How AI Helps Commercial Real Estate: A Task-by-Task Map

By Jude Lee · · Workflow

Brokers and an operations manager reviewing lease documents and a laptop in a downtown office conference room

Most AI-for-CRE content operates at the wrong altitude. “AI can analyze leases” is true and useless. What an operations lead needs is a map: for each recurring task in the brokerage, is the technology good enough to run unattended, good enough to draft with review, or not good enough yet?

The sorting rule we use is boring but effective: how expensive is a wrong answer, and how quickly would you notice? Extraction errors in a marketing flyer get caught on the first read. A wrong expense-recovery assumption buried in an underwriting model can survive to LOI. Those two tasks deserve completely different levels of automation, even though both are “AI reading documents.”

Bucket one: tasks that are production-ready today

These share a profile — structured or semi-structured input, an output a human sees immediately, and low cost of a mistake.

Bucket two: draft-then-review work

Here AI compresses the work but a qualified human must sign the output. This is where most of the value sits — and most of the risk if you skip the review step.

Underwriting is the flagship example. Models can populate a template, pull comps, and propose assumptions, but assumptions are the product. A tool that fills in a market vacancy rate from a national dataset without flagging that your submarket behaves differently is producing confident nonsense. If you’re building this workflow, our guide to automating CRE underwriting and financial models covers where to keep the human checkpoint.

Other draft-then-review tasks: offering memorandum assembly, DCF cash-flow model setup, lease abstract review for ambiguous or negotiated clauses, comp-set selection, and any client-facing analysis that carries your name.

The measure of a good CRE AI workflow isn’t how few humans touch it. It’s whether the human who signs it can spot an error in under a minute.

Bucket three: leave it alone for now

Relationship judgment, negotiation strategy, site-specific physical assessment, and anything where a fiduciary duty attaches to the judgment itself. Not because the models can’t produce plausible text — because plausible is the problem. Also worth keeping human: final pricing recommendations to a client, and any communication where tone carries the relationship.

This is also our answer to “will AI replace brokers,” and it rests on an assumption you should test against your own market rather than take from us. The document-handling and data-assembly half of the job is visibly compressing — you can watch it happen in your own abstraction queue. Our working assumption is that the other half doesn’t compress the same way, because the inputs aren’t documents: off-market inventory lives in private conversations, and trust accrues to people, not systems. If you see that changing in your submarket, weight this section accordingly. What we’d expect to shift first is the ratio — less time producing, more time in front of principals.

Sizing the payoff without making numbers up

Ignore any vendor headline that quotes a dollar figure for your brokerage. Nobody knows your inputs. Build the estimate yourself, in this order:

  1. Measure current effort. Hours per task × tasks per month, timed on three real examples. Not estimated from memory.
  2. Apply your own loaded hourly cost. Your comp data, not a survey average.
  3. Subtract what automation costs you. License fees plus the human QC pass you will still run.
  4. Then, separately, model the revenue side. Recovered broker hours × your historical revenue per producing hour — but only if those hours actually get reallocated to tours, calls, and pitches. If they evaporate into more email, that line is zero.

A worked example with numbers you should replace: say lease abstraction takes an analyst 90 minutes per lease, you process 20 leases a month, and automation plus review brings it to 30 minutes. That’s 20 hours recovered monthly. Multiply by whatever your loaded analyst cost actually is, subtract the subscription and the extra review load, and you have a defensible number. If the result is unimpressive, that’s useful information too — it means the bottleneck is somewhere else.

The third category — errors avoided — is real but hard to quantify honestly. A missed option-notice date or a commission miscalculation has a cost you can estimate from your own history. If it’s never happened to you, don’t put a number on it.

Buying a tool versus building one

In client conversations over the past year we’ve heard two competing stories: vendors consolidating toward single-platform, single-invoice pitches, and large institutional owners standing up in-house AI teams. That’s an observation from our own inbox, not reported market data — and neither story is a template for a 40-broker shop.

Configured off-the-shelf
Fastest path to value for standard workflows: CRM, marketing production, transcription, document extraction. Vendor absorbs model upgrades. You inherit their roadmap and their data model. Right answer for the large majority of brokerages.
Custom or in-house build
Worth considering when the workflow is genuinely idiosyncratic — an unusual split structure, a proprietary data asset, a compliance requirement no vendor serves. Costs recur as maintenance, not just build. Requires someone internal who owns it.

When you evaluate what people search for as “commercial real estate automation companies,” note that the category mixes three very different things: point tools, platform suites, and services firms that configure other people’s software. Ask which one you’re actually talking to. Our 10-point scorecard for vetting CRE AI tools is built for exactly this conversation.

A 30-day pilot that produces evidence

  1. Pick one bucket-one task

    Choose something with a fast feedback loop and no client exposure. Resist starting with underwriting — it’s the highest-value target and the worst first pilot.

  2. Baseline it for a week

    Have the person who does the task log actual minutes on real examples. You cannot prove improvement against a number you guessed.

  3. Run parallel, not replacement

    For two weeks, do it both ways on the same inputs. Log every discrepancy. Discrepancy rate is your real accuracy metric, not the vendor’s.

  4. Decide the review rule

    Write down, explicitly, who checks what and how long that takes. Add that time back into your savings math.

  5. Report in hours and error counts

    Present the pilot as measured hours and observed discrepancies. Let leadership apply the dollar figures using rates they already know.

If a workflow touches tenant financials provided under NDA, documents governed by confidentiality clauses in a PSA or LOI, or seller-side materials in a live process, confirm your data-handling and retention terms with qualified counsel before anything goes into a third-party model. Vendor marketing pages are not a substitute for reading the DPA.

What to do this quarter

Pick one bucket-one task and one bucket-two task. Automate the first properly. Pilot the second with a hard review gate. Measure both in hours and discrepancies, using your own baseline. That single exercise will tell you more about the future of AI in your brokerage than any market outlook report — because it uses your documents, your people, and your actual cost structure.

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