How AI-assisted review cut a portfolio's lease diligence from three weeks to four days

Start with the problem. What did diligence on a commercial real estate portfolio actually involve before this?

On a portfolio acquisition, the lease file is the deal. A buyer is paying for a rent stream governed by twenty, thirty, sometimes fifty separate contracts, and any one of them can contain a term that changes what the asset is worth. Someone has to read every lease and pull the material points: commencement and expiration, renewal options, rent escalations, co-tenancy conditions, exclusive-use clauses, assignment and change-of-control provisions, and landlord obligations that survive the sale. That gets reconciled against the seller's rent roll, which is frequently wrong. In parallel we review recorded documents: the CC&Rs, easements, and use restrictions that run with the land. Historically this was associate-intensive and slow. On a tight timeline you either staffed it heavily, which the client paid for, or you sampled and accepted risk. Neither is a good answer. The work is also monotonous, which is precisely when attention drifts and a clause slips through. High stakes paired with repetitive extraction is what made it a candidate for a different approach.

Extraction software has existed for years. Why is this viable now, and not five years ago?

The previous generation of extraction tools was rules- and template-based. They worked when a document matched a pattern and struggled the moment a lease was drafted in someone's own style, which is most of the time. Commercial leases are not standardized. A co-tenancy provision can be written a dozen ways, and older tools missed the ones that did not fit their template. What changed is that current language models read documents closer to the way a trained reader does, by meaning rather than exact phrasing. The model recognizes a renewal option or an exclusive-use restriction even when the drafting is unusual. On standard commercial lease terms, the accuracy we have seen in testing sits in the low-to-mid nineties, high enough to be genuinely useful as a first pass. I want to be careful, though. That figure describes standard terms. On unusual provisions, historical easements, layered covenants, bespoke drafting, the tool is weaker, and we treat its output accordingly. The shift is not that the machine became infallible. It became good enough to be worth verifying.

Walk me through how the system actually works on a live matter.

We load the lease set and the recorded documents into a controlled environment. The system produces a structured abstract for each lease against a defined field list: parties, premises, term, options, rent and escalations, operating-expense treatment, co-tenancy, exclusives, assignment, and default provisions. Those abstracts roll up into one normalized rent roll, which we compare against the seller's version to surface discrepancies. Separately, the system reads the recorded restrictions and title exceptions, categorizes them, and assigns a preliminary risk level based on how each item interacts with the buyer's intended use. High- and medium-risk items go to the top of an attorney review queue. A lawyer works through the queue, checking each flag against the source document and spot-checking a share of the low-risk items and the abstracts the model was most confident about. The output the client sees, the diligence memo and the issues list, is written and signed by an attorney. The model handles the extraction and the first sort. The legal judgment, what a finding means for the deal, stays with us.

This is privileged, confidential client material. How do you handle data governance and the ethics rules?

We treat it as client-confidential, privileged material, because it is. The governing guidance is ABA Formal Opinion 512, which makes clear that a lawyer's duties do not change because a tool is involved. Rule 1.6 requires reasonable efforts to protect confidentiality, Rule 1.1 requires enough technical competence to use the tool responsibly, and Rules 5.1 and 5.3 extend supervision to its output. In practice that meant real vendor diligence before any client document went near the system: where data is processed, whether client content trains the model, how long it is retained, and who can access it. We use an arrangement where our data is not used for training and stays within a controlled tenant, with access restricted by matter. We also handled client consent directly rather than relying on a line buried in an engagement letter. For a sophisticated institutional buyer, that conversation is straightforward and usually welcome. The point I make to partners is that adopting the tool did not create a new confidentiality regime. It applied the existing one, with more diligence on the vendor than we were used to.

Who is accountable when the model is wrong, and how do you catch it?

The accountable party is the attorney whose name is on the work, without qualification. The model produces a draft; it does not sign anything and it does not exercise judgment. Our validation is structured because the failure mode has changed. A few years ago an AI error was usually obvious. Now the errors that survive are confident and plausible: a clause summarized in a way that reads correctly but quietly drops a condition. Those do not jump out on a casual read, so we do not rely on one. Every high- and medium-risk flag is checked against the source. We sample the low-risk items and a portion of the abstracts the model scored as high-confidence, because confidence is not correctness. The reconciled rent roll gets a second set of eyes. And we document the review: what the model produced, what the attorney verified, what changed. That record matters for our own quality control and, if it ever came to it, for defensibility. The discipline is to verify by process rather than by impression.

