Trust and Governance could turn AI into a powerful force for decentralization, community, and value creation.

This is a big picture blog: my hypothesis on the forces shaping how AI and commerce will evolve. It builds on two posts I return to often. The New Economy: Small Wins (2016) argued for an economy where millions of small businesses thrive. Part 3: Internet 2.5, The Next Wave is Here (2022) argued that trust is the core of networks and commerce, how it changes is the key shaping force in the “next wave”. AI is now testing both ideas at once.

Technology Is Not the Bottleneck

Most of the AI conversation is about technology. The quest for electrical power and chips to run the physical engines. The alphabet soup of standards: MCP, UCP, ACP, AP2, TAP, x402. Model efficacy, benchmarks and hallucination rates. These problems are real, and capital is pouring in to solve them. My focus is elsewhere: commercial terms, trust, customer experience, data and privacy. IMHO, trust is by far the most powerful force shaping AI right now.

Consumers are saying so directly. Accenture’s 2026 Consumer Pulse, a survey of more than 25,000 consumers, found only 9% open to fully autonomous shopping (Business Insider, Sep 2026). The 2026 Edelman Trust Barometer describes a world retreating into insularity. People are shrinking the circle of who they trust. Better models will not fix that.

Bruce Schneier makes the same point: the promise of personal AI assistants “rests on a dangerous assumption: that we can trust systems we haven’t made trustworthy” (Building Trustworthy AI Agents, Dec 2025). As I wrote this week in Agentic Governance: Standards are worthless without a Commercial Construct, a standard needs a cop. Technology can carry trust. It cannot create it.

Trust Shaped Language

In Data Tipping Point (2018) and again in Wallets, APIs and Trust (2023), I defined trust this way:

“Trust” is an amorphous concept that represents an entity’s current and historical ability to operate in the context of a contract, process or relationship.

Individually, we interact and converse based upon our level of trust and the objective of the conversation. The need to communicate differently to different audiences shaped our language. 4,000 years ago, if a stranger asked “where is the best hunting ground?” you would not point to your favorite spot. You might say “where the water meets the grass and there is a place to hide.” The ability to throttle (and add color to) information allows the owner of information to control its flow and restrict its context (ie how it can be used). Richness of language permitted information to flow with different levels of trust in each interaction.

Today’s data architecture has none of this richness. Data is either shared or it is not, and once it leaves your control its use is unbounded (see Consumer Data Bureau?). An agent that acts for me needs the hunter’s discretion: what to say, to whom, for what purpose and at what price. Doc Searls, who coined the “intention economy” more than a decade ago, now expects that “we will indeed use AI agents to express our intentions (including the one to be left alone)” (The First Source of Personal Intent, Jun 2026).

The Small World Search

In Wallets, APIs and Trust I referenced MIT’s Solid Project. A search for the best red wine among your friends who are wine enthusiasts will likely beat the same search on Google. In Internet 2.5 I described one of my patents for this kind of decentralized search: group members bookmark the wines they love, a search for a Priorat returns your group’s curated results, and the member who created the winning result earns a penny. The challenge in this “small world” search across billions of data silos is threefold:

1. An organizing force to link them
2. A trust model to govern them
3. An economic model to share them

The inventor of the web is still on this path. His company Inrupt now has Charlie, a Solid based agent that works for the user, drawing on personal data to answer while protecting that data from platforms (The New Yorker, Oct 2025). The technology exists. The organizing force, trust model and economics remain open.

What PageRank Left Behind

Google’s success in search was a measurement innovation. PageRank measured links and clicks to rank relevance, and in doing so it indexed the internet. As I noted in Federated Models Need Measurement, Google started out as a measurement company. But small data sets were left untouched. Privacy, competition and uncertain ownership kept them out. Data owners could not trust, or control, how their data would be used. The public web was indexed. The private web of preferences, skills, inventory and relationships was not.

In Small Wins I wrote that web services failed between companies for a commercial reason, not a technical one: “there was no path to price a service based upon value.” Universal directories like UDDI died because no one had an incentive to publish anything beyond public data. MCP faces the same fate in the long tail unless it comes with a commercial construct.

Retailers Are Playing the Same Game

Top retailers learned this lesson quickly. They want to control their own proprietary data and deliver the value of AI within their own domain. Walmart’s Sparky is the clearest example. Walmart reported that shoppers using Sparky placed orders about 35% larger than those who did not (Digital Commerce 360, May 2026). Rather than cede the journey to third party chatbots, Walmart embedded Sparky inside ChatGPT and Gemini (Wing VC). Amazon took the “walled garden” approach and blocked outside agents (Amazon vs Walmart: Two Very Different Bets on Agentic Commerce). OpenAI shelved Instant Checkout after colliding with the merchant’s role in risk, data and CX (Explaining the Death of OpenAI’s Instant Checkout). I covered the data battle in Agentic: Intent and the New Data Games.

