How AI Is Changing Commercial Real Estate Work

How AI Changes Commercial Real Estate Work Before It Changes the Deal

Artificial intelligence is changing commercial real estate first by compressing the work around a deal. It can classify documents, extract lease and property data, assemble research, populate a model, draft marketing copy, compare offers, and track diligence. It does not remove the need to verify the source, choose the assumption, understand the market, negotiate the terms, or accept responsibility for the result.

The important unit of analysis is not the job title. It is the workflow.

A commercial real estate transaction is a chain of evidence, decisions, communications, and professional obligations. A fast draft at one stage can help the deal. A confident mistake can also travel into an offering memorandum, buyer model, negotiation, loan package, or closing file before anyone notices.

The best use of AI is therefore not to automate the deal. It is to reduce avoidable assembly work while making the review points more explicit.

A deal is not one task

Episode 81 of Venture Step begins with an information-movement argument. Luke Tatman describes a professional-services market in which firms once spent substantial time finding data, assembling it, building a model, and turning it into a story for the client. AI can reduce the effort attached to the first several steps.

The episode also preserves a crucial qualification. Approximate information may be useful for an early screen or a modest decision. A large lease, investment sale, financing, or development decision demands stronger evidence. The consequences of a bad assumption grow with the size, irreversibility, and legal significance of the transaction.

That is why the phrase "AI in commercial real estate" is too broad to guide a pilot. A transaction may include a client instruction, conflict checks, property records, financial statements, rent rolls, leases, market data, comparable transactions, underwriting assumptions, valuation, marketing, outreach, offers, negotiation, physical diligence, environmental work, title, survey, financing, contracts, and closing.

Different people own different decisions. The broker is not the surveyor. The analyst is not counsel. The model is not the valuation opinion. The offering memorandum is not the source record.

flowchart LR
    A["Client instruction<br/>authority and scope"] --> B["Property and document intake"]
    B --> C["Market research and comparable evidence"]
    C --> D["Underwriting and valuation support"]
    D --> E["Positioning and marketing materials"]
    E --> F["Outreach, offers, and negotiation"]
    F --> G["Diligence, financing, and close"]
    B -. "Source corrections" .-> C
    C -. "Assumption changes" .-> D
    D -. "Verified facts" .-> E
    G -. "Exceptions and new evidence" .-> D

The arrows run backward as well as forward. A survey exception can change usable area. A lease abstract can reveal an option that changes value. A buyer question can expose a weak comparable. A revised operating statement can alter the entire model.

AI is most useful when it helps the team see and manage those dependencies.

Start with the instruction, not the prompt

Before a team uploads documents or asks for a model, it needs to know what it has been engaged to do.

The current RICS Property Agency and Management Principles applies to RICS members and regulated firms across commercial and residential agency and management work. RICS client-care guidance also emphasizes a written brief, agreed scope, deadlines, outputs, limitations, conflicts, and client due diligence.

Those requirements are professional standards for a defined population, not universal law. The operating lesson travels well: authorization should precede automation.

At the instruction stage, AI can organize an intake form, compare a proposed scope with a standard engagement, identify missing fields, or route a conflict check. It should not decide that the firm has authority to act, that a conflict is manageable, or that a regulated activity falls within an individual's license.

The reviewer needs to know who the client is, who can authorize the work, which legal entity is engaged, which property interest is involved, what outcome is requested, and which exclusions matter. If the engagement does not permit a system to process confidential documents, a useful prompt does not create permission.

Document intake is the obvious first compression point

Commercial real estate work begins with uneven documents. A rent roll may use one unit convention while the model uses another. A trailing 12-month statement may combine expenses that the analyst needs to separate. Lease amendments may change terms that still appear in the original lease. A scanned survey may be legible to a person but difficult for an extraction system.

AI can classify files, recognize document types, extract fields, normalize names, compare versions, and flag missing records. That is valuable because the work is repetitive and the result can be tested against the source.

The safe output is not "the rent roll." It is a proposed structured representation of the rent roll with field-level provenance.

Every extracted figure should retain the source file, page, table, cell, or text span that produced it. The system should distinguish a blank field from a zero, a date from an inferred date, and an original term from an amended term. Confidence labels are helpful only if the team knows how they were produced and what action follows from them.

The reviewer should sample both obvious and difficult records. A system can perform well on clean leases and fail on amendments, handwritten notes, unusual expense recoveries, or a property type that was absent from the test set.

