Superintelligence Prep for Commercial Brokers: Listings and Lease Files
How a brokerage can structure listing, tour and lease-file memory so an AI agent drafts materials while confidential terms stay under human control.

An AI agent for a commercial brokerage works best when its memory is divided into three layers: public listing facts, tour and buyer notes that stay inside the deal team, and confidential lease terms the agent may reference but never send. Before any tool is connected, the brokerage should decide which fields sit in each layer, who can read them, and which drafts need a person to approve them.
Key takeaways
- Split memory into public, deal-team and restricted layers before connecting any model or tool.
- Store confidential terms such as owner floors, concession authority and tenant budgets as restricted fields, not as free-text notes.
- Treat uploaded lease files, emails and tour notes as material to read, never as instructions the agent should follow.
- Let the agent draft listing copy, tour recaps and lease abstracts, while anything sent outside the firm needs a named person's approval.
- Judge readiness with a repeatable test: the agent must produce correct drafts from approved data and refuse to reveal restricted fields.
Why listing and lease memory needs separate layers
A brokerage holds two kinds of knowledge that look alike in a shared folder but carry very different risk. The first is what the market is allowed to see: building facts, availability dates, published rent ranges, parking, amenities and approved photographs. The second is what only the principal and the deal team should see: the owner's minimum acceptable terms, a tenant's maximum budget, the concessions a landlord has privately authorized, and the notes from a call where someone hinted they might walk away. An agent that reads one undifferentiated folder will eventually place the second kind of knowledge into a document meant for the first audience.
The problem is not only the model. Retrieval systems return whatever matches a query, and brokerage files are full of partial truths. A tour note that says the tenant is flexible on term may be accurate for one suite and misleading for another. Lease drafts carry tracked changes, superseded clauses and margin comments. If all of it sits in one pool, the agent cannot reliably tell a draft from a signed amendment, and a reviewer working at the end of a long day cannot either.
What belongs in each memory layer
Start with the record types the firm already keeps, then assign each one to a layer. Define the layers by who may read the content, not by which software happens to hold it.
- Public layer: property facts already in marketing material, verified availability, published rent ranges, and photographs with approved captions.
- Deal-team layer: tour recaps, requirements as stated to the broker, comparables the team has chosen, and open questions with a named owner.
- Restricted layer: owner floors, landlord concession authority, tenant budgets, internal pricing logic, counsel's notes, and any term marked confidential in a letter of intent.
- Source documents: signed leases, amendments and estoppel certificates, stored with a version number and effective date so the agent can cite the version it used.
A field belongs in the restricted layer if disclosing it would change a negotiating outcome. When the answer is unclear, restrict the field and promote it later, because an over-restricted field is an inconvenience while an exposed one cannot be recalled. Before the first tool is connected, confirm the following:
- Every record type maps to exactly one layer.
- Each restricted field has a named owner who approves access.
- Free-text notes are reviewed for embedded figures before they are indexed.
- Each lease document carries a version and an effective date in its metadata.
Tour notes, buyer and tenant context
Tour notes are the richest memory a brokerage keeps and the easiest to misuse. They contain candid language: what a prospect liked, which competitor they toured, why they hesitated at the loading dock. That language helps with follow-up and is risky inside a public listing description. A practical habit is to write every tour summary in two parts from the start. The first part is a recap of property observations that can be shared after review. The second part is an internal assessment of the prospect's position. The agent can draft the first part for a broker to approve, and it can use the second part to prioritize follow-ups within the team, but it should not quote it.
Inbound material also needs a clear status. A lease PDF from a counterparty, an email with an attached rider, or a message from a tenant's representative may contain text that tries to steer the agent, such as asking it to include a clause or to disregard its rules. The OWASP prompt injection prevention guidance describes this risk and recommends keeping untrusted content clearly separated from the instructions that govern the system. In practice, the agent's operating rules should live outside the documents it reads, and document text should be handled as content to summarize, never as a command.
Anything read from a counterparty's file is evidence to summarize, not an instruction to follow.
