Don’t Give AI Agents Personhood — Secure People, Permissions and Plumbing

# Don’t Give AI Agents Personhood — Secure People, Permissions and Plumbing

Imagine your invoicing bot not only sends an invoice, but also pays a subcontractor, tweets about it, and books a return flight—without asking anyone first. That hypothetical isn’t sci‑fi. Agentic systems can transact, publish and act across systems independently, and they’re already doing so in ways that expose businesses to real operational and legal risk.

I want to be blunt: granting legal personhood to AI agents is the wrong neck to wring. It’s a philosophical solution that looks neat on a blog but performs poorly in courtrooms, insurance markets and executive suites. Personhood re‑frames accountability away from humans and organisations who design, deploy and profit from these systems, and places it on the software itself. That raises questions that are messy or nonsensical in practice: who insures an artificial person? How do you sue it? Where do civil fines go? And how do you stop vendors exploiting this framing to sell autonomous black boxes with “personhood‑ready” marketing?

Shruti Rajagopalan’s piece is a useful wake‑up call: agents are no longer toys. They act. That matters. But the policy leap from “agents can act” to “let’s make them legal persons” is like hauling out a sledgehammer to fix a dripping tap. For small and medium enterprises (SMEs) the better path is pragmatic and operational: fix people, permissions and plumbing.

Practical examples beat hypotheticals. I’ve seen a café almost lose cash flow because its inventory agent re‑ordered supplies every night at 2am based on optimistic demand forecasts. No human was watching. The fix was neither a new statute nor a philosophical treatise—it was configuration: set a permission boundary, add rate limits, and require a manager’s approval for orders over a sensible threshold. Cheap, fast and effective.

What SMEs should prioritise

– Clean your data and codify processes. Agents amplify whatever patterns exist in your systems. If your processes are undefined and your data is messy, agents will magnify the pain.
– Treat agents like power tools. Start in low‑risk areas (scheduling, internal summaries, draft copy) and only grant external actions—especially financial ones—gradually and with oversight.
– Enforce least privilege. Don’t hand broad permissions to an agent. Scope what it can do, where it can act, and for whom.
– Require human approval for money movement. Build approval gates for payments, refunds and contract signings.
– Keep immutable logs and ensure auditability. If an agent acts, you must be able to trace who authorised it, why, and what it did.
– Red‑team and test the messy scenarios. Plan incident response for when agents act unexpectedly. Run simulations and require vendors to support testing.
– Contractually pin liability back to operators and vendors. When buying agentic capabilities, demand transparency, versioning, and clear liability allocation.

Why personhood creates new problems

Proposals to give agents legal standing highlight a real problem: accountability gaps exist. But the remedy risks creating new gaps and perverse incentives. If an AI is a legal person, does it hold insurance? Can it be sued in practical terms? Who enforces a judgment? More worryingly, personhood could incentivise vendors to market ‘autonomous’ solutions while absolving human oversight, leading to hollow compliance and increased systemic risk.

What regulators and leaders can do now

Regulators should prioritise rules that force accountability back to humans and organisations: require explainability for material decisions, mandate auditable logs for systems that act externally, and insist on contractual clarity around liability. Insurers should demand stronger operational controls before offering coverage for agentic actions.

Leaders in SMEs should adopt a posture of cautious experimentation. Use agents to increase throughput and reduce repetitive work, but keep critical decisions under human control until you have repeatable, auditable processes.

Closing practical checklist

1. Start small—deploy agents in non‑customer‑facing, low‑risk functions.
2. Define scopes and permissions—least privilege for everything.
3. Add approval gates for any monetary or contract action.
4. Maintain immutable logs and audit trails.
5. Red‑team and run incident playbooks.
6. Contractually allocate liability and demand vendor transparency.

Don’t get distracted by philosophical labels. If you sort out the people, the permissions and the plumbing first, your bots will be helpers, not headline‑makers.

If you want to chat about building practical guardrails for your AI agents—no law degree required—I’m usually around, often with coffee and a checklist.

Source: [Governing agentic AI](https://marginalrevolution.com/marginalrevolution/2026/07/governing-agentic-ai.html)