AI workflow
AI Agents for Commercial Real Estate: 7 Parcel Workflows and the Controls They Need
What an AI agent is in plain terms, seven agent workflows that run on parcels and public records, how each one fails, and the controls that make agents safe to use on real deals.
An AI agent for commercial real estate is a language model that uses tools in a loop to finish a task. It looks up a parcel, reads the zoning code, checks a hearing agenda or updates a CRM record. Then it reads the result and picks the next step, until the job is done or a stop rule ends it. The agents that work in CRE today are narrow: one job, approved data sources, a citation on every claim, and a person who approves anything that leaves the building.
What is an AI agent?
An AI agent is a model, a set of tools, a loop and a stop condition. The model decides which tool to call, reads what comes back and repeats until it can answer or must hand off to a person.
Anthropic, the company behind Claude, separates two patterns. In a workflow, models and tools follow predefined code paths. In an agent, the model directs its own process and tool use. Anthropic describes agents as "typically just LLMs using tools based on environmental feedback in a loop." OpenAI's agents guide describes agents that plan and complete tasks with tools, work with other agents and keep context across steps.
In CRE, the tools are things like:
- a parcel search or county GIS query;
- a zoning code and comprehensive plan reader;
- a state entity lookup for LLC and corporation filings;
- a hearing agenda feed;
- read or write access to your CRM;
- a document drafter that cites its sources.
The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to tools and data like these, and both Claude and ChatGPT support it. It is plumbing, not a product: an agent still needs good data behind each tool.
| Workflow | Agent | |
|---|---|---|
| Who decides the steps | Your code, in a fixed order | The model, one step at a time |
| Best for | Repeatable checks on every parcel | Open-ended questions with unknown steps |
| Cost and risk | Predictable | Higher, since errors can compound |
| CRE example | Join zoning, plan and flood layers to 5,000 parcels | Follow a zoning question through cross-referenced code sections |
Anthropic's own advice is to use the simplest pattern that works. Many useful "CRE agents" are workflows with a model inside one or two steps, and that is a strength.
Which agent workflows work on parcels?
The workflows that hold up share three traits: the inputs are parcel IDs, the sources are public or licensed, and a person can check the output quickly.
| Agent | Trigger | Output | Human checkpoint |
|---|---|---|---|
| Zoning research | A parcel enters review, or someone asks a question | Cited answer on uses, standards and plan category | Analyst reviews any "not permitted", conditional or PD finding |
| Owner research | A parcel is shortlisted | Owner entity, decision-maker and contact candidates with sources | A person approves before any outreach |
| County-wide site screening | A new buy box or a monthly rerun | Ranked parcel list with reasons | Analyst reviews the top of the list and every conflict |
| New sale or permit watch | Daily schedule | Digest of changes on watched parcels or buy-box matches | A person decides who to call |
| Rezoning and agenda watch | A new hearing agenda is published | Cases near your parcels, with status | A person reads staff reports for the cases that matter |
| Pipeline enrichment (CRM) | A new or stale CRM record | Filled fields, each with source and date | Change log and a queue for conflicts |
| OM and comps drafting (custom build) | A broker starts a listing package | Draft sections and a sourced comps table | The broker signs off every number |
The seven workflows in detail
1. Zoning research agent
It answers "what can I build here?" for one parcel at a time, from the code itself.
- Trigger: a parcel moves to "under review" in your CRM, or an analyst asks a question.
- Tools and data: parcel lookup, the zoning layer, the zoning ordinance and comprehensive plan text, the future land use layer.
- Steps: resolve the parcel and its jurisdiction; get the zoning district and plan category; retrieve the use table and dimensional standards; answer with quoted text and section numbers; compare zoning with the plan; flag overlays and planned developments.
- Output: a short answer with cited sections, plus flags such as a zoning and plan conflict.
- Human checkpoint: an analyst reviews anything conditional, not permitted or on a PD, and the jurisdiction's zoning verification letter confirms before closing.
