A lease can be long, yet a property manager's team needs only part of its information day to day: who the tenant is, when the lease starts and ends, how much rent is paid, how the rent rises and by when someone has to be notified. Without a tool for it, somebody retypes that information by hand from a PDF into a spreadsheet. The same goes for addenda, handover reports and tenants' insurance policies.
AI models can pull this data out of a document faster than a person. They can also get it wrong, and not always in a way that catches the eye. So the more useful question is not whether AI will replace someone in the office, but how to organise the work so that AI proposes and a person approves.
We write as a small firm that builds bespoke systems. In this sector we have one project behind us: a system for managing a commercial real estate portfolio with a module that processes PDF documents using AI. The lease reading described below is something we offer to build, not a description of that system and not an off-the-shelf product.
What AI can extract from a lease
Leases differ in layout, but most contain the same kinds of information. That makes them a good task for AI: the model does not have to "understand" the whole lease, it only has to find specific values and show where it took them from.
- Parties: landlord, tenant, their registration details (NIP, the Polish tax number, and KRS, the court register number) and the people signing for them.
- Premises: building, unit, floor area, parking spaces.
- Term: signing date, handover date, expiry date, renewal options.
- Rent and charges: rate, currency, how it is calculated (per square metre or a flat sum), service charges.
- Indexation: the index the rent is linked to, and the date and frequency of reviews.
- Security: deposit, bank guarantee, the date by which they must be renewed.
- Notice periods: how far in advance the tenant or landlord must give notice to renew or end the lease.
From an addendum, AI can extract only what the addendum changes, and from a tenant's insurance policy the insurer, the sum insured and the end date of cover. The principle stays the same: data goes to review first and only then into the database.
Where a person cannot be left out
AI handles values that are stated plainly in the document well. It does worse with values that follow from several places at once or depend on interpretation. Here a human check is not a formality but a condition.
- Amounts and dates. One transposed digit in the rent or the expiry date changes the reports for the whole portfolio.
- Indexation clauses. The review formula is sometimes written out in words, with exceptions and thresholds, and a condition can get lost in the AI's summary.
- Addenda. An addendum (in Polish, aneks) changes part of the lease, so someone has to decide which value applies today.
- Unusual terms. Rent-free periods, rights of first refusal, contractual penalties, clauses added to the standard template.
- Poor-quality documents. Scans, handwritten notes, stamps over the text.
The final say stays with a person. This is not caution for show: the firm, not a language model, is responsible for the data in the portfolio.
The process: AI proposes, a person approves
Below is how we would design such a process. The details are settled case by case, because they depend on who in the firm is responsible for leases.
- A staff member adds the PDF of a lease or addendum to the tenant's file.
- AI reads the document and fills in a draft set of fields. Next to each field it stores the passage and page number the value came from.
- The draft is marked "to review". Nothing reaches the reports or the deadline calendar yet.
- The reviewer sees the reading and the original side by side. They correct wrong fields, fill in missing ones and approve.
- Only approved data feeds the tenant record, deadline reminders and portfolio reports. The system records who approved the reading, when, and what they corrected.
The review screen has to be convenient. If comparing the reading with the original means opening two windows and hunting for the page, people will start approving without looking, and then the whole check becomes a fiction.
We use existing AI models and do not train our own. Which AI services we use and what data reaches them, we agree with you before deployment, because leases contain tenants' data.
Reading documents is not the same as an assistant that talks to people. If the system is also meant to answer tenants as an AI assistant, then from 2 August 2026 the person talking to it must be told so (Article 50 of EU Regulation 2024/1689, as of 21.10.2026). This is not legal advice; it is worth checking the details of your situation with a lawyer.
What to do with errors and incomplete PDFs
Errors will happen. The question is whether the system lets you spot them before they reach a report.
An empty field instead of a guess
When a piece of information is missing from the document, AI should leave the field empty and flag it clearly, not fill in the "most likely" value. An empty field is visible at once; an invented value is easy to miss.
Scans and missing pages
A scan with no text layer, a cut-off page, a lease without its schedules: the system can mark such a document as incomplete and ask for a better copy instead of forcing a reading. Some documents will still have to be retyped by hand, and that is fine.
Corrections as a signal
Every correction the reviewer makes shows where AI goes wrong. If the same fields keep coming back with errors, the instructions for the model or the validation rules can be changed, for example by adding a check that the expiry date falls after the start date.
Addenda and versions
An addendum does not overwrite the lease blindly. The system shows which fields it changes, and a person decides from when the new value applies. The history is kept, so you always know where today's rent came from.
Where the data goes after approval
Reading a lease only makes sense when its data is used further on. In a property portfolio system that could be the building and tenant records, reminders about notice and indexation dates, portfolio reports and an Excel export for those who work with spreadsheets. Data entered once does not drift apart across copies.
The same single-source principle comes back in other topics in the sector, for example in price transparency for housing developers. Leases are not the only documents that land on a property manager's desk either: we write about supplier invoices that a housing community (wspólnota mieszkaniowa) receives through KSeF, the Polish national e-invoicing system, in KSeF for housing communities and property managers.
Where to start: a checklist
- List the fields you actually use in reports and the deadline calendar. Not everything in a lease has to go into the system.
- Gather a dozen or so sample leases and addenda, including the hardest ones: scans, old templates, leases in English.
- Decide who approves readings and who settles doubts about indexation clauses.
- Decide which fields are mandatory before approval and which may stay empty.
- Agree what happens to a document AI cannot read: who retypes it and by when.
- Settle which AI service you use and what data reaches it.
- On a trial set, compare the AI reading with a manual one before connecting it to reports.
Let's talk about your documents
If someone in your firm retypes data from leases, addenda or insurance policies every week, describe that process to us in a few sentences. We will reply with where it makes sense to start and arrange a free consultation.
We describe our project for a commercial real estate portfolio in the case study: 11 weeks of work from June to August 2026, four stages, each closed with a formal sign-off, and at the end user training and handover of the code. You can put together an approximate cost of the modules, including AI document reading, in the approximate price builder. It is approximate; the quote is always individual.
Sources
- Regulation (EU) 2024/1689 of the European Parliament and of the Council (AI Act), Articles 50 and 113, checked 21.10.2026
- Regulation (EU) 2026/1744 (Digital Omnibus), checked 21.10.2026
- ITCompass: a system for managing a commercial real estate portfolio, checked 21.10.2026