How did the team take to it, and what went wrong first?

There was skepticism, and it was reasonable. Two things went wrong early. First, we had to calibrate trust. On one of the first matters, a couple of people treated the abstract as nearly final and gave it a lighter read than they should have. Nothing reached the client, but it showed us that a polished-looking output invites over-reliance. We corrected by being explicit that the abstract is a first pass, and by building verification into the workflow rather than leaving it to discretion. Second, we underestimated document quality. Some leases arrived as poor scans, and the extraction inherited the errors in the underlying text. We fixed that by tightening intake and checking legibility before processing. On staffing, some associates worried this erased the work that builds their judgment. The honest answer is that it removes rote extraction, which no one enjoyed, and moves them into the analysis sooner, reading the flagged issues and weighing deal impact. That is the more valuable skill. Framing it that way, and training people on where the tool is weak, is what turned it around.

What did it change on that portfolio acquisition, in concrete terms?

On the acquisition I keep coming back to, we abstracted roughly thirty commercial leases with the tool assisting, alongside the recorded-restriction review. The part of diligence that had been running about three weeks came down to roughly four days of elapsed time, and we closed on schedule with full legal rigor, meaning every material finding was attorney-verified. Let me frame those numbers with care. This was one matter, and the document quality was reasonably good, which helped. The four days is elapsed time, not effort reduced to nothing: there were still meaningful attorney hours in the review queue, which is where they belong. The gain was not that we eliminated the work. It was that we moved hours from manual extraction, which is low-value and error-prone, to legal analysis and strategy, which is what the client is paying for. On a deal with a hard closing date, compressing the mechanical phase also removed the usual late-stage scramble. I would not present a single matter as proof of a fixed percentage. I would say the direction and the magnitude were clear.

What would you say to a skeptical peer, or to a client nervous about AI touching their deal?

To a skeptical peer, I would say the real exposure is not adopting the tool, it is using it without controls and treating an unverified output as work product. That is where the malpractice and privilege problems live, and it is avoidable. Used with verification and proper vendor diligence, this reduces the drudgery that causes human error rather than adding to it. To a nervous client, I make three points. Nothing the model generates reaches you unread; an attorney reviews and signs everything. The tool decides nothing about your deal; it extracts and sorts, and we interpret. And here is specifically how your confidential information is handled, including that it is not used to train anyone's model. On limitations, I am direct. The tool is weaker on unusual or historical provisions, precisely the items that tend to matter most, so those get more human attention, not less. Its output is only as good as the documents you feed it. And it does not replace judgment about what a finding means for the transaction. I do not sell it as a machine that does diligence. I describe it as a way to spend our time on the parts that require a lawyer.

Give me a concrete moment where it mattered.

One example stands out, and it makes the case better than any statistic. Among the recorded documents was a use restriction limiting a category of tenant use on part of the site. On its own it read as routine. But the buyer's plan for that parcel depended on a use the restriction arguably constrained. The system flagged the covenant as a medium-risk item and surfaced it in the first sorted pass, rather than in week three, when a manually assembled review might finally have reached it. That timing mattered. Because it came up early, we could analyze how the restriction interacted with the intended use, raise it with the client while there was room to negotiate, and price the risk into the deal instead of discovering it after closing. I want to be precise about the credit. The tool did not interpret the covenant or decide what it meant. It made sure a human saw the right document at the right time. On a portfolio with dozens of recorded instruments, getting the consequential item in front of a lawyer early is most of the value.

What does this mean for the profession and the client relationship?

Two shifts follow, and neither is only about technology. The first is commercial. If a phase of diligence that used to justify a block of associate hours now takes a fraction of the time, the hourly model comes under pressure, and clients are right to ask about it. I think this pushes real estate work toward defined-scope or value-based pricing for the mechanical parts, with the premium attached to judgment and advice rather than to hours spent extracting data. That is a healthier alignment. The second shift is about what distinguishes a firm. When capable tools are widely available, the differentiator is not which model you license; it is the quality of your verification and the soundness of your judgment on the hard questions. Diligence, quality control, and defensible review become core competencies rather than afterthoughts. For the client relationship, that is a good outcome. We spend less time as a document-processing cost center and more as advisors on the risks that actually affect the transaction. The technology handles the reading it is suited to. The relationship rests on the thinking it cannot do.