Take a vacant parcel on US 301 in Riverview, Florida. This agent should surface that the zoning is CG (Commercial-General) while the plan category is R-4, which caps nonresidential floor area at 0.25 FAR. The full parcel write-up is in AI due diligence.
2. Owner research agent
It turns an owner of record into a person you can reach, with a source behind each step.
- Trigger: a parcel is shortlisted.
- Tools and data: the property appraiser record, Secretary of State entity filings, business listings, people search, your CRM for duplicates.
- Steps: normalize the owner name; classify it as person, entity or trust; match the entity filing for status, officers and registered agent; pick the likely decision-maker; gather contact candidates; score each match; check the CRM for existing contacts.
- Output: a contact record with every source, a match score and the date.
- Human checkpoint: a person approves before any call, letter or email. Matches below your threshold go to review, since a registered agent may be a filing service rather than an owner.
For the manual version, see how to find commercial property owners and is the owner of an LLC public record. Use bulk files where a state offers them. Florida's Division of Corporations publishes free daily and quarterly entity data files, which is sturdier than scraping a search page. (Source: archived copy of its data downloads page, July 29, 2026.)
3. County-wide site screening agent
It applies your buy box to every parcel in a county, then explains each pass or fail.
- Trigger: a new buy box, or a scheduled monthly rerun.
- Tools and data: the county parcel layer, zoning, future land use, flood zones, wetlands, utility service areas.
- Steps: filter by use code, zoning and acreage; overlay each candidate with plan, flood and wetland layers; score against your criteria; write a one-line reason per parcel.
- Output: a ranked list with reasons and source dates, in a spreadsheet or CRM view.
- Human checkpoint: an analyst reviews the top of the list and every parcel where two sources disagree.
A test run on unincorporated Hillsborough County shows why the explanation matters. The county's parcel layer (last updated November 18, 2026) lists 204 parcels coded vacant commercial and zoned CG. Testing each parcel's center point against the plan and flood layers, 91 sit on a residential plan category (R-1 to R-20), 35 of them R-4. And 47 have their center point in a Special Flood Hazard Area. A filter on zoning alone would pass all 204. A center-point test is a screen; the shortlist still gets full polygon checks.
4. New sale or permit watch agent
It watches for changes on parcels you care about and tells you what changed.
- Trigger: a daily schedule.
- Tools and data: a county parcel layer with sale and permit fields, a permit portal, and a listing feed you are licensed to use.
- Steps: pull records changed since the last run; compare sale date, owner and permit fields; match changes to your watchlist and buy box; summarize each change with its record.
- Output: a daily digest, quiet on days with nothing new.
- Human checkpoint: a person decides who to call. Low-confidence matches stay out of the digest.
Hillsborough County republishes the appraiser's parcels as a layer that was last updated November 18, 2026. It carries last sale date, deed type, permit count and last permit date. In that layer, 3,366 parcels show a most recent sale dated October 2026, and 18 of them are coded vacant commercial or vacant industrial. Only 21 parcels show a last permit date since November 1, 2026, so check how often a field refreshes before you build alerts on it. Listing data should come from a feed you have rights to, not from scraping a portal.
Built for your team
Have Plotzy build this workflow on your criteria
Plotzy builds AI workflows that run on parcel, zoning and owner data: your buy box, your markets, your review steps, with every answer traced to its source. Start with a 30-minute call about the workflow you want automated.
- Working prototype in the first 2 weeks
- Answers cite the code section, record or dataset they came from
- Outputs land in your CRM, spreadsheet or deal pipeline
30 minutes, no pitch deck
5. Rezoning and agenda watch agent
It reads every land use hearing agenda so no one has to, and flags the cases near your parcels.
- Trigger: a new or revised agenda is published.
- Tools and data: the jurisdiction's agenda system, the planning commission's case layer, parcel boundaries, your watchlist.