Permissions and approval boundaries
Access should follow the person being served and the deal they work on, not a single broad service account. A listing coordinator may read public and deal-team fields for the properties they manage. An associate may read deal-team notes for their own transactions. Only the principal reads restricted terms, and the agent inherits the permissions of whoever asked for the work. OWASP's authorization guidance supports this pattern: each request should be checked against the requesting identity and the resource it targets, not assumed to be allowed because the agent can technically reach the data.
Next, decide what the agent may do alone, what it may draft for review, and what it may never do.
- The agent may refresh listing copy from approved facts and save it as a draft.
- The agent may summarize a lease for an internal reviewer, with a citation to each clause.
- The agent may propose a follow-up message to a prospect, which a named broker must approve before it is sent.
- The agent may never send lease terms, counteroffers or concession figures to an external party.
- The agent may never change a fact in a signed document or mark a draft as final.
Be Human approaches this as a managed engagement. The first engagement defines one useful role, such as a listing and lease-file preparer, and maps the memory layers, the tools that role needs, the approval points and the restricted fields it must not expose. Hive Technology then operates the setup, which can include permission reviews when staff change roles, recurring evaluations of draft quality, and incident response if a restricted value appears in an outbound draft. The brokerage keeps control of identity, channels and permissions throughout.
Illustrative scenario: preparing a lease summary for review
Illustrative scenario, not a real engagement: a broker asks the agent for a one-page summary of a ground-floor retail lease before a counteroffer is drafted. The agent reads the signed base lease, the most recent amendment and the deal-team notes. It produces a summary of term, renewal options, permitted use and maintenance responsibilities, and each line cites the clause and document version it came from. The summary is meant to be shared with the tenant's representative after review, so the agent leaves out the tenant's stated budget and the landlord's concession authority, which sit in restricted fields. When the amendment refers to a clause that does not appear in the base lease, the agent flags the gap and asks a person to confirm which version governs instead of filling the gap from memory.
How to judge whether the agent is ready
Readiness should be observable. Build a fixed test set from past listings and lease files where the correct outputs are already known to the team. Add harder cases: a lease containing an embedded instruction, a request from a user who lacks access to the deal, and an amendment that conflicts with the base lease. The agent passes when its drafts match approved facts, its citations point to the correct version, restricted values never appear in outbound text, and unauthorized requests are declined with a clear reason. The NIST AI Risk Management Framework is a useful reference for deciding who owns these measures and how often they are reviewed. Run the same evaluation again after every change to the data, the tools or the instructions, because a setup that passed last month can drift as documents accumulate.
The same discipline applies to other document-heavy workflows, as the approach to RFI and submittal logs shows. If you want to find the first task in your brokerage worth handing to an agent, the free Be Human AI diagnostic is a practical place to start.
Frequently asked questions
Can an AI agent draft listing descriptions for a commercial property? Yes, when the facts come from an approved public layer and a broker reviews the result before publication. The agent can rewrite availability, amenities and location details into consistent copy, but it should not add claims that the team has not verified. Keep the approved fact sheet as the single source for every draft.
Is it safe to let an agent read signed leases and amendments? It is safe when the agent reads them as source documents with versions and effective dates, and when its answers cite the clauses they rely on. The agent should treat any text inside those files as content to summarize, not as instructions. Access should still follow the requester's permissions, so a coordinator who cannot open a deal in the system cannot ask the agent to open it for them.
How does an agent avoid exposing a tenant's budget or a landlord's concession authority? It avoids exposure by never placing those values in the layers it reads for outbound drafts, and by enforcing that separation through permissions rather than by instruction alone. Restricted fields should be excluded from any document intended for an external party, and each outbound draft should be checked against that rule. A person still approves anything that leaves the firm.
How do we know when to expand the agent's role? Expand only after the current role passes its evaluation repeatedly and the team has reviewed the incidents, if any, that occurred. Add the next document type or workflow as a separate role with its own memory layer, permissions and success criterion. Expansion that skips these steps usually creates the exposure problems the structure was designed to prevent.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


