Nobody in your management team has read all three hundred leases. That is not negligence - each one runs to dozens of pages of rent, indexation mechanisms, options, guarantees, break points and one-off exceptions negotiated eight years ago that nobody remembers today. So when a simple question lands - in how many of our leases does the tenant have a break point in the next 18 months? - the honest answer is three weeks of analyst work, while the decision riding on it, sell or refinance or hedge, waits. An AI pipeline that reads documents has the answer ready by morning.
From binders to a single table
Lease abstraction is the most mature use of AI in income-producing real estate, and it is no accident that it shows up in global surveys among the five AI use cases with the most proven returns, at an average ROI of 171% across the portfolios measured. The mechanics are easy to describe. A PDF goes in; a structured row comes out - tenant, area, rent, indexation mechanism, option dates, guarantees, break points, unusual clauses. Three hundred leases become one table with alerts attached: which option opens next quarter, which guarantee is about to expire, which lease drifts from the company's standard language and needs a lawyer's eye.
Over my career I have built exactly these pipelines: a system that works through hundreds of documents overnight and returns the table, the exceptions and the alerts by morning. The marginal cost is one night of compute. The alternative is weeks of analyst and lawyer time, and even then with no forward-looking alerts.
How you build it matters. The system extracts and flags; a person approves. Extraction accuracy on standard fields is very high, but judgment on the exceptions stays with the lawyer. And that table earns its keep well beyond the asset management department - it is exactly what the bank and the appraiser ask for in every refinancing, and what the data room needs in every transaction.
Sales per square meter and tenant mix: the percent that is worth millions
The arithmetic here is brutally simple. Take an asset generating NOI of NIS 100 million a year. Every percent of improvement is NIS 1 million a year, and at a 7% cap rate that is roughly NIS 14 million on the value of the asset. Which is why a small operational gain across a large portfolio is worth more than most of the acquisitions sitting on the table.
Where does AI find that percent? In data you already own: tenant sales in the malls, footfall counters, seasonality, sales per square meter benchmarked against comparable tenants in comparable assets. The system surfaces what an experienced eye knows but never has time to work through asset by asset - a weak tenant in a strong location, a category that pulls traffic for its neighbors versus one that only feeds off them, a lease renewing at a rate the market has left behind. What management gets is not another dashboard. It is a list of recommendations ranked by money, in hand before each renewal negotiation. And as in every system I build, the AI recommends and the asset manager decides. Each recommendation arrives with the data underneath it, so you can test it, argue with it, and correct it.
Predictive maintenance: fix it before the tenant calls
Elevators, HVAC, pumps, generators. An emergency call costs several times what preventive service costs, and the bigger cost is what it does to the satisfaction of the tenant paying the rent. The good news is that you do not have to begin with an expensive sensor project. The service call history already sitting in your systems - which unit, in which asset, how often, at what cost - is excellent raw material for a model that ranks which systems are likely to fail in the coming months and where preventive work pays for itself in advance. For property managers that becomes a quarterly work plan instead of firefighting.
Take an office asset with twenty elevators. Preventing ten emergency calls a year is not only the direct saving - it is the difference between a tenant who renews quietly and a tenant who starts taking calls from other landlords. In malls it counts double, because an elevator out of service on a peak day is lost sales for every tenant on that floor, and the compensation clauses are already written into the leases.
The window is open because the sector is quiet
Here is the strategic point. While the banks and the insurance companies in Israel announce AI centers loudly, Israeli income-producing real estate is almost entirely silent in public. At the same time, the State Comptroller report of June 2026 found that 68% of AI projects in the economy are stuck in pilot and never reach deployment. Put those two facts together and you have an opening: in a sector where nobody has moved yet, the one company that actually deploys - not announces, deploys and measures - buys real differentiation with tenants, with investors, and with the analysts who cover it.
A pilot like this, in my experience, is a matter of weeks rather than quarters. It also stays clear of the core systems: it reads documents and data you already hold and returns its output alongside them - no risky integration, no dependence on a single large vendor.
Starting does not take a big strategic decision. It takes one narrow pilot: abstracting the lease file of a single asset, with the measures agreed in advance - how many work hours were saved, how many exceptions and alerts were caught that would otherwise have been missed. From there you widen to sales, to tenant mix, to maintenance.
Want to see this on your own portfolio? Send me ten leases, redacted is fine, and I will send back the table and a demo.
