CRE has adopted AI. So why hasn’t the work changed yet?

AI has reached the individual, but not the institution

We’ve held over 1,000 conversations and conducted comprehensive audits across commercial real estate firms to understand how they work, where AI is being used, and what is preventing broader adoption.

Across the board, we see individuals using ChatGPT, Copilot, Claude, and other tools for writing, market research, document abstracts, financial analysis, memo generation, modeling, and meeting prep.

All this is great, but most companies have not crossed the threshold from individual usage to organization-wide adoption and AI-assisted decision making.

Most CRE firms are stalled between individual skills and a firm-wide operating system

Level 1 — Individual Adoption of General AI Tools. Individuals are using AI for isolated workflows requiring new prompting and iteration every time; all the value stays with the person who created it.

Level 2 — Team Adoption of General AI Tools. Repeatable tasks become skills and agents shared across the firm, but the data is still unstructured and siloed.

Level 3 — Contextualized OS. Data from different operators, systems, and documents is standardized across the organization into a single governed source of truth. AI works from approved, structured records instead of raw files, which raises accuracy and token efficiency across the board.

Level 4 — Intelligent Decisions. AI learns how the firm operates, then proactively surfaces anomalies and recommends actions.

Almost every firm we audited sits in Level 1 or early Level 2. The true ROI begins when AI stops running on the same messy data and undefined processes and starts running on structured, validated data.

Firmwide AI requires governed data and repeatable processes

A firm’s edge lives on its proprietary data and the decisions built on it. Yet data compiled through years of underwriting and asset management remains scattered across OMs, models, leases, inboxes, shared drives, legal pads, and legacy software systems.

Achieving meaningful ROI with AI means making that information usable and accessible to AI.

Four things have to happen:

1. Turn fragmented data into a governed source of truth

The same number may appear differently in the underwriting model, accounting system, reporting file, and shared drive. No one has defined whether v5 or v30 is the final record. Every deal and property needs a single, authoritative record.

2. Standardize how information enters and changes the record

Connecting an LLM to an ERP, CRM, inbox, or shared drive does not resolve inconsistent formats, missing context, or poor data quality.

Every deal should have the same fields extracted and the same dates and commentary tracked. New opportunities have key property and deal points attached to them automatically. Monthly property financials should be mapped into a consistent chart of accounts, with those mappings maintained across the portfolio.

Every team must also define how information is validated and which version of key documents is final.

3. Define the process and capture the context behind it

Many workflows are not ready to be automated because no one has fully documented the required inputs, calculation and mapping rules, review process, exceptions, or final output.

Each deal record must also capture why a decision was made, not only the numbers behind it. Broker commentary, investment rationale, business-plan assumptions, and operator-call notes turn isolated data into institutional context.

AI can help execute a well-defined process. It cannot reliably resolve an undefined one.

This is why skills take much longer to build than teams expect. Even a simple lease-abstract skill can take more than 20 hours to create, iterate, and validate.

4. Layer AI on top of governed, traceable information

When AI works directly from raw files, it must repeatedly traverse Excel models, search long documents, and reconcile competing versions of the same information. That increases token consumption and makes outputs harder to verify.

Once the data is structured and a reliable source exists, skills can work from validated information instead. Employees no longer need to find the right files or manually add context before running a workflow. This is where token efficiency improves: the model retrieves the preprocessed information instead of processing entire folders and documents.

Every output should also link back to the supporting document, page, table, or model cell. Data is constantly reconciled and manipulated in CRE work, so teams must be able to verify where each answer comes from.

Here’s what governed data unlocks:

Acquisitions:

When a new deal comes in, AI matches the extracted criteria with your database by looking for similar properties across the same property type, market, size, or seller.

Within seconds, you get an analysis: have we seen this deal before? What did we underwrite it at? Why did we pass? Context-heavy notes, transcripts, and files provide rationale for why decisions were made. Other data, like lease comps, sale comps, and county tax rules can also be queried efficiently. Because the AI doesn’t have to traverse 100 random deals, it can retrieve the important info in seconds.

Asset Management:

For asset managers, the work moves away from data normalization and into driving higher NOI and stronger returns.

When financial statements that come in from different operators are mapped to a standard chart of accounts, there is no more time spent on normalizing data. Ask which properties are running over budget on payroll this quarter, and you get a ranked list across the portfolio, with each figure traceable to the operator statement it came from.

Because performance data sits next to the call transcripts and asset-management notes, the answer includes why: the operator flagged a turnover spike on the September call, and the make-ready costs land two months later.

Add in debt covenants, leasing info, and market reports, and the decisions get sharper: which assets are within 15% of a DSCR covenant trip, which renewals are pricing below market, which submarkets are softening before it shows up in the P&Ls.

The next year will reward firms that do the boring work to optimize their data

The common objection to building out systems of record is that context windows will get bigger, agents will get better, and people will be able to rely on the models alone for more work.

But a better model still cannot validate a number no human has validated, nor does it give a team a platform to collaborate on, and it will not have detailed auditability of what work has already been done and by whom.

Most issues today are not capability gaps; they are gaps in what the organization has systemized. Once AI is operating on governed, source-of-truth data, it can start learning how your firm makes decisions and what data it needs to make those decisions.

The goal is not simply to automate today’s administrative work. It is to make employees’ work more valuable and proactive. Instead of spending days finding, formatting, and validating information, they should be driving NOI, evaluating opportunities, and raising capital.

The last year rewarded individuals who were eager to learn; the next one will reward firms that have done the unglamorous work of making their own data legible.

The firms that invest in this will spend the next few years compounding on data they already own. The firms that do not will keep waiting for the magic-box solution that never arrives.

If you want to actually see what this means in practice, reach out to ziggy@broome.ai.