We build commercial real estate software on Anthropic's Claude models and run them in production every day — reading broker packets, structuring texted deal updates, drafting content. This is the practitioner's answer to "should our CRE team be using Claude?": yes, for four specific jobs, each with a guardrail.
What is Claude, in CRE terms?
Claude is a large language model that is unusually good at long documents and at following structured instructions. The long-document part matters most here: Anthropic's recent models accept up to a million tokens of context — enough to read an entire lease file, or a full transaction package, in one pass instead of in fragments.
Which four uses hold up in production?
- Reading lease and deal documents. Summaries, targeted questions ("is there a personal guarantee?"), and full abstraction against a schema. Fast, strong, and the guardrail is non-negotiable: verify extracted terms against the source clause before they touch your records.
- Structuring messy updates. A broker's email, a texted note, a 40-tab tracker — Claude turns unstructured updates into structured rows (tenant, space, SF, stage, date). This is the quiet killer app for leasing teams: capture without re-keying.
- Drafting. LOI cover notes, tenant communications, market summaries, listing copy. Guardrail: a person sends everything; the model drafts, never transmits.
- Portfolio Q&A. "Which leases expire within 18 months?" works when — and only when — the underlying data is structured and current. The model is the interface; the system of record is the truth.
What guardrails do these need?
One pattern covers all four: the model proposes; a person commits. Extractions land as reviewable suggestions with the source attached. Drafts wait for a human send. Nothing writes to the rent roll or reaches a tenant autonomously, and every acceptance is logged. (The reasoning is the same as in our note on governed AI vs. autonomous agents.)
Should a CRE team use Claude directly or inside a system?
Both, for different jobs. Direct chat (on a business tier, not personal accounts) is right for ad-hoc work: summarize this PDF, draft this email. But the recurring workflows — broker updates, lease abstraction, deal capture — belong inside a system that enforces the schema, the review step, and the audit trail, because consistency is the whole point of a record. Rule of thumb: if the output should land in your records, it should arrive through a governed pipeline, not a copy-paste.