AI Legal Review for CRE Deals: 4 Options Compared

By Jude Lee · · Comparison

CRE brokers and a transaction coordinator reviewing marked-up deal documents on screen in a conference room

The funding story worth five minutes of your attention

In early November 2026, SiliconANGLE reported that a startup called Athena raised $2.5 million in pre-seed funding to change how legal AI is priced. Business Insider’s markets desk covered the same raise, noting the founders came out of Legora and Big Law, are pointing the product at commercial real estate, and have a16z Speedrun backing.

The raise is small. The signal is worth reading — carefully. The reporting establishes the vertical (CRE) and the pitch (a different revenue model for legal AI). My own interpretation, not something the coverage proves: the positioning suggests a team that looked at the whole legal market and concluded the sharpest wedge in CRE is the billing model, not model quality. That matches what brokerage ops teams already feel — the PSA turn, the estoppel chase, the lease exhibit nobody can find, and an hourly invoice that lands three weeks after closing.

Strip the marketing away and three distinct jobs hide under one phrase:

  1. Extraction — pulling facts out of documents: rent, term, options, assignment language, title exceptions, estoppel dates. Mechanical and verifiable, and the job AI handles best. It is not solved. It breaks on scanned or photographed exhibits and handwritten margin notes; on an amendment four documents later that quietly resets a renewal notice window while the base lease still reads as authoritative; and on multi-document chains where the correct answer only exists across the base lease, two amendments and an assignment. Any extraction output you trust has to cite the source document and page, or you can’t catch these.
  2. Comparison — checking a document against a standard: your firm’s LOI template, the seller’s prior PSA, a buyer’s required-clause checklist. Reliable to the degree the standard is actually written down.
  3. Judgment — deciding whether a clause creates unacceptable risk and what to do about it. That’s legal advice. An agent can surface it; a licensed attorney has to own it.

Most disappointment comes from buying a job-3 promise and getting job-1 performance. Most of the real value comes from doing jobs 1 and 2 relentlessly well, so the expensive human hour starts at the right paragraph.

The win isn’t replacing the attorney’s judgment. It’s making sure the attorney never spends an hour locating the thing she’s supposed to judge.

Four ways CRE deal documents get reviewed now

1. Outside counsel, unchanged (hourly)

The baseline. Many firms now use AI internally — but unless your engagement letter says otherwise, you’re billed by the hour and see none of the efficiency.

Where it wins: complex, contested or unusual deals; anything needing someone insurable behind the opinion. Where it hurts: volume work — twenty small retail leases, near-identical estoppels, routine NDA turns.

Action item that costs nothing: ask your firm, in writing, what AI tooling they use and whether volume document work can move to a flat or capped fee.

The Athena-style category: agents handle extraction and first-pass drafting, priced per matter or per subscription rather than per hour. As of late 2026 this category is very young.

The category label does not guarantee attorney involvement, so treat these as things to verify per vendor, in writing: whether an attorney admitted in the relevant state signs off on each matter, which states they’re admitted in, whose malpractice policy covers the work and at what limits, whether an attorney–client relationship is formed and with whom, how conflicts are checked, and what happens to your documents including model-training terms.

Where it wins: repeatable transaction documents at volume with predictable pricing. Where it hurts: thin track record, jurisdiction gaps, and diligence you have to do yourself.

3. AI features inside software you already own

Your CRM, data room, document management system and lease platforms are all shipping AI. MRI Software, for instance, announced an expansion of its Agora AI for leasing and property operations in November 2026.

Where it wins: zero integration work, data already in place, no new vendor review. Where it hurts: the feature does what the vendor’s median customer needs, not what your deal checklist demands, and you can’t change it. Our breakdown of AI options for purchase agreement redlines goes deeper on where embedded features hit their ceiling.

4. A custom agent on your side of the table

An AI assistant connected to your own systems via MCP — the open protocol for giving a model governed access to your tools and data — reading your deal folders, precedent library and CRM. You define a skill: a written procedure saying “for every PSA, extract these 22 fields, flag deviation from our standard on these 9 clauses, produce a one-page memo, attach the source page for each finding.”