- Steps: poll for new land use meetings; parse each item for case number, folio, acreage, plan category, existing zoning and request; map the folio to a location; measure distance to your parcels; follow each case to its outcome.
- Output: a weekly digest of nearby cases with status and a link to the staff report.
- Human checkpoint: a person reads the staff reports for cases that matter. The agent never contacts applicants or speaks at hearings.
Hillsborough County publishes Zoning Hearing Master, Land Use Hearing Officer and Board land use meetings in Granicus Legistar, which has a read-only web API. The November 16, 2026 Zoning Hearing Master agenda held 32 application items: 15 planned development rezonings, 9 standard rezonings and 8 major modifications. Legistar records 18 as continued, 5 as withdrawn and 9 as heard and pending a recommendation. One item, RZ-STD 26-1580, asked to rezone 2.09 acres on Rhodine Road from AS-1 to CN(R) on R-4 land, about 1.6 miles from the Riverview parcel above. Its matter history shows it withdrawn that day. Per the agenda, recommendations are due December 9, 2026, and the Board is scheduled to hear those cases at its January 12, 2027 land use meeting. See how long rezoning takes for the usual sequence.
6. Pipeline enrichment agent (CRM)
It keeps every deal record filled with current parcel facts, so your pipeline reads the same way across the team.
- Trigger: a new CRM record, or a nightly pass over stale ones.
- Tools and data: your CRM's API, parcel and zoning data, the owner research agent, a duplicate check.
- Steps: match each record to a parcel ID; fill zoning, plan category, acreage, flood zone and owner entity; flag changes such as a filed rezoning or a sale; write each field with its source and date.
- Output: enriched records and a change log.
- Human checkpoint: the agent never overwrites a field a person typed. Conflicts go to a review queue.
7. OM and comps drafting assistant (custom build only)
It drafts the factual parts of an offering memorandum or comps table from your own data, with a source on every number.
- Trigger: a broker starts a listing package.
- Tools and data: your deal data, a comps source you license, parcel and zoning facts, your templates.
- Steps: assemble the facts with sources; draft property, zoning and location sections; build the comps table; mark every figure with its source.
- Output: a draft for the broker to edit, never a finished document.
- Human checkpoint: the broker signs off every number before anything reaches a client.
This one is a custom build wired to your licensed sources. It is not a standard Plotzy feature.
What goes wrong with CRE agents?
Agents fail in predictable ways. Each failure has a control that catches it, and a build should include every one.
| Failure | What it looks like | Control |
|---|---|---|
| Hallucinated citation | Cites a code section that does not exist or says something else | Require quoted text from retrieved documents and check that the quote supports the claim |
| Stale data | Reads a superseded ordinance or an old map layer | Store source dates, set freshness limits, rerun when sources change |
| Wrong parcel or jurisdiction | An address matches the parcel next door, or city rules get applied in the county | Resolve to a parcel ID first and confirm the taxing district |
| Rate limits | 429 errors from the model provider, or a public site that throttles or blocks bots | Honor retry-after headers, back off with jitter, batch work, use official APIs and bulk files |
| Runaway cost | A loop that keeps searching, or a full-county run on the largest model | Cap iterations and searches, set spend limits, batch work that can wait |
| Silent tool failure | A tool error read as "nothing found" | Check every tool result for errors and record missing data as missing |
| False negative | "No violations" read as "clean" | Write findings as "not found in [database] as of [date]" |
| Instructions inside documents | A web page or PDF with text aimed at the agent | Treat retrieved text as data, give tools least privilege, require approval to send or write |
Two provider details make these concrete. Anthropic's document citations return cited text with valid pointers into the documents you supplied, but a reviewer still checks that the passage supports the claim. Anthropic also returns web search errors inside a successful response, so a pipeline that checks only the HTTP status will miss them.
Which controls make agents safe to use on deals?
Five controls do most of the work. Build them in from the first prototype.
