Right to Send Your Agent: Citizen Power in the AI Services Economy

The next right is delegated action

Digital rights have usually been framed as access rights: see your data, export your data, delete your data, correct your data, or stop someone from selling your data. Those rights still matter. But AI agents move the frontier from reading to acting. A useful agent does not only download a file. It checks a deadline, compares options, fills a form, calls an API, asks for an explanation, uploads evidence, tracks status, and escalates a denial.

Without a delegated-action right, the agent economy will tilt toward institutions. Large firms will automate customer service, claims handling, procurement, compliance, collections, screening, and internal operations. People will still be stuck proving the same fact to five systems, translating notices they do not understand, waiting on hold, and missing deadlines because the form was buried in a portal. That is a political choice, not a technological law.

The Innovation Party should set the opposite default. When technology creates leverage, ordinary people should be able to carry some of that leverage with them. A parent should be able to authorize an agent to renew benefits, find the missing document, and warn before a deadline. A patient should be able to authorize an agent to gather records and explain a prior-authorization denial. A worker should be able to authorize an agent to find training funds, compare pay stubs against wage rules, and assemble a claim. A small business should be able to authorize an agent to test a compliance obligation against machine-readable rules.

AI legal personhood is outside this right

The policy should be explicit about what it is not doing. It does not make an AI system a citizen, lawyer, fiduciary, officer, or legal person. It does not let an agent bind a user beyond the authority the user granted. It does not erase professional licensing, privacy law, consumer protection, health privacy, banking rules, court rules, or anti-fraud enforcement.

The legal model is delegation. The user authorizes a tool, service, organization, or representative to perform specific acts for a specific purpose. The agent may be software, but the accountability chain must include a real user, a real provider, a real service operator, or a real representative. That chain is what lets the law assign responsibility when something goes wrong.

The architecture of a delegated-agent right

A usable right needs more than a slogan. It needs public rails:

1. Agent identity. Services should be able to distinguish a user-authorized agent from a botnet, credential thief, anonymous scraper, or ordinary human login. That requires standards for machine identity, provider attestation, authentication, and key rotation.
2. Scoped authority. The user should be able to grant narrow powers: read bank transactions for a mortgage application, check benefit status for 30 days, submit one appeal packet, retrieve health records for one provider, or schedule a public-service appointment. Broad permanent access should be treated as a warning sign.
3. Consent receipts. Every grant should produce a receipt stating who is acting, what it can access, what it can do, why, for how long, what data it can retain, and how to revoke it.
4. Revocation and expiration. Revocation should be as easy as authorization. Access should expire by default unless the user renews it.
5. Audit logs. Users, providers, agencies, and enforcement bodies should be able to see which agent acted, when, under what authority, and against which record.
6. API access and conformance tests. Covered services should expose documented, standards-based APIs where possible. Public conformance tests should keep access from becoming a paper right that each incumbent can quietly break.
7. Liability rules. Users should not be blamed for hidden system failures. Providers should not face open-ended liability merely because they accepted a properly authorized agent. Agent providers and covered services should be liable for unreasonable security, deception, scope violations, negligent design, or ignored incident reports.
8. Human fallback. A person must be able to reach a human or paper process when an agent fails, when a high-stakes decision is at issue, or when digital access would exclude them.

This is the consent firewall. It is not anti-automation. It is the safety casing that lets automation serve the person who authorized it.

Start where public rails already exist

The first phase should not try to force every website in America into one standard overnight. Start where public law already recognizes authorized access and where the stakes are high.

Finance is a natural starting point. Section 1033 of the Dodd-Frank Act creates a financial data-access right, and CFPB's personal financial data rights work has already wrestled with authorized third parties, data security, privacy, fees, and representative status. That rule area is contested, which makes the lesson clearer: the country needs a durable statutory and standards framework for user-authorized agents rather than a fragile patchwork that changes with each administration.

Health care is another starting point. CMS interoperability and prior-authorization rules already push payers toward APIs, patient access, provider access, payer-to-payer exchange, and prior-authorization process data. A patient-authorized agent should be able to gather records, track prior authorization, identify missing information, and prepare an appeal. It should also be constrained by health privacy, medical accuracy, and human review when care is at stake.

Public services are the third starting point. Login.gov, mobile driver's licenses, and verifiable credentials point toward reusable identity and attribute verification. A person should not have to upload a driver's license, pay stub, disability determination, veteran status, address proof, or student status again and again if a consented credential can prove the relevant fact. The agent should help the person reuse that proof lawfully, with logs and appeal rights.

Civic AI Defender as the public option

A right to send your agent should not be a luxury product. The people who most need help navigating institutions are often the least likely to buy a premium assistant from a major platform. The platform should therefore support a public-interest layer: a Civic AI Defender or Life Admin Autopilot funded through legal aid, benefits navigators, state digital services, libraries, consumer-protection offices, disability organizations, and public-interest technology groups.

Its job would be practical, bounded, and boring in the best sense. It would help a person understand a notice, assemble a timeline, find a missing document, draft a benefits appeal, prepare a court form, compare a medical bill, dispute a debt, renew a license, or route a case to human legal aid. It would cite its sources, warn when the issue needs a lawyer or specialist, protect confidentiality, and keep the user in control.

Legal aid is the right cautionary example. LSC's Technology Initiative Grant program has long funded technology for access to justice, and LSC guidance on generative AI emphasizes safeguarding personal information and testing AI responses for reliability. That is exactly the posture this issue needs: expand power for people while refusing to pretend that a chatbot can replace professional judgment, human counsel, or due process.

Guardrails against fraud and exclusion

The strongest implementation risk is fraud. A delegated-agent standard will fail if it becomes a high-speed channel for account takeover, elder exploitation, forged consent, benefits theft, phishing, or automated harassment of service desks. The standard should treat fraud controls as core infrastructure: strong authentication, device and agent attestation, anomaly detection, high-risk confirmation, incident reporting, account recovery, and special protections for people under guardianship, coercion, or domestic abuse threat.

The second risk is exclusion. A person who lacks a smartphone, has limited English, has a disability, distrusts digital identity, lives with unstable housing, or cannot pass automated identity proofing must not be locked out. Agent rights should supplement human service, community navigators, paper routes, phone support, and in-person assistance. The point is more agency, not a new gate.

A right that makes innovation legible

This issue ties several platform threads together. PRIV-01 gives people rights over their data. PRIV-02 and the mDL/verifiable-credential work point toward reusable identity. GOV-05 argues that government should ship usable services. GOV-06 argues that the country needs institutions ready for fast technology shifts. HEALTH-02 builds health interoperability and accountable health AI. ECON-10 says productivity gains should return time to people, not only more output to institutions.

Right to Send Your Agent is the connective tissue. It turns AI from a spectacle into a civic power tool. It says the frontier should not end at the enterprise dashboard. It should reach the kitchen table, the clinic waiting room, the small business office, the legal aid desk, the benefits appointment, and the worker trying to get through a system that was built to exhaust them.