AnalysisReal Estate

Underwriting at Machine Speed

AI is compressing site selection, underwriting and lease abstraction from weeks into days. What should change at the investment committee, and what must not.

Arclight Research5 min read

Fig. I · Real Estate

In brief

  1. ISplit underwriting into extraction, estimation and judgment. Automate the first, discipline the second, keep the third human.
  2. IIFaster screening helps only if the investment committee narrows the funnel faster. Otherwise it sees more deals, less carefully.
  3. IIIRequire provenance: every number in an IC memo should trace to a source document or a named, owned assumption.

Real estate has always been a document business. An acquisition arrives as a rent roll, a trailing twelve-month operating statement, a stack of leases and amendments, a title report, a zoning code and a model, assembled by analysts working late. AI can now do much of that assembly in hours rather than weeks.

That changes the economics of the deal team and what the investment committee should expect to see. It does not change who is accountable for the decision. The firms that gain most will redesign the committee as deliberately as the analysis that feeds it.

Where the hours are going

In site selection, geospatial machine learning screens parcels against zoning, flood and heat exposure, access, demographics and, increasingly, available power, which now gates data center and many industrial sites. Language models can read a municipal code and summarize years of planning commission minutes in an afternoon. They can also misread a setback table or miss the overlay district buried in an amendment, so a model’s entitlement summary is a brief for land-use counsel rather than a finding.

Lease abstraction is the clearest gain. Document intelligence extracts rent schedules, renewal options, exclusives, co-tenancy clauses, expense caps and termination rights across hundreds of leases, and it performs well on standard fields. Accuracy falls on bespoke clauses and amendment chains, where a later side letter quietly overrides an earlier renewal option. The control is a sampling protocol, with people reading the leases that carry the most rent and the most unusual terms.

In construction, computer vision on site cameras and drone imagery measures installed progress against the schedule, and models trained on past projects can flag change-order and delay risk earlier than monthly reporting. Owners and lenders can check draw requests against physical evidence instead of a contractor’s percentage complete.

In asset management, the applications are less visible and often more valuable: model-predictive control of building systems to cut energy cost and meet building performance standards such as New York City’s Local Law 97, renewal-probability scoring, tenant credit monitoring and operating-expense benchmarking across the portfolio. This is where post-close data accumulates, and that data is the raw material for the next underwriting.

Speed is only an advantage if the investment committee gets better at saying no.

Extract, estimate, judge

The useful way to think about AI in underwriting is as three layers, each with different rules. Confusing them is how firms end up automating judgment while hand-checking facts.

Extraction

Facts pulled from documents: rents, dates, areas, expense lines, clause terms. This layer should be automated aggressively and verified by sampling, because the facts are checkable and the errors are findable.

Estimation

Forecasts: market rent growth, downtime between leases, renewal probability, capital expenditure and exit cap rate. Models can propose these from comparables and history, but every assumption needs a named owner and a range. Models also learn the regime they were trained on, and one fitted mostly to years of cap rate compression learned a market that ended when the cost of capital rose sharply in 2022 and 2023.

Judgment

The thesis, the partner, the structure and the price. Whether a submarket is early or late in its cycle, whether a sponsor will behave well when a deal goes wrong, how hard to push for a guarantee: these are human calls, and they are where the firm earns its fees.

Pricing deserves particular care. Algorithmic rent-setting tools that draw on competitors’ data have drawn antitrust litigation in the United States, so any model that recommends rents should be reviewed by counsel for how its inputs are sourced and used.

What changes at the investment committee

The first change is volume. When screening becomes cheap, deal teams bring more opportunities forward, and a committee with the same calendar gives each one less attention. Speed is only an advantage if the investment committee gets better at saying no. Four practices help.

  1. Narrow the funnel earlier. Set explicit kill criteria that the machine screen applies before a deal reaches a person, and review those criteria every quarter.
  2. Require provenance. Every number in the IC memo should trace to a source document or a named, owned assumption, and the memo should mark which is which.
  3. Spend the saved time on the downside. When the base case is cheap to build, expect more work on stress cases: refinancing at higher rates, a tripped debt service coverage covenant, an anchor tenant leaving, an entitlement delayed by a year.
  4. Close the loop. Compare underwriting assumptions with actual performance across the portfolio, assumption by assumption, so the firm learns which of its forecasts are consistently wrong.

The fourth practice is the easiest to skip and the one AI makes cheapest. Abstracted leases and asset-level operating data, captured consistently after closing, turn a portfolio into a back-testing data set for every future acquisition.

Where judgment must stay

Some decisions should remain human by design, whatever the models can do. Four stand out:

  • The thesis. Whether a market, a property type or a vintage is mispriced. Models extrapolate, and the best investments are often the ones where history is a poor guide.
  • Counterparties. Sponsors, joint venture partners, contractors and anchor tenants are judged on character and incentives as much as on financial statements.
  • Entitlements. Approvals often turn on relationships with planning staff, elected officials and neighbors, none of which a model observes.
  • Accountability. Fiduciary duty sits with the people who vote, and it cannot be handed to software.

A committee that cannot explain a decision without pointing at a model has delegated something it was not entitled to delegate. The standard to hold is that the deal lead can defend every approved investment in plain language, with the model as evidence and never as the argument.

Questions for your next leadership meeting

  1. If our screening capacity doubled tomorrow, would our investment committee see better deals or only more of them?
  2. Can every figure in our last three IC memos be traced to a source document or a named assumption owner?
  3. Which of our underwriting assumptions has been most consistently wrong over the past five years, and how would we know?
  4. Where do our pricing and leasing tools source their data, and has counsel reviewed them?
  5. Which decisions have we agreed must always carry a human signature, and is that written down anywhere?

ARCLIGHT · RESEARCH

This analysis is general in nature and does not constitute investment, legal or engineering advice. To discuss how it applies to your organization, request a private briefing.

Begin

Apply this analysis to your own assets.

Tell us about your assets, your board and the decision in front of you. A principal replies within one business day.