AI Agents in CRE: 4 Autonomy Levels Compared

By Jude Lee · · Workflow

Commercial real estate brokers reviewing an AI-updated deal pipeline on screen in a modern office

The useful conversation in commercial real estate right now isn’t “is AI good at lease abstraction.” It’s “what is this agent allowed to do without me?” IBM’s Think coverage of research on why humans stay in the loop is a worthwhile companion read on where the automation boundary falls. The position I’ll state as my own, not as theirs: in brokerage, that boundary is drawn by who is accountable for the outcome, not by where the model’s raw capability runs out. An agent can read a 140-page lease faster than any analyst. It cannot absorb the consequence of a wrong renewal option date in a term sheet you sign.

So instead of a tool shootout, here’s a comparison of the four autonomy levels you can assign to any CRE workflow — and how to choose between them.

The four levels of agent autonomy, side by side

Level 1 — Suggest only. The AI reads what you paste or upload and produces text. No system access, no actions. Think Claude or ChatGPT drafting a callback email or restructuring your BOV narrative. Cheapest to adopt, near-zero blast radius, and it never touches your CRM.

Level 2 — Draft with data access, human sends. The agent has read access to real systems (email, CRM, the data room, a listing database) and produces an artifact you approve: an abstracted rent roll, a CIM section, a pipeline update, a follow-up sequence. Nothing leaves the building or changes a record without a click.

Level 3 — Act within a defined scope. The agent writes to systems and takes bounded actions: logging a call to the right deal record, moving a stage, scheduling a tour, tagging an inbound inquiry, filing a document into the deal folder. Every action is scoped, logged and reversible; humans audit exceptions rather than approve each item.

Level 4 — Closed loop, no routine human touch. The agent runs a whole cycle end to end — inbound message to qualified, scheduled, CRM-updated, confirmation sent — with people pulled in only on defined triggers.

Level 2: draft-and-approve
Best where the output is a document or a claim someone will rely on: lease abstracts, offering memoranda, pricing opinions, LOI terms, client-facing reports. The review step is the product — and it has known targets. Reviewers should assume the model is weakest on scanned or handwritten exhibits and faxed amendments; on amendments that supersede base-lease terms (base rent pulled from the original document when Amendment 3 reset it); on CAM language with ambiguous exclusions, caps, gross-up and admin-fee stacking; and on options with conditional triggers — renewal contingent on no ongoing default, or notice windows measured from an event rather than a fixed date. Check those four before you check anything else.
Level 3: act-within-scope
Best where the work is high-volume record-keeping with a known-correct shape: CRM hygiene, activity logging, file routing, tour scheduling, tagging. Reviewing each item costs more than the occasional fix. Requires write scopes, an audit log, and an undo path.
The autonomy level should be set by the cost of a quiet error, not by how impressive the demo was.

Mapping brokerage workflows to a level

This is opinion, not a benchmark — adapt it to your risk tolerance and your compliance posture:

  1. Level 1 — pricing judgment and negotiation strategy

    BOV conclusions, cap rate selection, counter-offer strategy, who to call and why. Use AI to structure the argument and stress-test it. The number stays yours.

  2. Level 2 — anything a client or counterparty relies on

    Lease and rent roll abstraction, CIM and OM drafting, LOI and proposal language, market surveys, listing activity reports. See our walkthroughs on rent roll and lease abstraction automation and automating CIM and offering memorandum production for what the review step should actually check.

  3. Level 3 — pipeline and record hygiene

    Email-to-CRM updates, deal stage moves, contact enrichment, document filing, tour scheduling, inquiry tagging. This is where agents quietly earn their keep, because the alternative is a broker who never updates the CRM at all. Our deal pipeline automation guide covers the plumbing.

  4. Level 4 — narrow, repetitive, low-stakes loops only

    First-touch acknowledgement of a portal inquiry, confidentiality agreement issuance from a template, tour confirmations and reminders. Each needs a named failure mode and an escalation trigger before it goes live: (1) an inquiry from a name that only partially matches an existing counterparty — “Smith Capital” versus “Smith Capital Partners” — gets threaded onto the wrong deal, so any fuzzy match below an exact-match threshold routes to a human instead of auto-linking; (2) a CA request on an asset with a restricted buyer list or a confidential seller instruction gets issued to an excluded party, so a restricted flag on the asset record blocks auto-issuance entirely; (3) a tour reminder fires after the space was withdrawn, leased, or cancelled by phone, so the reminder job re-checks listing status at send time and suppresses plus notifies when the status isn’t active. Anything smelling like a broker-of-record, agency or licensing question escalates by default.

How the level gets enforced, not just intended

Autonomy that lives in a prompt (“please don’t send anything without asking”) is not a control. Three mechanisms turn it into one:

Choosing tools once you’ve set the level

There is no single best AI tool for commercial real estate, and any list that claims one is ranking by marketing budget. What there is, is a fit between level and tool class. One time-bound caveat: vendor AI feature sets in this category change on a quarterly cadence, so re-check current product documentation before you rule anything in or out.

Folk heuristics AI will repeat back to you with unearned confidence

People search for a “30% rule in AI” as though it were a published standard. As far as we can tell, it isn’t one — there’s no governing body or peer-reviewed source behind that phrase, and it gets used loosely to mean anything from “AI should handle about 30% of a task” to “expect roughly a third of your prompts to need rework.” Treat any hard number quoted at you without a named source as marketing.

A more useful heuristic, offered plainly as opinion: if a human still has to check every output, you have a drafting tool, not an agent — and you should price it accordingly.

The 2% rule in commercial real estate is a related folk heuristic, mostly from residential rental investing: monthly gross rent of at least 2% of purchase price. It’s a screening filter, not underwriting, and it rarely survives contact with institutional CRE, where cap rate, DSCR, lease term and rollover risk do the real work. Worth knowing because AI assistants will happily repeat rules of thumb like this in a confident tone. If an agent cites one in a client-facing document, that’s a Level 2 review catch.

What AI is actually doing to brokerage jobs

JLL’s ongoing analysis of AI and its implications for real estate tracks the theme most practitioners are seeing: the task mix shifts before headcount does. Document assembly, data entry and first-draft research compress; relationship work, negotiation and judgment don’t.

Model the value transparently rather than trusting a vendor’s percentage. Each figure below is a formula you fill in, not a result:

hours × loaded rate
Recovered time — count only tasks you actually stop doing
Input: your own timesheet
deals × avg fee × lift
Captured revenue — from faster response or more prospecting touches
Input: your own pipeline numbers
errors × cost per error
Avoided rework — missed dates, wrong abstracts, stale CRM
Input: your own incident history

The honest caveat: recovered hours only become money if they get reallocated to revenue work. Run your own math with our broker-ops payback model before you commit to anything past Level 2.

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