What Mature AI Adoption Actually Looks Like in Commercial Real Estate
Commercial real estate has historically carried a reputation for being slow to adopt new technology. Now, however, AI might be changing the trend.
Interest in the new technology emerged quickly across the industry, and many CRE organizations have already moved beyond simple evaluation to full-on implementation. Employees are experimenting with tools, companies are testing new capabilities within their existing tech stacks, and purpose-built solutions are increasingly finding their way into operations.
The question then becomes whether these companies are adopting it in ways that actually improve the business.
Mature AI Adoption Starts With the Problem
One of the easiest mistakes to make with AI is starting with the technology itself.
“We need AI” is simply not a business strategy. It’s a way to check a box while wasting project time and budget (and solving 0 problems in the process).
A more mature approach starts with the problem: what are we trying to solve? Why is that problem difficult? And which technology is actually best suited to solving it?
In some cases, the answer will be AI. In others, it may be traditional automation, better reporting, functionality that already exists within an ERP, or simply a better-designed process. AI becomes particularly interesting in CRE because many of the industry's longstanding automation challenges stem from variation.
Real estate is inherently local, and real estate companies tend to evolve their operations over time. Every organization has its own processes, workflows, terminology, exceptions, and system configurations. Even implementations of the same major ERP can look dramatically different from one company to another.
That variation has made standardized automation difficult. Technology built around rigid rules struggles to accommodate the nuances of how an individual real estate organization actually operates. Those nuances aren't abstract. They're encoded in the core ERP, in Yardi or MRI or RealPage: the chart of accounts, entity structure, approval hierarchy, GL codes, and vendor records that define "correct" for that organization. If a tool can't read and write to that structure natively, someone has to keep the two in sync, and a process built to fix an integration gap is just a new manual process with a different name.
Well-designed AI solutions offer a completely different possibility. They can be trained and tuned around those nuances, allowing technology to adapt to the organization rather than requiring the organization to adapt to the technology.
That gives CRE an opportunity to address operational problems that previous generations of technology struggled to solve.
Move From Personal Productivity to Operational Impact
For organizations still early in their AI journey, personal enablement is a logical starting point.
Employees can use AI to organize information, manage tasks, refine written work, synthesize ideas, and eliminate some of the friction surrounding everyday knowledge work. Organizations should be creating safe ways for employees to develop those skills, with approved tools and appropriate policies around security and data use.
But individual productivity gains are only one level of AI adoption. The next step is identifying discrete operational problems where AI or automation can create measurable value. Narrower scope is usually the advantage here, not a compromise, because bounded systems concentrate learning and make error reduction visible in a way broad platforms rarely do.
Fortunately, that doesn’t require starting with a massive transformation initiative. In fact, smaller implementations can be far more valuable.
Solve a real problem. Integrate the solution into the actual workflow. Determine whether it works. Then build from there. Those incremental wins create organizational confidence and help teams understand where the technology is genuinely useful.
The opposite approach (accumulating demonstrations, prototypes, and proofs of concept that never become part of the production workflow) can create the illusion of AI adoption without delivering meaningful operational change. A prototype that works in a meeting but never makes it into the business is ultimately useless.
AI Isn't the Right Tool for Every Task
As organizations become more sophisticated about AI, one capability will become increasingly important: understanding when not to use it.
Large language models are extraordinarily useful, but they are not synonymous with AI, nor are they automatically the best technology for every process.
Consider a task involving two spreadsheets with matching data. Traditional logic, a macro, or a database query can perform the same operation repeatedly and produce the same answer each time.
An LLM may be capable of analyzing those spreadsheets, but because it is nondeterministic, the output may vary. Interestingly, that same LLM may be extremely good at helping someone build the deterministic tool that performs the task.
The same principle applies to financial reporting. Organizations need consistent definitions and repeatable outputs for core financial data. Traditional reporting tools remain extremely effective at producing that foundation. AI can then sit on top of reliable data to identify trends, analyze variance, summarize results, or surface insights that would otherwise require substantial manual analysis.
Ultimately, the goal should not be to find a way for AI to perform every task, but to select the right technology for each component of the workflow.
Keep Humans Where Judgment and Control Matter
The emergence of AI agents makes another truth increasingly important to recognize: just because technology can execute an action autonomously, doesn’t mean it should.
The emergence of AI agents makes another truth increasingly important to recognize: just because technology can execute an action autonomously, doesn't mean it should. In accounting and finance, organizations should be especially cautious about eliminating controls that exist for good reason. Part of the caution is structural. Large language models are nondeterministic, meaning the same input can produce a different answer on a second run, which is a feature when you're drafting and a liability when the output becomes a journal entry.
Segregation of duties is a prime example. Controls around bank reconciliation, payments, and approvals weren't created simply to add administrative work. They exist because these are critical points of financial risk, and while automating the repetitive work within those processes can create enormous value, automating away the control itself is a different proposition.
An autonomous payment process might appear safe when surrounded by rules, like only approving transactions below a particular threshold or only approving payments to established vendors.
But predictable rules can be exploited. A $500 threshold creates an opportunity for $499 transactions. A rule based on vendor history can create an opportunity to target an old but still-established vendor record.
As agentic technology becomes more prevalent, organizations should assume those systems will eventually be tested by bad actors. And because of this, human oversight remains an important part of responsible automation.
The objective should be to automate the repetitive work leading up to a decision while preserving human involvement where judgment, control, and accountability matter most.
A Prototype Is Not a Scalable Solution
Generative AI has made it remarkably easy to build something that works… once.
Unfortunately, building something that works reliably across an organization remains much harder.
As an automation expands from one employee's workflow to a process touching company-wide financial data, the requirements change dramatically. Permissions, access controls, integrations, security, data retention, maintenance, and continuity all become part of the equation.
The solution also has to fit within the broader technology environment.
Good automation shouldn't create another disconnected application that forces employees to manually move information between systems. It should integrate with core platforms, respect upstream and downstream workflows, and ideally become nearly invisible once implemented.
Organizations should also distinguish between building with AI and building a process that continuously runs on AI.
AI can make development dramatically faster. But a recurring process dependent on an LLM can introduce variable outputs, ongoing usage costs, model dependency, and maintenance requirements.
Sometimes those tradeoffs are worthwhile. Sometimes a deterministic solution built once and designed to produce the same result every time is a much better investment.
The Next Stage of CRE's AI Adoption
Over the next year, CRE is going to see increasing adoption of agents for internal tasks such as routing, follow-ups, triage, and task management, alongside continued growth in purpose-built solutions targeting specific operational problems.
But this means that organizations also need to prepare for the other side of AI adoption.
The same tools that allow legitimate businesses to operate more efficiently can make fraud more sophisticated. AI can produce more believable communications, conduct better research, and enable bad actors to operate at much greater scale.
That makes the preservation (and modernization) of financial controls particularly important as AI adoption accelerates.
Ultimately, the organizations furthest ahead won't necessarily be the ones using the most AI. They'll be the ones that know which problems they're trying to solve, which technologies are best equipped to solve them, where automation can remove unnecessary work, and where human judgment still adds essential value.
