AI and Guardianship 2026: When Your User Cannot Legally Agree to Your Terms

Q: Can a person under guardianship agree to terms of service?

It depends on the order, and the answer is usually more nuanced than yes or no. A guardianship or conservatorship removes decision-making authority in the areas a court specifies, and modern practice favours limited orders that remove as little as possible. A person under a limited order affecting only healthcare decisions retains full contractual capacity, and treating them as unable to consent is not caution — it strips a right the court deliberately left with them. A person under a plenary order that includes the estate generally cannot bind themselves contractually, and an agreement they accept may be voidable. The practical difficulty for a software company is that you almost never know. There is no register you can query, orders are not published, and the fact pattern arrives as a phone call from a relative. That is why the design answer matters more than the legal one: build a fiduciary role and a reversibility path, rather than trying to detect status you cannot see.

Q: Someone says they are a guardian and wants access to an account. What do we do?

Verify the document and the scope before doing anything, and treat the claim itself as unremarkable rather than as either trustworthy or suspicious. Three checks do most of the work. Ask for the order or instrument, not a description of it — the difference between a guardianship of the person and a conservatorship of the estate decides whether this individual has any authority over an account at all. Read for scope and date, since limited orders are common, powers of attorney are often narrower than the holder believes, and instruments get revoked. And record what you relied on: which document, which clause, who reviewed it. The failure mode in both directions is real. Granting access to a relative without authority is how a substantial amount of elder financial exploitation actually happens, and refusing a properly appointed fiduciary can leave a protected person locked out of their own care records. A defined intake process is what keeps a support agent from having to decide either way alone.

Q: Is data about a protected person handled differently?

The privacy statutes usually apply the same way, but the exercise of rights moves and the interests can conflict. A fiduciary with authority over the relevant domain generally stands in the protected person's shoes for access, correction and deletion, which means your rights-request process needs a path for an authorised representative that is more rigorous than an email address match. Two complications are worth designing for. First, the protected person retains interests of their own, and a guardian is not automatically entitled to everything — orders that expressly preserve rights of communication and association exist precisely because monitoring can be used against the person it is meant to protect. Second, deletion is genuinely fraught: a fiduciary may need records preserved for an accounting even while a family member wants them gone, and the estate's need generally outranks the preference. Where an order restricts contact or access, that restriction is a legal constraint on your product's behaviour, not a customer preference.

Q: What is supported decision-making, and why does it matter to us?

It is the alternative to guardianship that most states have now recognised in some form, and it inverts the assumption software products usually make. Under a supported decision-making agreement the person retains full legal authority and designates supporters who help them understand information and communicate a decision. Nobody signs for them. For a product this means the correct handling of a supporter is access and explanation without control — a supporter may need to see a bill, sit in on a session or receive a copy of a notice, and may not agree to terms, make a purchase or change a plan. Most software has no way to express that, offering only a full-access delegate or nothing, which quietly pushes families toward guardianship because it is the only status the software understands. Building a genuine view-and-assist role is a small feature that removes a real harm, and it is the same role a trusted contact needs.

Four Sources of Authority, and What Each One Actually Covers

These get treated as interchangeable in support queues, and they are not. The most consequential confusion in the whole area is between authority over the person and authority over the estate — one of them can pay you and one of them cannot.

Guardian of the person

Grants: Personal decisions — residence, care, services, sometimes association and communication.
Does not: Money, contracts and property, unless the same person also holds the conservatorship or the order says so.
For you: Can often consent to a care-related service. Frequently cannot agree to a paid subscription in the protected person's name.

Conservator / guardian of the estate

Grants: Financial decisions — accounts, contracts, income, assets, and paying for services.
Does not: Personal and care choices, and anything the order expressly reserves to the protected person.
For you: This is the signature that makes a paid account lawful. Verifying which role you are dealing with is the whole exercise.

Durable power of attorney

Grants: Whatever the instrument grants — which varies enormously and is often narrower than presented.
Does not: Anything outside the document's list, and everything after revocation or the principal's death.
For you: Not a court order and not a capacity determination. The principal may still hold full authority alongside the agent.

Supported decision-making agreement

Grants: Nothing. The person retains their own legal authority and chooses supporters to help them decide.
Does not: Substitute consent of any kind. A supporter cannot sign in the person's place.
For you: The person is your user and the account is theirs. Support means access and explanation, not authority.