The RICS Responsible use of artificial intelligence in surveying practice standard gives a current professional benchmark. It requires regulated firms using material AI systems to address privacy, confidential data, system appropriateness, supplier due diligence, erroneous output, bias, human judgment, quality assurance, and client communication.

For a non-RICS firm, those provisions are not automatically binding. They are still a strong checklist for what an accountable professional workflow needs.

Research becomes faster, but source quality becomes more visible

Market research combines public records, paid databases, firm records, broker observations, maps, planning information, demographic data, comparable transactions, and conversations that may never appear in a database.

AI can retrieve, deduplicate, cluster, and summarize this material. It can draft a market overview or prepare a comparable table. It can also merge records that refer to different properties, treat an asking rent as a completed deal, miss a concession, overlook a change in use, or cite a page that does not support the statement.

The research output therefore needs an evidence hierarchy.

The model can help find the answer. The model is not the answer's provenance.

Underwriting support is useful when the assumptions remain exposed

Underwriting turns property and market evidence into a view of future cash flow, risk, financing, and value. AI can help map a rent roll into a model, reconcile totals, generate scenarios, explain a formula, or compare versions.

The danger is not limited to arithmetic. A perfectly calculated model can still be wrong because the lease-up schedule, renewal probability, capital expenditure, exit cap rate, debt term, expense growth, or market rent assumption is wrong.

That makes a model audit more important, not less.

The system should show what it imported, what it transformed, what it inferred, and what a human entered. Material assumptions should be visible in a controlled input area rather than hidden inside generated formulas or narrative text. A change log should show when an assumption moved and who approved it.

AI can prepare a scenario. The investment lead still decides whether the scenario is plausible. An appraiser still applies the standards and professional judgment attached to a valuation opinion. Counsel still interprets the contract. A lender still applies its credit process.

RICS is developing specific guidance for AI in real estate valuation. The public scope emphasizes source reliability, verification, professional judgment, transparency, and accountability. That is the right direction. A more capable model does not dissolve those duties.

The offering memorandum shows both the opportunity and the risk

An offering memorandum, often shortened to OM, presents a property and investment opportunity to prospective buyers. The package may combine property facts, photos, location, tenant information, market context, financial performance, comparable evidence, and an investment narrative.

The work contains exactly the elements generative systems handle well: repeated layouts, prose, tables, maps, data transformation, and a recognizable firm style.

Henry is one current example. Its Compass Commercial case study says a broker reduced a basic broker opinion of value from roughly two hours and a full OM from as much as seven hours to about 20 minutes. The system helped with demographic research, amenity mapping, comparables, financial detail, and layout.

That result is a vendor-published customer account. It is not an independent benchmark or a promise that another firm will achieve the same time, quality, or economics.

Commercial Observer's launch coverage independently described Henry as a deal-deck copilot built around inputs such as trailing financials, rent rolls, photos, firm templates, and financial models. The article mainly reports founder statements and predates the later case studies.

The credible conclusion is narrow. Offering-memorandum production is a real, documented use case. The size of the benefit must be tested inside the firm's own process.

The review point is also clear. The broker and client must verify the property facts, financial figures, claims, images, comparables, and disclosures before distribution. Polished design can increase the perceived authority of a bad fact. Faster generation should create more time for review, not remove the review.

Outreach and negotiation stay attached to authority

AI can segment a buyer list, summarize prior interactions, draft an outreach message, compare offers, and prepare a negotiation matrix. It can help a broker see that one offer has a higher price while another has a stronger deposit, shorter diligence period, clearer financing, or fewer conditions.

It cannot know the client's priorities unless those priorities have been accurately captured. It cannot accept an offer, make a representation, or disclose confidential motivation without authority.

This stage also combines relationship context with legal risk. A contact record may include information collected for a different purpose. A generated message may make a claim the firm cannot substantiate. An automated follow-up may continue after a person has opted out or after the seller changes strategy.

The useful system prepares the work for an authorized person. It does not quietly become the authorized person.

Luke's E081 point about relationships fits here. A buyer or seller may value a trusted broker because that broker knows when to call, what the other party actually cares about, where a stated position may move, and how to keep a difficult deal alive. AI can improve the broker's preparation. The negotiation remains a human and organizational commitment.

Diligence exposes the limits of summary

During diligence, a transaction may involve leases, estoppels, environmental reports, engineering, zoning, insurance, title, surveys, tax, financing, and legal documents. AI can track missing items, compare defined terms, extract exceptions, and prepare questions.

