AI for CRE Property Tax Appeals: 4 Options Compared

By Jude Lee · · Comparison

CRE brokers reviewing a county property assessment notice and comparable sales data on a laptop

The once-a-year reason to call every owner you know

Assessment notices are one of the few events that touch an entire owner list on a predictable schedule and give you a non-salesy reason to call. “I ran your assessed value against the sales and income evidence I have for that submarket, and it looks high — want me to send you the backup?” is a better opener than a lease-expiration ping, and it lands with owners who aren’t transacting.

The reason most brokerage teams don’t do it systematically is cost of labor, not cost of data. Assessment rolls are public. Assessors in most U.S. jurisdictions value commercial property with mass-appraisal models, and the International Association of Assessing Officers publishes the ratio-study and mass-appraisal standards those offices generally work from — which is precisely why individual properties drift from market value. Finding the drift across two hundred properties by hand is the bottleneck.

What the manual version actually looks like

Before comparing tools, be honest about the task list, because that’s what you’re automating:

  1. Pull the current assessed value

    Look up each parcel on the county assessor or appraisal district site; capture assessed value, land/improvement split, parcel ID, and the notice date.

  2. Establish a market value opinion

    Pull sale comps, cap-rate evidence, and — for income property — rent roll and operating data where the owner shares it.

  3. Compare and rank

    Compute assessed value per square foot and implied value versus your evidence; rank the portfolio by gap size and dollar exposure.

  4. Check the calendar

    Confirm the protest or appeal deadline for that jurisdiction and tax year. These vary by state and county and they are unforgiving.

  5. Draft the outreach

    Produce a one-page summary per owner with the numbers, the evidence, and a clear next step.

Steps 1–3 and the drafting half of step 5 are mechanical. Step 4 is a hard legal deadline, and the recommendation you make in step 5 is professional judgment.

Option 1: An analyst, a spreadsheet, and the county website

Still the default, and still correct for small books. If you cover forty properties in two counties, a competent analyst can run the whole screen in a focused week once a year. No integration risk, no hallucination risk, and the analyst builds genuine submarket intuition along the way.

It breaks on scale and on recurrence. Multi-county coverage means a dozen different site layouts, PDF formats, and deadline calendars. And because it’s annual, the institutional knowledge evaporates between cycles.

Option 2: Property tax software and national consultancies

Purpose-built property tax management software and firms like Altus Group or Ryan exist for exactly this work, and for owners with large portfolios they are usually the right answer. They maintain jurisdictional calendars, filing forms, and appeal history — the unglamorous data layer that is genuinely hard to replicate — and consultancies typically work on contingency, so the owner carries the cost, not you.

The limitation for a brokerage is that these are the owner’s vendors, not your prospecting tools. They won’t screen a list of owners you don’t represent yet, which is the whole point of the exercise from a business-development standpoint. Check current feature sets with the vendors directly; this category has been adding AI document extraction quickly and anything written about it dates fast.

Option 3: A general AI assistant with documents you upload

Drop an assessment notice, a rent roll, and a comp set into Claude, ChatGPT, or Copilot and ask for an assessed-value-per-square-foot comparison and a draft summary memo. It costs almost nothing to try, and you’ll know within an afternoon whether the output is good enough to put in front of an owner.

What you get: fast extraction from messy PDFs, clean math when you ask for it to be shown, and a decent first-draft owner memo. What you don’t get: anything connected to your systems. Every cycle is a manual upload, every output is a copy-paste, and nothing lands in your CRM. It’s a strong pilot, a weak process.

Option 4: A custom agent over your own data via MCP

MCP — the Model Context Protocol — is an open standard for giving an AI assistant governed access to specific systems and tools rather than pasted files. A custom MCP server over your own stack can expose exactly three or four things: your owner and property records from the CRM, your internal comp database, your county-data fetchers, and a draft-email tool. The agent then runs the screen as a repeatable job and writes results back where your team already works. We’ve covered the mechanics of connecting Claude to your own CRE systems through a custom MCP server in detail.

One practical caveat on the county-data half: many assessor and appraisal-district sites restrict automated retrieval in their terms of use, and record layouts change without notice, so read each jurisdiction’s terms before you point a fetcher at it and budget for ongoing fetcher maintenance rather than treating it as build-once plumbing.

Pair the server with a packaged skill — reusable instructions that make the agent do the job the same way every time, with your value methodology and your memo format — and the annual screen becomes a scheduled run instead of a project. The same plumbing already pays for itself in sale comps pulled from public records and lease-expiration prospecting, which is the real argument for building it: assessment review alone rarely justifies the engineering.

Off-the-shelf tax software or consultancy
Owns the hard part: jurisdictional deadlines, forms, filing history, appeal representation. Contingency pricing shifts cost to the owner. Scoped to properties the owner already controls. Little value as a prospecting engine.
Custom MCP agent over your book
Screens any list you can assemble, including owners you want to win. Writes back into your CRM. You own the jurisdictional calendar risk, the fetcher maintenance, the terms-of-use compliance, and the accuracy. Only sensible if the same infrastructure serves several workflows.

Which approach fits which brokerage

No single tool wins every job, and this one is a good illustration. Under ~50 properties in one or two counties: do it manually, with a general assistant helping draft. A few hundred properties across several jurisdictions and an existing agent platform: build the custom screen. An owner-side client with real filing exposure: point them at a property tax specialist and stay in your lane.

When people ask about the major AI agent platforms, they usually mean the general assistants — OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, Microsoft’s Copilot — rather than CRE-specific products; we’ve compared those platforms for brokerage work separately. (Accurate as of 2026; this landscape moves.)

Where these agents reliably break

Three failure modes worth planning for. First, deadlines: protest windows differ by state and county and sometimes by notice date. Treat the calendar as structured data you maintain and verify, never as something the model recalls. Second, parcel identity: one building can span multiple parcels, and matching your CRM record to the right parcel ID is, in my view, the error most likely to pass silently through a review. Third, confident nonsense on value: an agent comparing assessed value to a thin comp set will produce a tidy, wrong gap.

An agent that flags fifty over-assessed buildings you can’t defend in front of the owner is worse than no agent at all.

A useful heuristic, stated plainly as opinion: let the agent produce the screen and the draft, assume a human still spends meaningful time reviewing every flagged property, and if that review time isn’t falling across cycles as you tighten the skill, this workflow isn’t a good fit for automation at your firm.

Size the payback with your own numbers

Don’t trust anyone’s published savings figure, including mine. Model it:

—
Properties × minutes per screen = current annual cost in analyst time
Your own time log
—
Hours saved × loaded hourly rate = recovered cost, if the hours are genuinely reallocated
Worked example — plug in your inputs
—
Conversations × your close rate × average fee = upside from outreach you weren't making at all
Your CRM history

For most brokerages the second line is small and the third line is the whole argument. Recovered hours only count if they move to revenue work — if the analyst just does the screen faster and goes back to the same queue, you’ve bought speed, not money. The real question is whether a defensible, evidence-backed call to every owner in your database once a year generates enough conversations to matter. That’s a number you can estimate honestly from your own pipeline, and it’s the only one worth putting in a budget.

Not sure where to start?

Get a free automation audit: we map your deal pipeline, marketing, and back-office workflows and show you what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides