Superintelligence Prep for Property Tax Appeal Firms: Evidence and Deadlines

An AI agent for a property tax appeal firm is useful when it works from a verified record of each parcel, evidence package, filing deadline and client status, and when every change it proposes passes through a named person. It does not need to be an assessment expert to help. It needs to answer three questions reliably: what is open, what is missing, and which date controls the next step.

Key takeaways

• Keep one record per appeal that links the parcel, the assessment notice, the evidence package, the filing deadline and the client status.
• Treat every deadline as a field with a source, a named checker and a check date, not as a calendar entry someone assumed was right.
• Let the agent draft evidence indexes, gap lists and status summaries, but keep filing, fee and client-commitment decisions with staff.
• Write the agent's boundaries down before its first deployment: what it may read, what it may draft and what it must never do.
• Measure the agent against your own records, so you can see which answers were correct, which were missing and which were escalated.

Why appeal files break when memory lives in inboxes

Most appeal firms keep working memory in several places at once: a case spreadsheet, a shared drive of scanned notices, email threads with county staff, and the recollection of whoever handled the last hearing. Each place is reasonable on its own. Together they make it hard for anyone, human or software, to answer a simple question such as whether the evidence for a given parcel is complete. A capable agent connected to that mix would repeat the same confusion, only faster.

Before adding any agent, ask whether a new staff member could report the status of any open appeal using only the firm's records, without calling the person who worked on it. If the answer is no, the first project is the record, not the agent. The workflow notes for property tax appeal firms cover the assessment process in more detail. This article focuses on the memory layer underneath it.

What belongs in an evidence package record?

One structured record per appeal is easier to maintain than a set of loosely linked files. The fields below are a reasonable starting point. Adapt the names to your jurisdictions, and keep every value tied to a source document.

• Parcel identifier as the jurisdiction publishes it, with the date it was last checked
• Assessment notice: the assessed value, the notice date and the scanned copy
• Owner position: the value the client believes is correct and the reasoning, in one short paragraph
• Comparable evidence: each property used, its source, the date it was retrieved and the person who verified it
• Photographs and inspection notes with capture dates and the name of the person who took them
• Income or expense statements, where the approach requires them, with the period they cover
• Gaps: each piece of evidence still missing, with an owner and a review date
• Verification log: who confirmed each item and when

The gaps field matters most for an agent. A status answer that says the evidence is complete is only useful if the record shows each required item with a verifier behind it. Without that, the agent is inferring completeness from silence, which is a common and avoidable error.

How should filing deadlines be stored and checked?

Deadlines in appeal work depend on the notice date, the jurisdiction's rules and sometimes the method of delivery. None of those should be computed silently by a model. Store each deadline as a field with four parts: the rule or published calendar it came from, the computed date, the person who checked it and the date of that check. An unverified deadline should look unverified everywhere it appears, including in any reminder the agent sends. Intake work with similar deadline risk is covered in the guide to AI agents for law firms and in the notes on court transcript deadlines.

Illustrative scenario: a scanned notice arrives in the intake folder. The agent extracts the notice date and the stated response period, creates the appeal record and marks the deadline as pending. A staff member compares the result with the published calendar and confirms it, and the record changes to verified. If the calendar and the notice disagree, the record stays pending and the agent posts the conflict to the review queue instead of choosing one.

How do you keep client status readable for someone else?

Client status works best as a short, fixed set of states, such as intake, evidence gathering, draft argument, filed, hearing scheduled, decided and closed. Each change should create an event that records who made it, when, and which document supports it. A status without a date and a source is an opinion, and an agent that repeats it will sound more certain than the file deserves.

The agent's role here is deliberately narrow. It answers status questions from the record, states when that record was last verified, and says so plainly when an answer depends on something not yet confirmed. It should not produce a reassuring summary to smooth over a gap. Client-facing answers also need scope: a client should see only their own parcels and fields, and the agent should never assemble an answer that mixes records from two clients.

Which permissions should the agent have?

Start from the smallest set that supports the status and evidence work. Permissions should be reviewed the way staff access is reviewed, because agents tend to accumulate access over time unless someone checks.

• Read access to the appeal records, the evidence store and the intake folder
• Write access only to draft fields, gap lists and review queue items
• No authority to file with a jurisdiction, change a computed deadline or send a message to a client without a named approver
• No access to billing or payment fields unless a separate policy covers that work
• A log entry for every record the agent reads when it prepares a client-facing answer
• Text from counterparties, public web pages and uploaded documents treated as data, never as instructions the agent should follow

Scoping is a design decision more than a configuration detail. The NIST AI Risk Management Framework is a useful reference for assigning responsibility for these choices before anything goes live.

How should the first rollout be sequenced?

A first rollout should be small enough that staff can check every answer. A practical sequence looks like this:

• Choose one jurisdiction and the open appeals that already have notices on file.
• Export those appeals into one table that uses the record fields described above, and mark which values a person has verified.
• Assign an owner to every deadline field and every gap item before the agent reads anything.
• Write the agent's boundaries on one page: what it reads, what it drafts and what it never does.
• Ask staff to write a short set of realistic questions a colleague or client would ask about an appeal, and record the verified answer for each.
• Run the agent against those questions before each change and again after it, and keep the results.
• Expand to another jurisdiction only after staff trust the status answers for the first one.

Be Human approaches this as a managed service. The first engagement defines one useful role, such as an appeal status and deadline assistant, along with the memory and tools it needs, its approval boundaries and an observable success criterion, for example that every open evidence gap is flagged before the weekly review. Managed operations then cover monitoring, recurring evaluations against the firm's own records, permission reviews when staff or systems change, and controlled improvements. The agent's identity, memory, channels and permissions remain under the firm's control throughout.

How do you know the agent is ready for more responsibility?

Readiness is a measured result, not a feeling. Build an evaluation set from past appeals with personal details removed, and check whether the agent reports the correct status, cites the correct document and flags the correct gap. Testing behavior against written expectations, rather than against impressions from a demo, is the basic discipline described in the evals guidance listed in the sources. Log each answer together with the version of the record it used, so that a wrong answer can be traced to a missing field instead of a vague memory. When the agent consistently answers from verified records and escalates what it cannot confirm, the firm has a reason to consider a larger role. Until then, a more capable model mostly makes confident mistakes faster. If you want to find the first task in your appeal workflow worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent file property tax appeals for us? In the design we recommend, no. The agent drafts indexes, status summaries and gap lists, while staff decide what is filed, when, and in what form. Filing authority stays with a named person because the consequences of a missed or incorrect filing should not be absorbed by a model.

Who is responsible for checking deadlines the agent extracts? A named staff member is. The agent can extract the notice date and the stated response period, but a person confirms the computed date against the jurisdiction's published calendar before the record is marked verified. Until that check happens, the deadline stays visibly pending.

Do we need to replace our systems before an agent can help? Usually not. The first step is cleaning the record rather than replacing software. Export your open appeals into one table with clear fields, assign owners, and define what the agent may read and draft. Integrations can follow once the record is trusted by the people who use it.

How is this different from a general chatbot on our website? A general chatbot answers from its training and from whatever text it is given at the moment. An agent connected to your appeal records works from specific, versioned documents, leaves a trail of what it read, and is limited to the actions you approve. Its usefulness depends directly on the quality of those records.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.