Five Rungs, and What the Product Owes at Each

Capacity is not a switch. Your users are distributed across this ladder right now, and the rung with the most people on it is the second one — where nothing legal has happened yet and nothing in your product accommodates it.

Full capacity

No order, no instrument, no concern.

Ordinary consent. The only obligation is that your terms and purchase flows are actually comprehensible, which is a design question rather than a legal one.

Observed difficulty, no legal status

Support tickets, repeated confusion, a family member calling on the account's behalf.

The largest and least-governed group. You have no authority to restrict the person and no basis to accept a relative's instruction. What you can do is design for reversibility — confirmations, cooling-off, and a trusted-contact channel that notifies without granting control.

Agent under a power of attorney

A private document exists and someone is presenting it.

Read the instrument for the specific power being exercised rather than accepting it as blanket authority. Record which power you relied on. The principal usually retains their own rights simultaneously, so the agent's instruction does not override them.

Limited guardianship or conservatorship

A court order removing authority in named areas only.

The order is the source of truth and it is narrow by design. Authority in one domain says nothing about another, and treating a limited order as global is itself a harm — it strips rights a court deliberately left in place.

Plenary guardianship with conservatorship

Broad authority over person and estate in one fiduciary.

Consent, contracting and payment run through the fiduciary, and your records become part of an accounting a court will review. Build for the fiduciary as a distinct account role, not as a password shared with the protected person.

Your Records Are Going Into a Court Accounting

A conservator files periodic accountings and is personally accountable for every expenditure from the estate. When your product is one of those expenditures, your records become the evidence — and the format you keep them in was designed for a billing dispute, not for this.

Artefact	What an accounting needs	What products actually store
Recurring subscription charge	Must appear as an expenditure with a purpose, on behalf of the protected person, that a fiduciary can justify as being in their interest.	A line item on a card statement with a product name and no description of what it bought or who authorised it.
In-product purchase made by an agent or assistant	Requires an identifiable authorising decision and a person accountable for it.	A transaction attributed to the account, with the instruction living in a chat log that is retained for thirty days.
Price or plan change	A change in the estate's ongoing obligations, which the fiduciary is responsible for reviewing.	An email to the address on file, which is the protected person's, which nobody reads.
Refund or credit	Must return to the estate, not to a personal card or a stored balance controlled by someone else.	Store credit issued to the account, or a refund to the original payment method whoever owns it.
Account closure and data	The fiduciary must be able to obtain records for the accounting and, at termination, hand the estate over intact.	Export tied to the login of a person who may no longer be able to complete a verification challenge.

The Rung Nobody Builds For

Almost all of the harm in this area happens on the second rung, before any court is involved. A person is struggling, a family member wants to help, and the product offers exactly two options: full account access via a shared password, or nothing. Families choose the password, which destroys the audit trail, defeats every fraud control, and makes exploitation by that same family member undetectable — and it is the most common form this harm takes.

Four features that address it, none of them large

01A view-and-assist role: sees activity and receives copies of notices, cannot purchase, change plans or accept terms. This serves supporters, trusted contacts and guardians of the person equally.
02A fiduciary role as a distinct identity with its own login, so an accounting can show which human authorised which expenditure without anyone sharing credentials.
03Authorisation records retained on an estate-accounting timescale — years — rather than on a support-log timescale, and exportable by the fiduciary rather than only by the account holder.
04Reversibility by default on high-consequence actions: confirmation, a cooling-off window, and a documented unwind path. This is the only control that helps on the second rung, where you have no authority to do anything else.

Frequently Asked Questions

Can a person under guardianship agree to terms of service?

It depends on the order, and the answer is more nuanced than yes or no. Guardianship removes authority in the areas a court specifies, and modern practice favours limited orders removing as little as possible. Someone under a limited order affecting only healthcare retains full contractual capacity, and treating them as unable to consent strips a right the court deliberately left with them. Someone under a plenary order including the estate generally cannot bind themselves, and an agreement they accept may be voidable. The difficulty for a software company is that you almost never know — there is no register to query, orders are not published, and the fact pattern arrives as a phone call from a relative. That is why the design answer matters more than the legal one: build a fiduciary role and a reversibility path rather than trying to detect a status you cannot see.

Someone says they are a guardian and wants access to an account. What do we do?

