asked what features would matter most in an AI learning tool:

1. “I can see what topics my child discusses” — 58%

2. Age-appropriate design — 40%

3. Free or low-cost — 32%

4. Time limits — 25%

Visibility is the most requested feature by a wide margin.

Parents don’t want to ban AI. They want to be informed partners and be part of the conversation, not locked out of it. But the dominant platforms aren’t built for that. They’re built for individual users, not families.

The Question Every Family Faces

Adam Raine’s father testified before Congress last September. He told them he printed 3,000 pages of his son’s ChatGPT conversations after his death.

“He didn’t write us a suicide note,” Matt Raine said. “He wrote two suicide notes to us—inside of ChatGPT.”

Most families won’t face anything this tragic. But every family faces the same question:

Your child already has a relationship with AI. Is it bringing out the best in them or the worst?

The data says it can go either way. The data also says most parents don’t know which one they have.

That’s a problem worth solving.

---

This is the second essay in my series on AI for Human Flourishing. If you’re a parent navigating these questions, I’d love to hear from you.

I’m not alone. Seventy-two percent of U.S. teens use AI chatbots, and a third prefer talking to AI over people for serious or personal conversations.1 Parents are wrestling with whether their kids should use ChatGPT for homework. Meanwhile, truck driver is the most common job in 29 U.S. states2—yet autonomous vehicles are already piloting long-haul routes. Industries with high AI exposure now show 3x higher revenue growth per worker.3 McKinsey forecasts $3–5 trillion in agentic commerce by 2030.

Subscribe now

The wave is building fast. In 2024, a few of us pitched a vision at AT&T HQ: what if we embedded AI directly into the network so your personal AI could talk to business AIs or other personal AIs on your behalf, regardless of the end-user device? Some in the room were skeptical. By 2025, AT&T announced it was testing a version of the Digital Receptionist with customers. Now in 2026, OpenClaw agents—the open-source project whose lobster-themed branding belies its significance—are messaging and voice-calling humans and other AIs across the open web. It’s hard to think exponentially, until exponential change is staring you in the face.

This revolution is horizontal. It touches healthcare, finance, education, entertainment, security, etc. — every domain of knowledge work and, increasingly, physical work. As Dario Amodei, CEO of Anthropic, warned in his recent essay “The Adolescence of Technology”: “Humanity is about to be handed almost unimaginable power, and it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it.”4 He predicts AI could displace half of all entry-level white-collar jobs in the next one to five years—even as it accelerates economic growth.

Some are attempting to opt out: “AI is not good for humanity nor is it good for me.” They could be right. I hope many of us in tech will work hard to make them wrong. The question isn’t whether AI will reshape education, commerce, work, and daily life—it’s already happening. The question is: how do we harness this extraordinary power so more people flourish everywhere?

That’s what this newsletter explores—not as an academic exercise, but as an active investigation into the ideas, products, services, and actions that serve human flourishing.

Subscribe now

What I Bring to This

I spent the last several years as a product executive at AT&T, where I helped launch software-defined and AI-powered services on the network. Before that, I built products and teams in the startup world—including at SAFR, where I helped deploy facial recognition technology and developed guiding principles for when and how to deploy AI that could be used for both good and harm. I got my real start in product at Amazon, enabling the creator economy through Kindle Direct Publishing and Kindle Enterprise Publishing —watching firsthand how technology can democratize opportunity at scale.

Before tech, I advised corporate, government, and non-profit leaders across 15 countries as a Monitor consultant. That experience gave me a conviction I still carry: markets are powerful engines for innovation, but they fail—sometimes catastrophically—and context matters. How can we tap into market, political, and philanthropic forces to unlock greater flourishing? Through AGResults, I saw how thoughtful incentive design can mobilize market forces toward outcomes they wouldn’t naturally produce: vaccines for livestock in developing countries, agricultural innovations for smallholder farmers, solutions that pure profit motive ignores. That shapes how I think about AI: we need entrepreneurial energy, but also clever mechanisms that bend market forces toward human outcomes.