Where it wins: your checklist, your audit trail; cost scales with usage, not headcount; it runs before counsel sees the file. Where it hurts: you build and maintain it, including evals — and it must never emit anything resembling a legal opinion. If this is new, start with how to connect Claude to CRE systems through a custom MCP server.

AI-native legal service
Pitched as covering judgment, not just extraction, at predictable per-matter pricing. Verify rather than assume: who signs off, admitted where, under whose malpractice policy, with what conflicts and data terms. You adopt their process; the vendor is young.
Custom broker-side agent
You own the process and the checklist. Output is deal-readiness, never advice. Cheap at volume, works across document types. You own maintenance, evals, and the discipline to keep a human in the loop.

These aren’t mutually exclusive — option 4 can feed option 1 or 2: your agent produces a structured abstract and exception list, counsel bills for judgment on a pre-organized file.

Matching the tool to the document type and risk profile

There is no single best tool, and anyone naming one is selling. The useful question is narrower: for this document type, at this volume, with this risk profile, what’s the cheapest path to a correct answer? Low-stakes and high-volume (NDAs, estoppel chasing, lease abstraction) favors automation. High-stakes and one-off (a ground lease restructure, a contested title exception) favors a human who can be sued. Our 10-point scorecard for vetting CRE AI tools is a reasonable filter for categories 2 and 3.

Modeling the cost without making up numbers

Don’t start from a vendor’s savings claim. Start from your own invoices.

Your count
Documents routed to counsel per month
Hours × blended rate
Current monthly legal review spend
0
Legal opinions an agent should issue unsupervised

Pull three months of legal invoices and split the line items into reading and judgment. Reading hours are the addressable pool. Estimate what share of that pool a structured abstract would eliminate — be conservative, and use your own figure, not a vendor’s. Multiply by your blended rate, then compare against the all-in cost of the option you’re considering, including the ops hours to run it. The second-order gain is usually larger: faster diligence clearance, fewer re-trades from a missed clause, and coordinators freed for work that advances closings — a theme we cover in AI agents for CRE deal file compliance.

A pilot you can run without touching a live deal

  1. Pick one document type

    Choose the highest-volume, lowest-variance document on your desk — estoppels, NDAs, or standard-form leases. Not your most complex PSA.

  2. Write the checklist down

    Whatever your best reviewer checks, in order, in plain language. This document is the actual asset, and it’s what makes a vendor comparison fair.

  3. Run it blind on closed files

    Take 20 closed deals where you know the answers and run each option against the same checklist. For a vendor trial, ask for a fixed evaluation scope in writing: 20 documents of one type, your checklist, your output format, a named turnaround and a named price (including zero). Most will do a bounded evaluation; few will do open-ended free work.

  4. Score misses, not wins

    Count missed clauses and fabricated findings. A tool that’s right most of the time and confidently wrong on the rest is worse than no tool, because it erodes the reviewer’s attention.

  5. Set the human gate before you scale

    Define in writing who signs off, what the agent may never output, and where the source citation for each finding lives. Then expand one document type at a time.

What AI will and won’t do to CRE deal work

My forecast, offered as opinion: extraction and comparison keep moving toward machines, pricing for routine document work shifts from hourly toward packaged, and the document volume one transaction team can absorb rises. What doesn’t change is liability. Someone licensed still owns judgment, and your brokerage still can’t give legal advice no matter how good the agent’s output reads.

For teams with real deal volume and someone who owns operations tooling, the sequence I’d argue for is: write the checklist, run an agent against it on your side of the table, and renegotiate the engagement letter with the firm you already use.

But the build is genuinely the wrong call for a lot of shops. If you close a handful of deals a month, or you have no ops or engineering capacity to maintain prompts, skills and evals as document formats drift, a half-maintained agent is worse than none — it decays quietly and nobody notices until it misses a clause. In that case, do only the free part: write the checklist down for your own reviewers, and renegotiate the engagement letter. Skip the build until volume justifies the maintenance.

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