Extracting Party Information From a Power of Attorney Document

Everything downstream of a power of attorney depends on one field. A document that grants authority over a single vehicle sale and a document that grants broad authority over a person’s financial affairs are the same length, use the same vocabulary and name the same two parties. Extracting principal and agent from both and calling the job done produces two records that look identical and mean completely different things.

Scope is the field everything else depends on

The first classification is general versus limited. A limited or special power of attorney names a transaction or a subject and is confined to it; the confining language is usually a single clause (“solely for the purpose of…”, “limited to…”) and it can appear anywhere, including after the list of granted powers. A general power of attorney grants authority across subject areas without transaction-level confinement.

Extract this as a classification plus the verbatim clause that supports it. The classification is what a downstream system filters on; the quoted clause is what a person checks it against. Never emit the classification alone. A record that says { "scope": "general" } with nothing behind it is an assertion your pipeline made, and if it is wrong there is no way to notice without reopening the document.

The other half of scope is subject area. Many United States forms follow a statutory short form derived from the Uniform Power of Attorney Act, which enumerates categories — real property, tangible personal property, stocks and bonds, banks and other financial institutions, operation of an entity or business, insurance and annuities, estates and trusts, claims and litigation, personal and family maintenance, benefits from governmental programs, retirement plans, taxes. The categories are printed on the form whether or not they are granted, which is the trap in the next section. The Uniform Law Commission publishes the act text and the list of enacting states; see the Uniform Law Commission for the authoritative version and note that state enactments vary.

The powers list is a grid of marks

On a statutory short form the principal grants authority by making a mark — initials, a tick, an X — beside each category they wish to grant, or by striking through the ones they do not. This means the extraction target is not text. Every line on the form is present in the document regardless of the outcome, and the only difference between a granted power and a withheld one is a few pixels of ink in a box.

The consequence for the schema is that each category has three possible values, not two:

• granted — a mark is present beside the category.
• withheld — the category is printed on the form and no mark is present, or it is struck through. This is a deliberate act by the principal and it is a value.
• not_on_form — the category does not appear at all, because this document is not a statutory short form or uses a different enumeration.

Collapsing the second and third into “null” is the single most damaging simplification available on this document. A system reading that record cannot tell the difference between “the principal declined to grant authority over real property” and “we do not know”, and those two lead to opposite decisions.

There is a further layer worth capturing. Under the uniform act, certain powers — making a gift, creating or changing rights of survivorship or a beneficiary designation, delegating authority, waiving a survivor benefit, disclaiming property — must be granted expressly rather than being carried along by a general grant. Forms following the act put these in a separate block with their own marks. An extraction that reports only “general authority granted” has silently dropped the block that a bank or transfer agent is going to ask about.

{
  "subject_powers": [
    { "category": "real_property",        "status": "granted",  "mark": "initials", "page": 2 },
    { "category": "stocks_and_bonds",     "status": "withheld", "mark": null,       "page": 2 },
    { "category": "claims_and_litigation","status": "granted",  "mark": "x",        "page": 2 }
  ],
  "express_grants": [
    { "power": "make_a_gift", "status": "withheld", "page": 3 }
  ]
}

Durable, springing, and when authority ends

Durability is a sentence, not a checkbox on most forms: language to the effect that the power is not affected by the principal’s subsequent incapacity. Its presence or absence is a fact about the document and should be extracted as a boolean plus the supporting quote.

Effectiveness is harder, because it is not always a date. A springing power becomes effective on a determination of incapacity, often by one or two named physicians or by a named person. So effective_date is the wrong field type: the correct shape is a discriminated union of an absolute date and a stated condition, with the condition kept as text. Forcing a condition into a date field produces either a null that hides the condition or a hallucinated date, and the second is worse.

Termination has the same shape. A stated end date, revocation by the principal, the principal’s death, and in several states the automatic termination of a spouse-agent’s authority on divorce. The document will state some of these and be silent on others; report what it states and do not fill in the rest from general knowledge. The page is about parsing, and the effect of any of these on a particular transaction is a question for the organisation’s own counsel.

Co-agents, successors and the word jointly

Two named agents can mean two entirely different arrangements. Acting jointly means every act requires both signatures. Acting severally or independently means either can act alone. The distinguishing word is usually one adverb in one sentence, and it is trivially lost by an extraction that returns an array of agent names.

Successor agents are ordered and conditional in the same way a trustee chain is — the second serves only if the first is unable or unwilling — and the same modelling applies, which is worked through in detail on extracting trust terms. Store position, name, and the verbatim condition.

What breaks

• A struck-through category read as granted. A single horizontal line through printed text survives most binarisation, and a model reading recognised characters never sees it. If your form uses strike-through as the withholding convention, the mark has to be detected visually, not textually.
• The limiting clause on page 4. A form that reads as general for three pages can be confined by one sentence near the signature. Scope classification must consider the whole document, not the first block of granted powers.
• An unexecuted form. Blank signature, no notarial acknowledgment, no date. These circulate as drafts and templates and they extract beautifully. Capture the execution block — signature present, date, witnesses, notary acknowledgment — as required fields, so an unsigned document cannot enter the system as a valid one. That is the missing required field path.
• The agent’s address treated as the principal’s. Two names and two addresses in adjacent blocks, one label each. Anchor on the label, and keep the label text in the record.