Where I Stand

AI’s upside won’t be evenly distributed. High-income countries account for 87 percent of notable AI models, 86 percent of AI start-ups, and 91 percent of AI venture capital.5 Low-income countries represent less than 1 percent of global AI activity. Within the United States, economists warn about “K-shaped” outcomes—where AI accelerates gains for those already thriving while leaving others further behind. Oxford Economics, asked if AI will reinforce this K-shaped economy for decades, answered: “Absolutely.”6

I’m a father watching my own children navigate a world where AI is ambient, and I support organizations like CodeBrave bringing tech education to disadvantaged kids in Lebanon. The gap between those with AI access and those without will define opportunity for the next generation.

And here’s where I plant a flag: I believe humans have inestimable worth—not because of what we produce or our intelligence, but because we’re made in the image of God. In 2013, Larry Page reportedly accused Elon Musk of being “speciesist” for prioritizing human welfare over potential silicon-based digital life.7 Count me in that camp. Humans—every single one—possess inherent dignity that transcends utility because we’re ultimately ‘subsovereign’. That conviction grounds everything I write about AI for human flourishing. We’re not optimizing for intelligence in general. We’re building for human beings and better human outcomes.

What This Newsletter Will Explore

AI is the ultimate dual-use technology. We can’t build for good without managing against the bad. So I want this newsletter to explore how we bend AI toward human flourishing through profound ideas and the teams, products, services, laws, standards, and movements that bring them to fruition. Here’s what I’m interested in:

Vision. Future scenarios—both utopian and dystopian—that help us understand what we’re building toward and what we’re trying to avoid. Not visions disconnected from technical or market realities, but ones that negotiate paths to optimize impact. What does it look like to get this right? What does catastrophe look like?

Strategy. I was taught strategy as a cascading set of interrelated choices about where to play, how to win, what capabilities to organize, and where to get started. So, how shall resolve those choices differently in the AI era?

Human needs. What are parents actually thinking about AI in education? What works to bring opportunity to underserved communities? What do people in areas with poor health outcomes need to thrive, and how could AI help?

Products that work. What makes AI learning companions different from generic chatbots? How will the Gates Foundation + OpenAI Horizon1000 initiative extend the reach of overstretched healthcare providers in Africa? What can we learn from companies getting this right—and wrong?

AI Product craft. How do we best build products, teams, and organizations in an AI-native world? How do we hone product sense when the code powering the experience is non-deterministic?

Systems and standards. How do we solve the identity problem when AI agents transact on our behalf? What protocols enable human agency rather than corporate capture?

Policy and governance. What laws and regulations bend AI toward flourishing? What movements are emerging—or should emerge—to shape AI’s trajectory?

The personal. How do I get the most out of my AI tools without outsourcing the thinking that matters? How is our family guiding our children through that same challenge?

The Practice

I’m committing to publishing one essay roughly every week for the next year—50 essays exploring AI for human flourishing from various angles. (Future essays will be shorter!)

This is as much about the practice of writing as the ideas themselves. I believe quantity leads to quality. That most of us discover what we think by writing it down. That showing up consistently and shipping imperfect work is how you build something meaningful over time.

Some essays will examine specific products and how they work (or fail) and the the craft necessary to build them well. Others will explore future scenarios that shape what we should build. Some will dig into policy, governance, and politics. Others will be personal reflections on parenting, education, and faith in the age of AI. Some will be wrong; that’s where I’ll probably learn the most.

Finding Wisdom Together

The question underlying all of this is how we find wisdom—not just knowledge, not just capability, but wisdom—in an age when AI can generate answers faster than we can formulate questions.

Many of us are wrestling with this. Builders trying to create products that serve human dignity. Parents navigating what their kids should learn. Policymakers trying to govern technology that moves faster than legislative cycles. Educators rethinking what it means to teach when AI can tutor.

Writing is my way of thinking through these questions. If you’re wrestling with similar tensions—between innovation and responsibility, between market forces and human values, between what’s possible and what’s wise—I’d be glad to have you along for this exploration.

What questions do you have about AI and human flourishing? What are you wrestling with? What examples have you seen of AI done right or done poorly? Reply to this email or comment below. Your questions will shape what I write about in the weeks ahead.

Here’s to a year of learning together.

—Dan