Before EstateMate existed, Ari ran his family office’s portfolio in Excel. He had looked at property software. It would have cost more than half his staff, and probably only two or three people in the office would ever have used it. So Excel it was, with tenant lists spread across files and loan agreements spread across folders. And a tenant list is only worth as much as the lease and the last rent increase behind it.
Ari and Christof told that story at BUILTWORLD on 6 October. The webinar asked whether AI can run a multi-million portfolio better than Excel. Their answer was more careful than the title. Running the portfolio stays with the owner. What AI can do is take on the work that sits between the owner and their numbers, as long as it has the right foundation underneath.
Why a chat on its own falls short
The obvious move today is to drop the tenant lists and leases into a chat. Christof explained in the webinar why that breaks down on a large portfolio. Give a model 300 leases and it reads some of them, then answers as if it had read them all. Start a new chat and the overall picture is gone. A year or two ago, the models were also still happily making up figures.
Excel has the opposite problem. It calculates reliably, but only with what someone has typed in by hand and checked.
Excel
Calculates whatever you give it. A person enters the data and keeps it up to date.
An AI chat
Reads single documents quickly. With many leases it reads only some, and the context ends with the chat.
A platform with AI
Keeps the data, its history and the checking rules in one place. Every number leads back to its document.
Christof summed up what both are missing:
“AI works best when there’s a platform underneath it with the data already prepared, so it can simply draw on it.”
Christof, co-founder (translated from German)
That platform is what we built. It has three parts, each resting on the one before.
Every source in one place
The property manager knows the tenants and the payments. The bank knows its loans but not the portfolio. The tax adviser knows the annual accounts. Only the owner needs all of it together, and as current as possible.
So in EstateMate, properties, units and leases live in one database, history included. When a lease changes or an old tenant list is added, the rent history of every unit is kept. For that database to fill up, though, the documents have to be read first.
Documents that have been read and checked
In the webinar, Christof walked through the route a document takes before any figure from it lands in the database.
The system scans the document, recognises what it is, extracts the values and attaches it to the right unit. A second AI checks the result, including whether the values make sense. A person takes the last step: they see the document next to the extracted figures, with each source passage highlighted, and confirm.
We still require that review for every document. The AI classifies about 97% of documents correctly. For software that lives or dies on whether people trust its numbers, the remaining 3% is too much to leave to chance.
“Better to click once more than to read the whole contract yourself and type in the numbers.”
Christof, co-founder (translated from German)
For the difficult extraction work we use high-performance models hosted in Europe, and smaller tasks run on self-hosted models. When reading documents, open models were at times 20 to 30% behind the strongest ones, so at EstateMate the task decides which model does the reading.
Every number with its source
The review makes sure the numbers are right. For the owner to trust them, they also need to be able to check them. That’s why you can click any number EstateMate shows. If it’s a rent, the lease sits right next to it. If it’s a metric such as net operating income, a small button shows the whole calculation: the values from the database, the formula and the documents behind them.
What changes once the foundation is there
Ari described what happened at one client once their leases had been read in full for the first time. A commercial lease for a shop on a good street in Berlin had a break clause, and the window to use it closed in a month. Left alone, the lease would have run for another three years at €16/m². The client served notice and re-let the shop at €40/m².
The clause turned up because every lease was finally in one database. From there it takes one filter: every lease that ends in the next 24 months, with or without an option to renew.
The same foundation helps when looking ahead. Many loans from the low-interest years are coming to an end, and every refinancing starts with the question of what the portfolio will earn over the next few years. We model those scenarios on more than 90 parameters. Nobody has to invent the assumptions. The tool looks at how vacancy, for example, actually developed across the portfolio over the past five years and learns from those figures.
And on an acquisition, the foundation shows what the sales particulars leave out. A client wanted to buy a larger building with about 55 units and sent us 1,800 documents, including more than 35 stepped-rent leases. On paper the building looked good, with a healthy average rent and room to grow. The acquisition report was ready within hours. It flagged 16 flats let above the local rent index, two of them already in dispute under the rent cap (Mietpreisbremse), and most of the stepped rents had been raised beyond what the cap allows. At the planned price, the building was not a viable purchase.
The owner decides
Back to Ari’s spreadsheets. The work in them was preparation: merging lists, hunting for leases, reconciling figures. That’s the work AI can take on today, provided it sits on a clean foundation. Whether a building is held, sold or refinanced is still the owner’s decision. They just make it on numbers whose origin they know, with the time that used to go into preparing them.
