The Fiduciary Wedge: Why You Still Need a Company

Suppose the cost of running a company really does collapse. Coordination gets cheap because agents talk to agents; execution gets cheap because building the feature is now cheaper than holding the meeting about it. If you take that seriously, the next question is whether you need the company at all — or whether a few people and a swarm of agents can transact directly with the market and skip the corporate overhead.

The answer is no, and the reason is narrow and load-bearing. There is one thing the agents cannot do for you: be accountable. That residual — the gap between what an autonomous system can execute and what a human or legal entity must answer for — is what the framing I’m borrowing here calls the fiduciary wedge. I took the term from Peter Diamandis’s Organizational Singularity episode, where the company is “a purpose container, a fiduciary legal container, a liability container.” That’s the seed. The rest of this is about whether the wedge holds up under the law as it stands in 2026, and where it’s thinner than it looks.

The wedge is not a metaphor — it’s the default rule #

Start with the hard constraint. AI systems have no legal personhood in any jurisdiction. They are classified as property, so they cannot sign a contract, cannot be sued, and cannot be held liable for anything (SCBC Law). An agent is, legally, a very capable tool — and tools don’t bear risk. Someone behind the tool does.

That someone is the company. Under English corporate law and analogous US doctrine, liability stays with the legal person whether a decision was made by a human or a machine; the non-delegation principle means you cannot offload the legal consequences of a decision onto the thing that made it (St Andrews Law Review). This is not a policy anyone has to enact. It’s the default. Automate the whole pipeline, and the corporate entity is still standing there as the named party when something goes wrong. That standing-there is the wedge.

The cleanest illustration is Moffatt v. Air Canada (February 14, 2024). Air Canada’s website chatbot told a grieving customer the wrong thing about bereavement fares. When the customer sued, Air Canada ran the “we don’t need the org for this” argument in reverse — it claimed the chatbot was “a separate entity” and the airline wasn’t on the hook for what it said. The BC Civil Resolution Tribunal threw that out, finding Air Canada “had failed to exercise reasonable care to ensure the information’s accuracy” and ordering damages (American Bar Association). The amount was trivial — CAN$812 — but the precedent is the point: AI autonomy is not a liability escape hatch. Mobley v. Workday pushed the same logic onto vendors, with a federal judge applying agency theory to an AI hiring tool in 2024 and a class certified in May 2025 covering rejected applicants over 40. The courts keep converging on one instinct: when the AI causes harm, reach for the legal person behind it.

Regulators are encoding the wedge into statute #

If the case law establishes the wedge by default, regulators are now hard-coding it. Article 14 of the EU AI Act legally mandates that high-risk AI systems be designed so a human overseer can understand the system’s limits, detect anomalies, override outputs, and halt the thing (EU AI Act Article 14). High-risk provisions become fully applicable August 2, 2026. This is the state writing “a human must remain accountable” directly into the design requirements — not as governance advice but as a condition of deployment.

The same expectation is hardening in financial regulation, though here you have to be careful who actually said what. Regulators and practitioners increasingly treat AI accountability as non-delegable: ESMA has stated that financial institutions “must take full responsibility for the actions of AI systems they deploy” (a Reuters-reported characterization, not a formal ruling), and practitioners summarizing the regime — the law firm Cullen and Dykman, not the SEC itself — put it bluntly: “delegating decisions to a machine does not absolve the human fiduciary from oversight” (Venable LLP, Dec 2025). The SEC’s own contribution has been narrower and proposal-stage. The attributions matter because the wedge’s strength depends on whether oversight is genuinely compelled or merely expected.

Boards inherit this personally. Risk Management Magazine argues that “passively deferring to an algorithmic decision may violate the duty of loyalty because such abdication constitutes bad faith,” exposing individual directors to personal liability (RMMagazine, Feb 2025). This is the underrated half of the wedge: the corporate wrapper doesn’t only shield stakeholders from the AI, it concentrates accountability onto identifiable humans who can’t rubber-stamp their way out of it. The container has names attached.