If you are OpenAI or Shopify, your priority is to get top retailers’ inventory and skills integrated into your model via MCP. Shopify moved first at scale. Its Agentic Storefronts let merchants set up their data once and surface it in ChatGPT, Copilot, Google AI Mode and Gemini (Shopify, Mar 2026). That solves catalog syndication. It does not solve trust, risk or economics for a merchant with no brand power.

The numbers are large. There are over 1 million retail establishments in the US (US Commerce data) and, per the National Restaurant Association, more than 1 million restaurant and foodservice outlets. How do you integrate these? As I noted in Agentic Commerce: Killing Small Retailers?, platforms take the path of least resistance. They integrate the top 10 first.

Every Person Has Something to Offer

The “commerce” opportunity is far greater than just the 1-2 M retailers. Every person has skills and assets to bring to a need, a “job to be done” in Clayton Christensen’s language (see the milkshake problem in Walmart and Open AI).

As a car guy, imagine me describing the problem of fixing my son’s 2006 Audi A4. A general model would give me excellent recommendations on diagnosis and repair. But what if someone who solved this exact problem last week lived down the street and could help? What if I needed a specialized tool and could borrow it from a stranger, because an intermediary managed the risk? Or how could local farmers take on a direct to consumer (DTC) role? In 2022, 116,617 US farms sold directly to consumers for $3.3 billion in sales (USDA Census of Agriculture). That is a rounding error in US food spending, and exactly the kind of fragmented supply AI could organize.

None of these are products in a catalog. They are capabilities held by small entities whose trust is local. In Reputation: Commerce Implications (2013) I described how eBay learned this in China: reputation has community context, and merging the communities sent fraud through the roof.

Can Demand Be Orchestrated Differently?

Taking this a step further in my hypothetical future: why do we have cars that only last about 10 years? Why not a standard, interchangeable base structure with well defined component interfaces (ie like a PC)? Of course, the answer is that cars aren’t sold that way. Canoo built its vehicles on a modular skateboard chassis designed to simplify repair, and it filed for Chapter 7 in January 2025 (TechCrunch). The supply side could not finance the idea.

But if this is a supply problem, can demand be orchestrated differently? Small Wins made the case that modularity is how small works: PCs succeeded because standardization let investors trust a collection of interdependent companies to deliver a finished product. Agents that aggregate intent (say, ten thousand owners who want a 20 year chassis) could change the calculus for suppliers. In Data Tipping Point I asked the same question about bread: “How could a farmer trust in the demand for a local flour?”

Web3 Had the Tech, Not the Trust

So much of today’s internet has led to centralization, powered by the virtuous cycle of data. Web3 was about decentralization. It had the technology right, but trust and economics were not there. As I wrote in Internet 2.5, “No commercially successful community in history operated anonymously.” In Why Closed Chains Work Best for Banks I made the related point that trust in banking is a commercial construct, not a consensus algorithm. Industries change when commercial agreements are rewritten or trust is reallocated (Tokenization and Trust).

Can we fix trust and economics, and enable the trusted interaction of very small entities to decentralize and federate interactions that improve community?

Economists are asking a version of the same question. In The Coasean Singularity? (NBER, 2025), Shahidi, Rusak, Manning, Fradkin and Horton argue that agents lower the costs of preference elicitation, contract enforcement and identity verification, which expands the feasible set of market designs. That is Coase and Williamson, whose transaction cost economics anchored Small Wins and Agentic Commerce Economics and Governance. Lower transaction costs are what make small work.

Big Agents for the Generic, Small Agents for the Trusted

From my perspective, agentic AI that attempts to take on Amazon or Walmart is a pipe dream. Even if successful, the cost savings will be de minimis. The real opportunity is around the edge (Agentic’s Real Economic Opportunity: Edge Use Cases).

In this future world, big monolithic agentic platforms will do well in the generic. Small specialized models will be best for trusted interactions and specialized knowledge. NVIDIA researchers make the technical version of this argument in Small Language Models are the Future of Agentic AI: most agent work is “a small number of specialized tasks” done repetitively, which favors smaller models. Citi GPS described the same fork, with agentic systems built from “smaller, specialized language models” on one path and foundation model developers embedding agents on the other (Citi GPS, “Agentic AI: Finance & the ‘Do It For Me’ Economy,” Jan 2025). Google Cloud (GC) enables a world of specialist model partnerships rather than one central brain (Strategic Innovation Era: Part 1). Although Google Gemini is looking to be everything to everyone (and maintain the ad business).

