An AI agent is software that pursues a goal you give it — not by writing one answer, but by breaking the goal into steps, using tools to carry them out, checking its own progress, and coming back when it is done or when it needs you. That last part is the whole difference. A chatbot answers the question in front of it and stops. An agent takes a job like “find every lease in this folder that expires in the next twelve months and draft a renewal email for each” and works the job end to end. For a commercial real estate professional, the useful version of this explanation is not the computer science of how it plans. It is what an agent can actually be trusted to do across your deals, and — more importantly — the line past which a human has to stay in the loop.
The word “agent” is doing a lot of work in commercial real estate right now, and most of it is marketing. Every proptech vendor and general tool now advertises an “AI agent,” and the term has stretched to cover everything from a chatbot on a listing page to a genuine multi-step assistant. Before you decide whether any of it belongs in your firm, you need the plain definition — what an agent can do that a chatbot cannot, and what it still cannot do at all. That definition is the vocabulary that makes the small CRE firm manifesto and the CRE AI training playbook usable rather than abstract.
What an AI Agent Actually Is
An AI agent is a language model given a goal, a set of tools, and permission to use them in sequence until the goal is met. Strip out the jargon and it runs a simple loop: understand the goal, make a plan, take a step, look at the result, decide the next step, and repeat until finished. A person does the same thing when handed a task with no instructions — figure out what “done” looks like, then work toward it one action at a time. An agent automates that loop.
The engine underneath is the same generative AI that powers ChatGPT, Claude, Gemini, and Microsoft Copilot. On its own, that model only produces text: you ask, it answers, the exchange ends. An agent wraps that model in two extra abilities. The first is tools — the model can search a folder, read a PDF, send an email through your account, update a record in your CRM, or pull a figure from a spreadsheet, rather than only talk about doing those things. The second is iteration — it can take the result of one step and decide what to do next, instead of stopping after a single reply.
Put those two together and the model stops being a very fast writer and becomes something closer to a junior assistant who can be handed a multi-step task. “Draft me a renewal email” is a chatbot job. “Go through this folder of leases, find the ones expiring within a year, pull each tenant’s contact and current rent, and draft a personalized renewal email for each” is an agent job — several steps, several tools, one goal. The model plans; the tools carry it out; the loop runs until every lease is handled.
That is the entire concept. Everything else — “agentic AI,” “autonomous agents,” “copilots that act” — is a variation on a model that can plan, use tools, and iterate toward a goal.
Agent vs Chatbot vs Automation
The fastest way to understand an agent is to place it between two things you already know: the chatbot you have tried and the automation you may already run. They occupy three rungs of the same ladder, and the difference is how much they decide for themselves.
Rules-based automation is the bottom rung — Zapier, an Outlook rule, a CRM workflow. It follows a fixed recipe you wrote in advance: when this happens, do that. It never decides anything, which makes it reliable for one job and brittle the moment reality does not match the recipe.
A chatbot is the middle rung — ChatGPT, Claude, or a listing-page assistant. It understands a request in plain language and produces a response, a huge step up from a rigid rule. But it is reactive: it answers what you asked and waits. It does not take actions in your systems or carry a task across multiple steps on its own.
An agent is the top rung. It combines the language understanding of the chatbot with the ability to act like the automation — but unlike automation, it decides the steps itself. You give it the goal, not the recipe.
| Rules automation | Chatbot | AI agent | |
|---|---|---|---|
| You give it | A fixed recipe | A question | A goal |
| It decides the steps | Never | No, one reply | Yes, as it goes |
| It takes actions in your tools | Only the coded ones | No | Yes, the ones you allow |
| Handles a multi-step task | Only if pre-built | No | Yes |
| CRE example | Auto-file emails by folder | Draft one LOI | Screen a folder of deals and flag the three worth a call |
The practical read: an agent is worth reaching for when a task has several steps, touches more than one tool, and would otherwise sit on someone’s list because it is tedious rather than hard. For a single draft or a quick answer, a chatbot is faster and simpler. For a task you can fully describe as a fixed recipe, plain automation is cheaper and more predictable. The agent earns its place in the messy middle — real work that needs judgment about the next step but is too repetitive to deserve an hour of a broker’s afternoon.
What an AI Agent Does in a CRE Firm
The generic articles list agent use cases like “scheduling” and “lead follow-up.” True, and thin. Here is what an agent can take on across the work a small commercial firm produces — and why each is an agent job rather than a chatbot job.
Lease portfolio sweeps. Point an agent at a folder of executed leases and ask it to find every one expiring in the next twelve months, extract the tenant, current rent, and renewal option terms, and assemble a renewal-outreach list. Multiple documents, multiple extractions, one deliverable — the definition of a multi-step job.