Why the firm survives Coase even when costs collapse #

The economic version is the interesting one for builders. The singularity argument is Coasean at its core: firms exist because internal coordination once beat market transaction costs, and AI inverts that. So does the firm dissolve?

Not cleanly, and the reasons are organizational before they’re legal. California Management Review’s analysis of agent-heavy firms argues that uncoordinated agent proliferation recreates market-like chaos inside the org, likening it to “increasing the Brownian motion of particles within a system” (CMR, Apr 2025). The three failure modes it names are organizational entropy, platform lock-in (where a platform “becomes a gatekeeper” and “the organization is at risk of being dismembered”), and digital fragmentation from departments running incompatible agents. The firm matters here not as some abstract “trust device” but as the thing that keeps a hundred autonomous agents from tearing the org apart and chaining it to whoever owns the platform. Coordination doesn’t vanish when it gets cheap; it changes shape and resurfaces as a coherence problem.

The liability literature points the same direction without over-claiming. A January 2025 paper in AI & Society maps four distinct liability-gap types autonomous systems create — no actor liable, the wrong actor liable, blame shifting between actors, and actors escaping across legal regimes (PMC). The honest reading is not that the corporate form resolves all four; the authors actually favor distributed, proportional responsibility shared across the humans, vendors, and deployers in the chain. What the paper does establish is that liability has to be anchored to identifiable parties — and the corporate form is one durable way to provide that anchor. That’s a weaker, truer claim than “the company is the answer,” and it’s enough to keep the wedge standing.

The honest counter-case #

The wedge is real, but it’s not invulnerable, and a builder should know exactly where it gives.

Someone could just give the AI personhood. The gap could close from the other side — grant AI systems limited legal personhood, a new quasi-agent category subject to direct regulation, the way corporations were themselves once a novel legal fiction. A 2025 UK Law Commission paper floated it as a “radical possibility.” The counterargument is sharp: AI personhood becomes the ultimate liability shield, letting firms hide behind judgment-proof shells with no assets to make victims whole. Tellingly, the EU withdrew its proposed AI Liability Directive in 2025 under industry pressure, leaving the corporate-accountability model in place by default rather than by design. The wedge holds partly because the alternative is worse and partly because nobody finished building it.

The wedge can persist legally while hollowing out functionally. This is the failure mode that should worry you most. The whole thing assumes humans actually exercise the oversight the law demands. But as of 2025 only 36% of boards had a formal AI governance framework, and just 6% had any AI-related management reporting metrics (NACD survey). If oversight degrades into rubber-stamping under automation bias, the container stays legally intact while emptying out — present on the org chart, absent in practice. Article 14 anticipates exactly this and legislates against it, but a checkbox is not judgment.

For commodity transactions, the market may route around the firm. If agents eventually hold cryptographic identities, sign smart contracts, and face programmatic penalties, agent-to-agent contracting could reduce the marginal value of firm boundaries for low-stakes, automated micro-transactions — the “Coasean Singularity” scenario modeled in a 2025 NBER paper. The wedge is strongest in high-stakes, regulated, relationship-sensitive domains and weakest for commodity flows. It is not a universal law. It’s a gradient.

What to actually do with this #

The takeaway for anyone building an AI-native operation: keep the company, but be precise about what it’s for. It is not the coordination layer anymore — agents do that, cheaper than your middle managers did. It’s the accountability shell. The scarce, non-automatable functions live in that shell: who signs, who is named when it goes wrong, who holds the duty of loyalty, who keeps the agent swarm coherent and off a single platform’s leash. Design those deliberately. The wedge is thin — the narrow strip of human accountability that survives when everything else gets automated — but thin is not weak. It’s the part the law, the regulators, and the courts all reach for first, and the one piece of the org you can’t refactor away.