- Citations required. An answer without a source and date is rejected, not shown.
- Confidence thresholds. Owner matches, parcel matches and zoning reads below a set score go to a person.
- Review queues. Flags land in one queue with an owner and a due date, not in chat threads.
- Audit logs. Every tool call, source, model version and output is stored, so any answer can be traced and rerun.
- Read-only by default. Agents draft; people send. Write access to a CRM is limited to fields the agent owns.
Add a test set: 20 to 50 parcels with known answers that the agent reruns after every change. If accuracy drops, the change does not ship.
How much does a CRE agent cost to run?
Cost is model tokens plus tool fees plus the data behind the tools, multiplied by how many steps the loop takes. Anthropic notes that agents trade higher cost and compounding errors for flexibility.
- Batch what can wait. Anthropic's Message Batches API charges 50 percent of standard prices, with most batches done in under an hour. OpenAI's Batch API also offers a 50 percent discount, with a 24-hour completion window. Overnight county screens fit this well.
- Count tool fees. Anthropic's web search costs $10 per 1,000 searches plus tokens, and a
max_usessetting caps searches per turn. - Set spend limits. Both providers place organizations in usage tiers with monthly spend limits, and Anthropic lets you set a lower limit of your own.
- Use the smallest model that passes your test set for extraction and formatting steps.
The data often costs more than the model. Parcel, zoning, owner and listing data each come with their own licenses and limits, so price those first.
Which model provider should you use?
Use the one that scores best on your own test set at a cost you accept. Anthropic (Claude) and OpenAI both offer tool use, web search with citations and domain allowlists, batch pricing and published rate limits. Both return a 429 error when you exceed a limit, and both document how to back off and retry.
The model is the easiest part to swap. The parcel data, the code text, the controls and the review process are what make an agent useful, and they carry over if you change providers. See Claude for commercial real estate and build vs buy for the next decisions.
How Plotzy builds this for your team
Plotzy builds production agents for CRE teams on its own parcel and zoning data. It runs zoning Q&A agents, owner contact research and automated report generation in production today. A build includes:
- Your buy box and criteria, encoded so every parcel is judged the way your team judges it.
- Runs across your pipeline or a whole county, on a schedule or on demand.
- Data pipelines to county assessor feeds, GIS layers, hearing agendas, state filings and your internal systems.
- Outputs into your CRM or spreadsheet, with source and date on every field.
- Human review points at each checkpoint above, plus audit logs.
- Citations back to the code section, map layer or record behind each answer.
An engagement starts with a working prototype in the first 2 weeks. A proof of concept follows within 30 days, and production within 8 to 12 weeks, as described on custom software.
Parts of this are self-serve today. Plotzy covers 150M+ parcels across 3,200+ US counties, with a zoning category and designation on every parcel. Site selection filters by zoning, lot size, land use and geography, with folders, notes and CSV export. Zoning AI (Production plan and up) answers zoning questions from the ordinance text, with the section cited, in the Florida jurisdictions where Plotzy has loaded zoning. Prospecting covers owner contact lookup from Secretary of State filings, business listings and people search.
If one of these seven workflows eats your team's week, book a 30-minute call below to scope it.
Questions to ask before you build or buy
- Which one workflow comes first, and what does a correct answer look like?
- Does every answer carry a quoted source and a date?
- Is there a test set of known parcels, rerun after every change?
- What can the agent write or send without a person?
- How does it handle rate limits, blocked sources and tool errors?
- What does a full county run cost, and is it batched?
- Can you switch model providers without rebuilding the data layer?
Frequently asked questions
What is the best AI agent for commercial real estate?
There is no single best agent. The best one does one job your team repeats, on data you are licensed to use, with citations and a review step. Start with zoning research or rezoning watch, where answers are checkable against public records.
What is the difference between an AI agent and an AI workflow?
A workflow runs fixed steps in code, with a model inside some of them. An agent lets the model choose its next step. Workflows are cheaper and more predictable; agents suit open-ended questions.
