Agents are getting wallets before enterprises get guardrails

The spending agent economy

Town's co-founder came on the a16z podcast with a clear thesis: the neutral AI agent — one that works for you rather than for an advertiser or a marketplace — should be funded solely by the user's wallet. The company says the breakage model used by loyalty programs like American Express and Chase is becoming obsolete under AI optimization, while Amazon has adopted an adversarial approach toward AI agents to protect its advertising business model. If the agent shops for you, the agent's incentives have to be yours. m:…9873

The pivot to this thesis was not obvious. The speaker and his co-founder walked away from their tax business after one year because they concluded it could not scale quickly enough. In his own words: "we stopped believing that we could have as much impact in the world as we ought to have because we didn't feel like we could scale the business quickly enough." A year in, they decided the business could not reach their desired level of global impact, and they pivoted away. The pivot was driven by a gap the co-founders saw: no high-quality AI agents existed to manage email and calendar workflows. The product then found its fit by focusing on common knowledge-worker workflows like meeting preparation and email drafting. m:…9873

The interview maps where the shopping economy breaks. The co-founder argued that Walmart may hold a competitive advantage over Amazon in the agent era because it relies less on advertising revenue — the revenue stream that makes Amazon fight agents rather than serve them. A neutral, wallet-funded agent routes around the ad tax; a marketplace funded by ads treats that routing as an attack. m:…9873

Enterprise-ready means you can't look

The same interview sets the enterprise bar: adoption requires audit logs, data privacy postures, and cost controls for CFOs. And the data posture is strict: "part of being enterprise ready is you can't look at user data." The wallet-funded agent and the enterprise agent are converging on the same requirement — the user has to trust who sees what. The company maintains a model-neutral strategy to avoid competing against the AI labs that supply its models, does not train on customer data, and lets customers own their business data, memories, and context. The trust stack is: don't look, don't train, don't lock in. m:…9873

LeanData's research says enterprises know they are not ready. The company says: "LeanData research found that 94% of GTM teams don't believe their infrastructure is fully ready for AI." Undocumented processes are the named culprit — 37% of GTM leaders cite them as a root cause of AI underperformance — and the company says only 21% of enterprises have mature governance for AI agents, per Deloitte. Clean, connected, complete data tells the AI what to work on, the company says; business process documentation tells it when to start, what steps to take, and when to stop. That playbook is the part most teams never wrote. The failure mode is concrete: one observer noted AI summaries built on deprecated or unused old CRM fields, and undocumented processes splinter into multiple versions of the same operating model across marketing, sales, and customer service. m:…6903

The security gap is already live

Meanwhile the deployment numbers are running ahead of the governance numbers. Cyberhaven says endpoint agentic AI adoption reached 60% of enterprises in May 2026, according to Cyberhaven Labs. The risk model is simple: agents inherit the access of the person who runs them, and "An AI agent can turn one customer file into a far more sensitive dataset without any humans in the loop." Speed plus inherited access is the whole threat, in one sentence. m:…9930

Cyberhaven's answer is an Agentic Data Security product. The company says: "Cyberhaven prevents sensitive data from reaching an unsanctioned LLM, such as one used through a personal account." And it adds: "Cyberhaven blocks data moving to and from unsanctioned MCP servers." The pattern is unmistakable: agents are being handed money, data, and workflows while the guardrails are still being built. The companies selling the guardrails are counting on that lag. The company frames the underlying tension honestly: AI agents become more effective as they gain access to business data and APIs in a secure, privacy-preserving manner. More access means more value and more risk — the product exists because that trade-off has to be managed at the data layer, not the prompt layer. m:…9930 m:…9873