Fund accounting is one of the highest-pressure entry-level jobs in finance. Every day, you calculate the Net Asset Value of the funds in your book. Get it wrong badly enough, you’re gone. When the market closes, there’s roughly 45 minutes to gather all necessary information, reconcile it, and produce an accurate NAV. The margin for error is zero.
On a floor of about 250 accountants sharing a single Bloomberg Terminal, two decades ago the post-close rush meant forming a line, waiting your turn, manually typing in dozens of CUSIPs without error, writing down rates by hand, then rushing back to re-enter everything. Five minutes of the most critical 45 minutes of the day, gone to pure manual process.
The fix was straightforward: pre-code the CUSIPs on a CSV file saved to a 3.5-inch floppy disk, walk up to the Bloomberg, run a macro, and retrieve every value in a single batch. Five minutes became twenty seconds.
But the more important move was what came next. Rather than keeping the shortcut personal, the solution was brought to the IT team. The business problem was explained, and they engineered a technical solution. Working together turned a clever hack into a production-grade fix that eventually rolled out across the entire floor. That collaboration, between the user and the technical team, was the difference between a workaround and a real change.
The tools have changed beyond recognition since then. The lesson hasn’t.
The Moment We’re In
It should be common knowledge by now that an easy AI demo is a trap. Large language models are dangerously good at appearing to solve problems they can’t actually solve. That means building a production-grade knowledge system from scratch when a purpose-built vendor has spent years compounding intelligence across thousands of customers is usually a losing bet. That argument stands.
But there’s a massive category of problems inside every real estate organization where building is not just the right answer, it’s the only answer. Companies that figure out how to build well, with the right guardrails, are going to operate circles around the ones waiting for a vendor to solve every problem.
Three data points frame why this matters right now:
Shadow AI is already here. Microsoft and LinkedIn’s 2024 Work Trend Index reports that 78% of AI users bring their own tools to work.
It’s financially material. IBM’s 2025 Cost of a Data Breach research found that one in five organizations experienced a breach linked to shadow AI, with high shadow AI exposure correlating with roughly $670,000 in additional breach cost.
Sensitive data is already in the prompts. Cisco’s 2024 Data Privacy Benchmark Study found that 62% of respondents had entered internal process information into GenAI tools, and 48% had entered non-public company information.
Your teams are already building. The question is whether they’re doing it in the open, with support and buy-in, or in the shadows without guardrails.
The Micro-Problem Gold Mine
There is a massive class of problems inside real estate organizations that vendors will never solve because they’re too specific and too fragmented to monetize. These are operational micro-problems: recurring reconciliations, copy-paste routines, the “export doesn’t match the format the VP wants” pain, and the glue work between systems that quietly eats thousands of hours a year.
This is where building wins, because the advantage isn’t cross-customer intelligence. It’s proximity to the workflow. It’s the person who understands the exception on Tuesdays, the reporting nuance investors expect, and the “we always do it this way because…” context that no vendor can encode because it’s completely unique to your shop.
These micro-problems are everywhere in proptech and real estate operations. A few representative examples, with rough estimates of annual hours at stake:
Governance levels: Low = internal formatting/automation, no regulated data. Medium = internal operational/financial data, human review remains primary. High = tenant/financial/compliance-impacting workflows, or anything feeding investor reporting, lender covenants, or the GL.
No venture-backed software company will ever build a product for “reformatting the Monday maintenance summary at a 200-unit garden-style portfolio in the Southeast.” The addressable market is exactly one customer. But collectively, these micro-problems are eating your organization alive.
The Data Quality Problem Isn’t New, The Failure Mode Is
Here’s the nuance most people miss: data quality has always been the hard problem, and it has always required controls.
Before anyone worried about LLM hallucination, the same class of error was everywhere. An analyst copies a formula down a column in Excel and doesn’t notice the cell reference didn’t shift, and now an entire quarterly report is wrong. Someone hardcodes a value that should have been a lookup, the source changes, and the error runs silently for months. A macro pulls from a range that was accurate when built, but a new row gets inserted and the totals are off by an entire property. Every real estate finance team has war stories like these
What’s different with AI-assisted building isn’t the category of risk: it’s the speed and confidence of the failure mode. An Excel error is wrong quietly. An LLM-powered tool can be wrong fluently — producing outputs that look polished and authoritative while being subtly incorrect. Non-determinism adds a new wrinkle: the same input doesn’t always produce the same output, which is unfamiliar territory for teams accustomed to spreadsheets that at least fail consistently.
But the control principle is identical to what it’s always been: tie the data out. Reconcile outputs against known sources. Build validation checks that catch when a number doesn’t match the upstream system. Don’t trust any tool, including Excel, Python script, or an in-house AI agent, without a control that independently verifies the result.
The organizations that are disciplined about data quality controls in their spreadsheets will be disciplined with AI tools. The ones that aren’t will struggle with both. The tools change. The requirement of discipline does not.
A Decision Rubric: What to Build vs. What to Buy
The most useful framework for this decision draws on Dan Hockenmaier’s essay “The Software Shakeout: What Is Durable and What Is Not in the Age of AI,” which argues that software durability depends on compounding value and switching costs. For real estate organizations, that translates into two questions:
Workflow specificity: Is this problem uniquely shaped by your organization, your portfolio, and your internal conventions?
Compounding value: Does the “right answer” improve materially with broad, cross-customer learning and long-tail edge cases, or is it mostly contained within your local data and preferences?
Build when specificity is high and compounding value is low. These are the micro-problems: transformations, reconciliation automation, document normalization, scheduled reporting, alerts, workflow glue between systems. The correctness criteria are mostly visible, the scope is bounded, and proximity to the workflow is the advantage.
Buy when the system must learn continuously from long-tail exceptions that no demo ever shows; when it must remain auditable across version changes — what data went in, what model ran, what rules applied, who approved, and how to roll back; when it must handle model drift as business rules, data distributions, and integrations evolve; and when it must carry vendor accountability and security controls (SOC 2, lender audit requirements, investor reporting standards) that an internal team cannot hand-wave into existence.
NIST’s AI Risk Management Framework treats lifecycle governance and ongoing measurement as core risk management functions, and for good reason. These aren’t one-time launch problems. They’re permanent programs.
The property accountant who has built a rent roll reformatter knows, from direct experience, that reformatting is straightforward and GL-level coding is hard. They’ve felt the difference. They’re not going to watch an AI-powered invoice coding demo and ask “why are we paying for that?” Because they understand what’s underneath.
Build where only you can build. Buy where the market has already solved the hard parts.