Deal screening triage. When ten offering memoranda land in a week, an agent can read each, pull the asset type, price, cap rate, and location, compare them against the criteria you specify, and hand you a short list ranked by fit — so a person spends judgment on three deals instead of skimming ten.
Inbox and follow-up management. An agent with permission to read your inbox can draft replies to routine inquiries, flag the messages that need you personally, and queue follow-ups to the prospects who have gone quiet — each as a draft you approve, not a message it sends on its own.
Listing and marketing production. Give it the property facts and an agent can draft the listing description, a matching email blast to a buyer segment, and the social copy, then stage them for your review in one pass rather than three separate requests.
Investor and owner reporting. For property management and investment work, an agent can pull the month’s figures from your rent roll and reporting sheet, draft the owner or investor update in your firm’s format, and leave it for a person to check the numbers and send.
The pattern across all five matches the definition: several steps across more than one document or tool, aimed at a single outcome, where a person reviews the result before anything reaches a client. That review is not optional politeness — it is the load-bearing part, and the next section is why. For the wider picture, our look at how AI is changing commercial brokerage puts the agent in context.
Where to Let It Act, and Where a Human Must Stay
An agent is only as safe as the boundary you draw around what it may do without asking. Draw that boundary well and an agent saves real hours. Draw it badly — or not at all — and you have handed a confident, tireless junior the authority to make mistakes at speed. The rule that keeps you on the right side is simple to state and worth enforcing without exception.
Let an agent act on its own only where a mistake is cheap and reversible. Sorting an inbox, drafting internal summaries, assembling a shortlist, pulling figures into a working sheet — if it gets one wrong, you notice and fix it, and nothing left the building. This is where an agent’s speed is pure gain.
Keep a human in the loop wherever a step touches money, law, a contract, or a client. Sending an email to a prospect, committing a number in an LOI, filing anything with legal weight, moving on a deal term — an agent may draft and prepare these, but a person approves each one before it goes out. The reason is the same one that governs any generative tool: the model produces the most plausible next step, not a verified-correct one, and plausible is not good enough when a wrong cap rate or a misread renewal clause reaches a client.
The workable setup for a small firm is an agent that does the tedious middle of a task and stops at the doorway. It sweeps the leases and drafts the emails; you read and send. It screens the deals and ranks them; you decide which to pursue. It removes the gathering, the sorting, and the first draft — the parts that eat time without needing a broker’s judgment — and returns the decision to the human every time it matters. A firm that holds that boundary comes out ahead. A firm that switches on full autonomy and walks away is running an uncontrolled risk with its own name attached.
Confidentiality When an Agent Touches Your Systems
Confidentiality is a sharper question for agents than for a chatbot, and it is worth being precise about why. When you paste a document into a chatbot, you control exactly what it sees. An agent is more useful precisely because it does not need to be spoon-fed — you connect it to your inbox, your files, or your CRM and let it fetch what it needs. That access is the source of both its value and its risk.
Two questions decide whether that is acceptable, and they are the same two that govern any AI use with deal data. First, does the vendor train its model on what the agent reads? The business and enterprise tiers of the major AI products state they do not train on business-customer content by default — which is why a paid business tier, not a free consumer login, is the floor for any firm handling NDA-bound material. Second, and separately, does your NDA even permit the material to pass through a third party’s system at all? That is a firm-level judgment, not something to decide document by document in the moment.
The practical safeguards follow from those questions. Scope an agent’s access to the folders and accounts it genuinely needs, not your entire drive. Keep the most sensitive material — seller financials, NDA-bound memoranda, rent rolls with named tenants — behind a human handoff. And decide these rules once, at the firm level, so no individual broker is guessing. The full treatment of how to classify deal data lives in the training playbook; the one-line version is that an agent’s access should be a deliberate grant, never a default.
Why This Is a Training Problem First
Notice what every section above has in common: none of the value came from the software being clever, and none of the risk was solved by buying a better tool. The value came from knowing which tasks are genuine agent jobs, and the safety came from a person who understood where to draw the boundary and what to keep out. That understanding is the asset. The agent is just the instrument.
This is the mistake that sinks most small-firm adoption. A principal reads that agents are the next thing, buys a subscription, hands it to the team, and waits for hours to appear. They do not, because a powerful tool given to people who cannot yet tell an agent job from a chatbot job — or a safe action from a risky one — produces a few nervous experiments and then quiet disuse. The firms that gain from agents invest first in fluency: teaching the team what an agent is, how to supervise one, and what never to hand it. The ordered, ninety-day version of that sequence is the CRE AI training playbook.
The prize is genuine. An agent lets a fifteen-person firm carry the administrative load that used to require hires it could not justify — the sweeping, sorting, drafting, and assembling that fills a week. But it goes to the firm whose people understand the tool well enough to aim it and rein it in, not the one with the biggest software budget. The definition you just read is the first step toward being that firm.
