The AI Workflow, and Where It Breaks
The best acquisitions teams currently run deals through the leading AI models. Within an hour of a deal hitting the desk, an analyst has generated a BOE, a list of red flags, and the initial memo.
However, as we have learned, conviction requires more than an LLM. The model can't cite its numbers, it doesn't know your firm, and it has no market understanding.
No citations. Investing runs on provenance: the document, the line, the cell. Today's models can extract data, but every figure gets re-verified by hand before it goes into the IC memo.
No memory of your firm. A general-purpose assistant doesn't keep structured records of the deals and parties it sees. There's no file on a tenant that grows as deals come through: every lease of theirs that's ever crossed your desk, the write-up your team saved to SharePoint, the space they went dark on in 2023. What it pieces together in one chat doesn't carry to the next, and nothing holds ten years of deal history. So when that tenant shows up on a new deal's rent roll, nothing flags it, and the risk surfaces after you buy the building.
No market understanding. A frontier model knows what a T-12 is the way a textbook does. What it doesn't have is a sense of normal: the hundred rent rolls your team has seen in this market. Which is why the broker's numbers go through unchallenged: the pro forma says renewals at $1,450, your own record in that submarket says $1,275, and the question never gets asked. An underwriter thinks. The model summarizes.
The Context Engine: Extract, Connect, Cite
None of these are prompt problems. They're architecture problems, and the architecture that fixes them is a context engine. It's built in three steps.
Extract. When a document enters Starboard, it isn't treated as text. It's decomposed into typed datapoints: financial figures, property addresses, tenant names, lease terms, risk disclosures, Excel formulae. Starboard allocates attention the way an analyst would: a light pass over sparse text, heavy reasoning on dense tables and spreadsheets.
Connect. Those datapoints load into the context engine, a network of entities and the relationships between them. "ABC Industrial LLC" is a tenant. It appears in the rent roll, the lease abstract, and a risk disclosure, so Starboard connects them to each other, and to every relevant deal that came before. No analyst has to think to make the links.
Cite. Every fact keeps a provenance link to its source: the document, the page, the cell. When the system surfaces a finding, it can tell you exactly where it came from. "Tenant ABC's lease contains a co-tenancy clause." Click it, and the lease abstract opens at page 14 with the clause highlighted in the original PDF. Every claim arrives with its evidence attached.
This is true for all document types, including Excel. Provenance runs to the cell level, capturing not just a formula's output but the formula itself and the inputs it draws from, and no predefined template is needed to understand that column K is "Current Rent" and column M is "Lease Expiration." The system isn't just reading the rent roll. It's understanding it, and flagging when the numbers don't hold up.
Cross-Document Reasoning: Where Real Value Lies
That answers two of the three failures. Citations come from provenance. Memory comes from the context engine. Market understanding is built once the system reads across market sources and your entire firm's internal context together.
Take a portfolio acquisition across 12 assets. Each asset has its own OM, its own financial model, its own environmental report. A traditional workflow means 12 separate review processes, with synthesis happening manually at the end, if it happens at all.
The context engine treats all 12 assets as a connected dataset. Entities that appear across multiple documents are automatically linked. The same lender, the same environmental firm, the same tenant appearing in multiple assets becomes visible instantly. Patterns that would take weeks of manual review to surface, like shared structural risks, common counterparties, and correlated lease expirations, can be queried directly: "Which other assets share a lender with Riverside Logistics?" "Which tenants have lease expirations within 18 months across more than two properties?"
This is grounded situational reasoning. It's the difference between a system that answers questions and a system that discovers what questions to ask. The pattern matching that separates great investors, built deal by deal over decades, becomes something the whole firm can query. It turns the art of real estate into a science.
What Changes for the Team
Speed to conviction. The time between document receipt and initial underwriting conviction compresses from days to hours. The prescriptive work is done before the file is opened. Analyst hours go to critical thinking, not the checklist.
Risk surface coverage. Manual review has blind spots by design. There are only so many hours and so many analysts. The context engine leaves no stone unturned. It processes every document with rigor and flags relationships and risks.
Institutional memory. The context engine is cumulative, and it starts from what you already have: integrate it with your file storage systems and it builds a structured repository of every asset, entity, counterparty, and relationship your team has ever reviewed. New deals are evaluated against precedent automatically. "This rent roll looks similar to the structure we saw in the Denver industrial acquisition in Q3. Here's what we flagged then."
Scalability without headcount. The constraint in growing an acquisitions pipeline has historically been analyst capacity. A document intelligence layer removes that bottleneck. The same team can analyze more deals, in more depth, faster than they ever could before.
A System of Work, Not a Demo
Most AI products in this industry are demos dressed as products. They shine on a single PDF in a controlled setting and fall apart under real workloads: concurrent uploads, mixed file types, live deals with documents arriving out of order.
The distinction is between a tool you visit and a system you work in. A demo answers a question when someone thinks to ask. Starboard is where deals are triaged: it screens every deal, draws on your internal and external knowledge stores, and delivers the memo and the model with your assumptions already in them. That's a system of work.
Your Competitive Picture Today
The shift has started, and the early movers are pulling away. The difference shows up in what everyone else misses: the deal you never saw, the risk you caught late, the bid you lost. By the time the gap is visible in returns, it has been compounding for years.
Having AI is no longer the differentiator. Most firms do, in some form. The differentiator is whether that AI knows your firm.
Legal already ran this experiment. Two years ago, deal work and litigation looked like underwriting does today: smart people reading documents against the clock. The first firms to adopt purpose-built AI didn't just collect a marginal efficiency gain. They reset the pace of an entire profession, and their competitors are still recruiting against it. Real estate is standing at the same fork. The head start lasted about two years before it became table stakes. There's no reason to expect real estate to be more forgiving.
AI belongs in real estate investing. The open question is: does your system make your team better investors, or just faster ones?