The technical case and the trust case point the same way. A specialist that keeps its knowledge local can participate without surrendering it.

The Consumer Control Panel

To succeed, consumers need a control panel that shows what data they have permissioned to which agent, for which use and at which price:

• My wine list goes free to my friends.
• My automotive tool lending has a price.
• My healthcare data goes to defined endpoints with significant compliance.

This is the “right to have, right to use, right to share” framework that Visa’s former Chief Privacy Officer Russ Schrader taught me, applied to agents (Consumer Data Bureau?). It is the hunter’s discretion rebuilt in software.

On the consumer side, my view is that Apple is best positioned: the most privacy focused company in consumer technology, the best container, with the most personal data and the best CX. Apple’s new Siri runs on device and in Private Cloud Compute (Apple, 2026), and Apple says personal data is used to complete a request rather than to train or build profiles (SiliconANGLE). That is the outline of an orchestrator that decides which agent has which rights to which data. In September Apple updated its iOS 27 privacy policy to let users opt in to improving its models with their Siri interactions (AppleInsider, Sep 2026). I see this as consistent with the thesis, not a departure from it. A personal agent that knows you has to learn from you. If the end state is a model that lives on your phone and is your property, then training it on your own interactions, with your explicit consent, is exactly the kind of permissioned data use I am arguing for. I described this role in Scenario: Agentic Wallets and Federated Data: wallets will not become the agent, but they will be the most important interface to all agents. See also WWDC 2026: Apple Wallet.

Scaling Trust in the Long Tail

The challenge in going after edge use cases is that they are the hardest to tackle. How do small data sets add value? AI has the potential to solve this problem. Where Web3 was about decentralizing the storage of data, what we need is a framework for scaling trust in the long tail.

eBay did it with reputation. Resnick and Zeckhauser’s controlled experiment found that established sellers earned a measurable premium over new sellers for identical goods (The Value of Reputation on eBay). Payments did it with operating rules. In Federated Models Need Measurement I outlined what a specialist needs to participate: access control, pricing, commercial terms and measurement.

Who will be the enablers?

• Supply side: the last mile platforms from Internet 2.5 (Toast, Shopify, Square) already hold the merchant relationship, the catalog and the payment. They can certify small suppliers and carry risk. Individuals have no equivalent yet, and lending a tool or fixing a neighbor’s porch raises insurance and liability questions that no agent protocol answers.
• Demand side: the wallet and device (Apple, Google) that hold consumer identity and permissions.
• In between: a neutral party that writes the rules, measures contribution and enforces penalties. Today the card networks are the only entities with that commercial construct across all parties (Retail, Roles and Know Your Agent).

Where I Could Be Wrong

• Aggregation may win again. Ben Thompson argues that with agents “the ability to do stuff is becoming abundant; what is scarce is volition,” and whoever solves inspiration gains power over every supplier (Stratechery, 2026). Goldman sees agents expanding the long tail by surfacing smaller merchants that best fit a need, but warns that AI platforms will gain pricing power and turn lower acquisition costs into “a new distribution toll” (Eric Sheridan et al., Goldman Sachs, “Entering the Agentic AI Era for the Commerce Landscape,” Sep 2026).
• The edge is getting hurt first. My informal survey of small retailers found traffic down 50 to 70% since the rise of AI search (Agentic Commerce: Killing Small Retailers?). Edge players must survive the short term.
• Peer tool lending has been tried before. Tracxn counts 329 peer to peer rental startups, including Peerby and Fat Llama, and only 5 have raised Series A or beyond (Tracxn, Sep 2026). A tool lending market lives or dies on density in a single neighborhood. AI lowers transaction costs. It does not create local liquidity. My bet is that agents help here because the demand already exists as intent (a stated need), so supply can be found on request rather than listed in advance.
• Best positioned is not the same as executing. Apple has the container and the privacy credibility. The open question is pace: how quickly it opens agent permissions to third parties.

What to Watch

• Whether Apple can create the dashboard which other agents report to
• Whether a platform offers risk underwriting for individuals (the tool lender, the weekend handyman), not just merchants
• Whether networks extend operating rules to agent attribution and specialist payment
• Small model adoption inside vertical platforms
• Consumer willingness to delegate moving from 9% toward a majority for bounded tasks

Small Wins

Per my blog on Small Wins, a platform that makes progress here will not be competing with Amazon. The demand will be completely different. The opportunity to decentralize AI and value creation is immense. Creating trust is hard.

Thoughts appreciated.