Can an AI agent contact property owners for me?
It can find owners and draft outreach. A person should approve and send every message. Matching errors and outreach rules make unsupervised contact a poor risk.
How much does it cost to run an AI agent on parcels?
It depends on the steps per parcel, the model, tool fees and data licenses. Batch pricing at both Anthropic and OpenAI cuts model cost by 50 percent for work that can wait, and spend limits cap the downside.
Do I need MCP to build a CRE agent?
No. MCP is an open standard that makes tools easier to connect to AI applications such as Claude and ChatGPT. It helps when several apps need the same tools. The data and controls matter more.
Sources
The documentation, records and agendas below were checked on November 19, 2026. Provider features, prices and limits change often, so confirm them at the provider's documentation before you rely on them. Parcel counts come from public layers on that date.
- Anthropic: Building effective agents (December 19, 2024), accessed November 2026. Workflows versus agents, the tool loop, stopping conditions, cost and compounding errors.
- Anthropic: Rate limits, accessed November 2026. RPM, input and output token limits, 429 errors, retry-after, tiers and spend caps.
- Anthropic: Batch processing, accessed November 2026. 50 percent pricing, completion times and expiry.
- Anthropic: Web search tool, accessed November 2026. Citations, max_uses, domain filtering, errors in successful responses and pricing.
- Anthropic: Citations, accessed November 2026. Cited text and valid pointers to supplied documents.
- OpenAI: Agents, accessed November 2026. How OpenAI describes agents and the Agents SDK.
- OpenAI: Rate limits, accessed November 2026. RPM, RPD, TPM, TPD, tiers, 429 errors and backoff.
- OpenAI: Batch API, accessed November 2026. 50 percent discount and the 24-hour window.
- OpenAI: Web search, accessed November 2026. Inline URL citations and domain filters.
- Model Context Protocol: What is MCP?, accessed November 2026. Definition and client support.
- Granicus: Legistar Web API examples, accessed November 2026. Endpoints, tokens and query options.
- Hillsborough County Legistar: November 16, 2026 Zoning Hearing Master meeting, with its agenda items and the RZ-STD 26-1580 matter history, accessed November 2026.
- Hillsborough County: Parcels layer (HCPA data), accessed November 2026. Use codes, zoning, sale and permit fields, and counts.
- Plan Hillsborough: Hillsborough County future land use layer, accessed November 2026.
- Hillsborough County: FEMA flood zone layer, accessed November 2026.
- Florida Division of Corporations: Data Downloads (archived July 29, 2026), accessed November 2026. The live page blocked access.
Related reading
- AI Real Estate Due Diligence: What to Automate and What to Verify
How AI handles land and pre-acquisition due diligence from public records, which checks still need a surveyor, title insurer, environmental professional or planner, and a cited summary for a real parcel in Hillsborough County, Florida.
- AI Zoning Analysis: Screen Every Parcel Against Your Buy Box, With the Code Section Cited
How AI zoning analysis reads the zoning code against a parcel layer, why that beats a district lookup table, where it still fails, and a worked screen of 618 industrial parcels in Hillsborough County, Florida.
- AI Site Selection for Commercial Real Estate: A Land-First Screening Workflow
How AI site selection screens parcels for a specific asset across one or many markets, which data layers it runs on, how scoring works, where retail trade-area tools fit, and a car wash screen in Hillsborough County, Florida.
- Claude for Commercial Real Estate: What It Does Well and the Data It Needs
What Claude is good at for CRE teams, how Projects, Skills, connectors, Claude Code and the API fit together as of November 2026, and why any answer about a specific property needs a parcel data connection and a citation.
- Build vs Buy AI for Commercial Real Estate: A Sourced Cost Model
What it costs to build AI in-house at a CRE firm, from published wages, parcel data prices, model API rates and hosting, and when buying software or partnering with a specialist builder wins instead.