Where to Start
You do not need to hire a technologist to put an AI agent to work, and you do not need to guess which of your tasks are the right first ones. A free AI-readiness assessment is a short working session that looks at your firm’s actual mix of brokerage, management, and acquisitions work, identifies the handful of multi-step, document-heavy tasks where an agent pays off soonest, and sets the supervision boundary that keeps it safe with your deal data. Book a free AI-readiness assessment and you will leave knowing exactly where an agent belongs in your week — and, just as usefully, where it does not.
Frequently Asked Questions
What is an AI agent in simple terms?
An AI agent is software that takes a goal you give it and works toward it in steps — planning what to do, using tools like your files or email to carry each step out, checking the result, and continuing until the job is done or it needs your approval. The plain contrast: a chatbot answers one question and stops, while an agent handles a whole multi-step task. For commercial real estate, think of it as a fast junior assistant that can sweep a folder of leases or screen a stack of deals, then hand the result to a person to review before anything reaches a client.
What is the difference between an AI agent and a chatbot?
A chatbot is reactive — you ask, it answers in plain language, and the exchange ends. An AI agent adds two things a chatbot lacks: it can use tools to take actions in your systems, and it can carry a task across several steps, deciding the next move as it goes. Asking ChatGPT to draft one email is a chatbot job. Asking an agent to find every lease expiring this year, pull each tenant’s details, and draft a renewal email for each is an agent job — multiple steps and tools aimed at one outcome.
How is an AI agent different from automation like Zapier?
Rules-based automation follows a fixed recipe you wrote in advance: when X happens, do Y. It never decides anything and breaks when reality does not match the recipe. An AI agent is given a goal instead of a recipe and works out the steps itself, adapting as it goes. Automation is cheaper and more predictable for tasks you can fully script; an agent is the better fit for multi-step work that needs judgment about the next move but is too repetitive to deserve a broker’s afternoon.
What can an AI agent do for a commercial real estate firm?
It takes on multi-step, document-heavy work: sweeping a lease portfolio for upcoming expirations and drafting renewal outreach, triaging a week of offering memoranda into a ranked short list, managing routine inbox replies, producing listing and marketing copy in one pass, and drafting investor or owner reports from your figures. In every case the agent does the gathering, sorting, and first draft, then a person reviews before anything is sent. It removes the tedious middle of a task, not the judgment at either end.
Are AI agents safe to use with confidential deal data?
They can be, with deliberate limits. An agent is riskier than a chatbot here because you connect it to your inbox, files, or CRM rather than pasting in one document, so it can reach more than you intend. Two questions govern safety: whether the vendor trains its model on what the agent reads — business and enterprise tiers of the major AI products state they do not by default — and whether your NDA permits the material to pass through a third party at all. Scope the agent’s access to only what it needs, and keep the most sensitive material behind a human handoff.
Do AI agents replace commercial real estate brokers?
No. Agents are good at the routine, multi-step work — gathering, sorting, drafting, assembling — and poor at the things that define the job: judgment on deal terms, negotiation, relationships, and accountability for the facts. The realistic outcome is that a broker using agents carries more administrative load with the same headcount and spends less time on paperwork, while the client-facing work stays firmly human. The competitive risk is not the agent; it is the broker across town who has learned to aim one.
Do I need to code or hire IT to use an AI agent?
No. Modern agents are set up and directed in plain language — you describe the goal and grant access to the folders or accounts it needs, the same way you would brief and equip a new assistant. The general-purpose business tools most firms already use are adding agent features a non-technical principal can configure. The real prerequisite is fluency, not technical skill: knowing which tasks are genuine agent jobs and where to keep a human in the loop, which is a training question rather than a hiring one.
Which AI agent tools should a small CRE firm use?
For most small firms the sensible starting point is the agent features inside the general-purpose business tools you already run — ChatGPT, Claude, Gemini, or Microsoft Copilot — chosen to match your existing Microsoft or Google setup. Standardizing the whole team on one tool matters more than which you pick, because it keeps training and data rules enforceable. Specialized proptech tools with agent features can come later. Verify any vendor’s current capabilities before you buy; agent features change quarter to quarter.
How does a small firm actually get started with AI agents?
Start with understanding, not a purchase. Learn the difference between a chatbot job and an agent job, put the team on one business-tier tool, and agree on a firm-level rule for what an agent may do on its own versus what needs human approval. Then pick two or three multi-step tasks — a lease sweep, deal triage, investor reporting — and build the habit there under supervision before widening access. The first investment that pays off is training the team to that fluency, not buying specialized software.
Dirk Jan van Veen, PhD