Verify the document and its scope before doing anything, treating the claim as unremarkable rather than as trustworthy or suspicious. Three checks do most of the work. Ask for the order or instrument rather than a description of it — the difference between a guardianship of the person and a conservatorship of the estate decides whether this individual has any authority over an account at all. Read for scope and date, since limited orders are common, powers of attorney are often narrower than the holder believes, and instruments get revoked. And record what you relied on. The failure mode runs both ways: granting access to a relative without authority is how a great deal of elder financial exploitation actually happens, and refusing a properly appointed fiduciary can lock a protected person out of their own records. A defined intake process is what stops a support agent deciding alone.

How does this affect AI agents that make purchases?

It concentrates the problem, because an autonomous purchase collapses the authorisation record exactly where it matters most. In an ordinary account a disputed charge is resolved by looking at who clicked what. In a fiduciary context the questions are whether the expenditure was authorised by someone with authority and whether it was in the protected person's interest — and the evidence is a prompt, a model decision and a timestamp. Three controls help: bind spending limits to the payment layer rather than to instructions given to a model; attribute every purchase to an identified authorising human with their role, so an accounting can show who decided; and retain the authorisation record on the timescale an estate accounting needs, which is years rather than the ninety days a chat log usually survives.

Is data about a protected person handled differently?

Privacy statutes usually apply the same way, but the exercise of rights moves and interests can conflict. A fiduciary with authority over the relevant domain generally stands in the protected person's shoes for access, correction and deletion, so your rights-request process needs an authorised-representative path more rigorous than an email match. Two complications: the protected person retains interests of their own, and orders that expressly preserve rights of communication and association exist because monitoring can be turned against the person it is meant to protect. And deletion is genuinely fraught — a fiduciary may need records preserved for an accounting while a family member wants them gone, and the estate's need generally outranks the preference. Where an order restricts contact or access, that is a legal constraint on your product's behaviour, not a customer preference.

What is supported decision-making, and why does it matter to us?

It is the alternative to guardianship that most states now recognise in some form, and it inverts the assumption software usually makes. Under a supported decision-making agreement the person retains full legal authority and designates supporters who help them understand information and communicate a decision. Nobody signs for them. The correct handling of a supporter is therefore access and explanation without control — they may need to see a bill, sit in on a session or receive a copy of a notice, and may not accept terms, purchase or change a plan. Most software cannot express that, offering a full-access delegate or nothing, which quietly pushes families toward guardianship because it is the only status the software understands. A genuine view-and-assist role is a small feature that removes a real harm.

Do we have a duty to act if the product observes signs of incapacity?

A general software company usually has no statutory duty, but the analysis changes with what you do and who you sell to, and no duty is not the same as no exposure. Financial institutions and their vendors operate under frameworks permitting holds and trusted-contact outreach where exploitation is suspected. Care, health and social service settings pull in mandated-reporter obligations attaching to the organisation using your tool even where they do not attach to you. And everywhere, contractual promises create duties statutes did not: a product marketed as detecting cognitive decline or preventing financial exploitation has undertaken something. The safer posture is specificity in marketing about what the system observes, routing observations to a person who can act rather than to a dashboard, and never treating an alert with no recipient as a control.

Does refusing service to someone we think lacks capacity protect us?

Rarely, and it creates its own exposure. Acting on a suspicion you have no basis to form means denying service on the basis of a perceived disability, which is its own legal problem, and the suspicion is usually formed from signals — slow typing, repeated password resets, a confused support call — that correlate with age far more reliably than with capacity. There is also a substantive harm: someone cut off from a service they depend on because a model flagged them has lost something real, with no notice and no appeal. The defensible posture is not to make capacity determinations at all, which you are not equipped or permitted to do. Make the actions safe instead — confirmation on consequential steps, a route to a human, a trusted contact who is notified rather than empowered, and an unwind path. Those help every user and require no judgement about anyone's mind.

The Two-Year Question

Pick a purchase your product processed eighteen months ago. Can you produce, today, the identity of the human who authorised it, the role they held, the limit in force at the time, and the instruction they gave?

A conservator filing an accounting has to answer exactly that, about every line, under personal liability. If your system cannot, the fiduciary's only option is to stop using you — which is the commercial version of this problem, and the one that arrives first.

Your Product Has One Consent Model. Capacity Law Has Five.