The summary is not the professional deliverable.

The 2026 ALTA/NSPS Land Title Survey Standards show why. Those standards align expectations among clients, lenders, title insurers, and surveyors. They address records research, fieldwork, mapping, certifications, and deliverables related to boundaries, easements, encroachments, access, and title matters.

An AI system can summarize a survey or compare it with a title commitment. It cannot perform the fieldwork, sign the certification, or assume the surveyor's responsibility.

The same distinction applies elsewhere. A model can extract language from an environmental report. It does not become the environmental professional. It can compare lease clauses. It does not become counsel. It can track closing conditions. It does not decide that a condition has been legally satisfied.

Choose a pilot that can fail safely

The first pilot should not be the task with the most impressive demo. It should be a task where the team can measure the result and catch an error before the error reaches a client or counterparty.

A strong candidate has meaningful volume, a clear baseline, structured inputs, a reversible output, a knowledgeable reviewer, and a way to compare the output with an authoritative source. Document classification, first-pass extraction, comparable-table preparation, or an internal OM draft may fit.

A weak first candidate combines confidential data, unstructured authority, a hard-to-reverse decision, an unclear standard of correctness, and no independent reviewer.

The pilot should record time, rework, field accuracy, source-link accuracy, exception detection, reviewer effort, failure severity, and user behavior. If the system saves 60 minutes but creates 50 minutes of difficult review, the net value is small. If it produces a beautiful draft that causes one material misstatement, average time savings do not answer the risk question.

The Venture Step guide [[How to Practice Entrepreneurship Inside a Company]] provides a full pilot method. The shorter version is simple: define the task, preserve the baseline, obtain permission, use authorized data, decide what counts as correct, keep a named reviewer, and make scale a new decision.

Closing should improve the next deal

A completed transaction contains some of the firm's best learning material. It shows which assumptions changed, which buyer questions repeated, which diligence items caused delay, which source records were wrong, and which work product actually helped the client decide.

Most of that learning is lost if the final file is treated only as an archive.

An AI-assisted closeout can compare the original instruction, marketing package, underwriting versions, offer log, diligence tracker, and closing record. It can identify corrections, trace material changes, and prepare a short internal post-transaction review.

The review should not train a general model on confidential client material by default. The firm first needs to decide what it owns, what the client owns, what contractual restrictions apply, what must be retained, what should be deleted, and whether a proposed future use is compatible with the original purpose.

The useful output is a controlled update to the firm's operating system. A corrected property alias can improve entity matching. A newly documented exception can strengthen the next extraction test. A repeated buyer question can improve the next diligence room. A model assumption that repeatedly misses can become a required review point.

This is how an information advantage compounds after AI lowers the cost of basic assembly. The moat is not the untouched archive. It is the governed loop that converts corrections and outcomes into a better next decision.

Measure reviewer effort, not only generation time

Time-to-first-draft is easy to celebrate and easy to misuse.

A proper measurement separates generation, review, correction, approval, and downstream rework. It records the severity of each error and whether the reviewer could find it from the output alone. A system that creates five harmless formatting mistakes is different from one that quietly substitutes the wrong rent roll.

The team should also watch automation bias. If reviewers accept more errors as the document becomes more polished, the workflow is becoming less safe even while the average production time falls.

For a document-extraction pilot, useful measures include field-level accuracy, material-field accuracy, source-link accuracy, exception recall, review minutes per document, and the number of corrections that propagate downstream. For an OM pilot, add client corrections, marketing-fact changes, model-to-document reconciliation, and time from approved data to approved package.

The decision to scale should use the full result. Fast generation is one input. Reliable review, lower total effort, controlled data use, and fewer downstream errors are the outcome.

The durable advantage is the verified workflow

AI can reduce the cost of moving from a pile of documents to a usable first pass. That changes staffing, cycle time, and the amount of work a team can evaluate.

The firm still needs to know which evidence is authoritative, which assumption is material, which professional owns the decision, what must be disclosed, and how a correction moves through the downstream work.

That is the deeper shift in commercial real estate. The information does not disappear. The value moves from manually assembling every page toward designing and governing a system that turns evidence into an accountable decision.

Readers interested in the competitive implication can continue with [[When Information Stops Being the Moat]] and [[Why AI May Squeeze the Middle Market]].

AI assisted with research organization and drafting. Dalton Anderson remains responsible for the analysis, source boundaries, and publication decision.