Users, not systems.

Something has shifted in real estate investment over the past few months. Where AI was, until recently, a topic that came up mostly in board decks and trade press, it now appears in actual workflows — extracting data from offering memorandums, drafting investment committee memos, populating asset management dashboards, summarizing LP communications. The release of capable, general-purpose agents like Claude Cowork has compressed the gap between curiosity and adoption from years to weeks. Practically every firm we speak with is using AI in some form.

The natural question is how these new capabilities fit alongside the enterprise software firms already use. Some have argued that agents will replace traditional platforms as the surface on which work gets done. Others see the two as complementary. The productivity gains from recent AI deployments are real; the question is what they imply for how firms should organize their data, their processes and their systems over the medium term.

This question is not academic. The way firms deploy AI now will shape their data, their processes and their competitive positions for years afterward. Productivity gains that look impressive in the short term are already accumulating in places that are difficult to audit, govern or build on. The gap between firms whose AI deployments compound on strong foundations and firms whose deployments do not is likely to widen meaningfully over the next several years.This paper considers the relationship between AI agents and enterprise software platforms — what each is for, why neither replaces the other, and how to think about deploying them together.

The most important thing to internalize about an AI agent is that it is not analogous to a software application. It is analogous to a teammate. The products are even named accordingly — Claude Cowork, Microsoft Copilot, etc. Claude is a user, not a system. It receives instructions, executes tasks, produces outputs, and moves on. Unlike most new hires, it works quickly, scales easily, and costs a fraction of a person. But, like any new hire, it has no memory of the business beyond what it has been told, no instinct for process discipline, and no obligation to anyone but the person prompting it.

Almost everything Claude can do, a human could also do, given enough time. A team of summer interns could read 200 OMs and extract the relevant terms into a spreadsheet. An analyst could build any model or deck Claude can build. What Claude offers is throughput: more work, faster, at lower cost. This is the same value proposition that automation has always offered, and it is genuinely transformative — but it is also, on its own, incomplete.

If AI agents are simply a more efficient form of automation, then enterprise software begins to look redundant — a more expensive, less flexible way of getting work done. Some have made the argument that the agent era will render traditional platforms obsolete. The argument turns on what enterprise software is for, and on that question, the case for substitution is weaker than it first appears.

Efficiency was rarely the primary reason firms adopted platforms like Yardi, Salesforce, MRI, Argus or Altrio. These systems are bought because they create consistency across teams, geographies and time. They establish a single source of truth that survives staff turnover. They enforce process discipline. They produce audit trails. They control who can see, change and approve information. They turn fragile, person-dependent activities into durable, institutional ones.

None of this is something an agent can provide on its own. Having a controller manage a paper ledger is faster than doing it yourself, but no CFO would run the books on a paper ledger, no matter how many employees they could delegate the grunt work to. The same logic applies to agents.

Industry veterans will recognize the shape of this debate. The temptation to bypass enterprise systems with a faster, cheaper, more flexible tool is not new — it just takes a new form each cycle. Excel-as-database in the 1990s. Shadow IT in the 2000s. Robotic process automation in the 2010s. Each promised to let firms work around their enterprise systems. Each delivered a real near-term productivity win. Each, at sufficient scale, ended in a reckoning when the workarounds calcified into infrastructure that no one could trust and no one could replace.

The risk in this moment is not that firms will fail to adopt AI. The risk is that they will adopt it so successfully that they convince themselves they no longer need anything else. The productivity gains are real and visible. An analyst who used to produce two deal memos a week may now produce eight. A team of three may be doing the work of a team of ten.

The harder question is where the resulting output goes. In firms without strong platform foundations, the answer is typically: into personal OneDrives, ad hoc Excel files, email threads and chat messages. The work is being done, but it is dissolving into the same silos that platforms were originally built to eliminate.

Eventually, someone will ask where a particular number came from. It might be an LP doing diligence. It might be an auditor at year-end. It might be a regulator. The answer "Claude generated it" is not sufficient, and neither is an agent's chat log. The institutional answer is an audit trail in a system of record: who entered this, when, under what assumptions, approved by whom, modified how. Platforms produce that trail by design. Agents do not.

The strongest case for enterprise platforms in the agent era is not defensive. It is that platforms are the substrate that makes agents materially more useful. An agent reading from a clean, well-modeled deal database produces work that is more accurate, more consistent and more reusable than an agent parsing PDFs out of a SharePoint folder. The same agent, given the same task, will perform meaningfully better when the underlying data is structured than when it is not.

The Model Context Protocol (MCP) is the bridge that makes this concrete. MCP turns an enterprise platform into a well-described surface of tools, queries and actions that an agent can read and act against safely, with permissions and audit logging applied by default. Firms with strong platform foundations are not just protected in the agent era; they are advantaged. Their agents have more leverage, produce better work and operate within guardrails that less-disciplined competitors cannot easily replicate.

The dependency runs in both directions. Enterprise platforms have, for decades, struggled with low utilization. Data entry is painful, reports are clunky, workflows are rigid. Firms pay for sophisticated systems and then run their actual business in Excel because the platform is too cumbersome to live in.

Agents change this equation. An agent can populate a platform from source documents in seconds, answer questions about its data in natural language and drive workflows on the user's behalf. The cost of using a platform falls, and the latent value of systems that firms have already paid for is finally unlocked. Platform and agent form a flywheel: better data makes the agent more useful, and the agent then puts more high-quality data into the platform, making the next run better still.

The strategic impliction of all of this is uncomfortable for laggards. AI does not flatten competitive advantage in this industry — it compounds it. Firms with disciplined data, clean platforms and well-governed processes will see AI multiply those strengths. Firms with fragmented data, shadow systems and weak governance will see AI multiply those weaknesses. The gap between operationally mature firms and the rest is likely to widen, not close, over the next several years.

The implication for capital allocators is that AI fluency, on its own, is no longer a sufficient diligence question. The deeper question is what sits underneath the AI. A manager who talks fluently about agents but cannot demonstrate a clean system of record is generating outputs without the institutional substrate to sustain them.

The strongest objection to this argument is that AI agents will eventually evolve to become platforms themselves. With persistent memory, multi-agent orchestration and richer state, the argument goes, the agent layer will subsume what we currently call enterprise software. Why invest in platforms that the next generation of agents will make obsolete?

Two responses. First, even granting this argument, firms cannot run their business on the assumption that infrastructure they need today will be made unnecessary by software that does not yet exist. Second, even if the interface changes, the underlying needs do not. Whatever form the next generation of platforms takes, it will still need to provide structured data, governed workflows, audit trails and a single source of truth. Those requirements flow from the nature of the business, not from the technology of the moment.

The relationship between AI agents and enterprise software is not one of substitution. Agents perform work — they are users, in the same sense that human employees are users. Platforms provide the structured environment in which that work is captured, governed and made durable. The two technologies do different jobs, and the framing that they compete obscures more than it reveals.

The more useful observation is that each makes the other materially more capable. Agents perform better on top of well-modeled data with well-defined actions to take. Platforms get used more when agents lower the friction of operating them. The relationship is symbiotic, not merely complementary.

For real estate firms, this means that the productivity gains now being reported from AI adoption are likely to compound in firms that invest in both layers, and to plateau or unwind in firms that invest in only one. For those building the platforms underneath them, it means that the foundational work of structuring real estate data and bringing transaction processes online remains as important in the agent era as it was before, since the productivity of agents depends directly on the quality of the substrate they operate against.