AI for lease management: what it handles, and what still needs a human

The short answer. AI is good at reading a lease and turning it into structured data: parties, term, rent, review dates, break conditions, service charge provisions, repairing obligations. It is not good at deciding what to do about any of them. Use it to build the register you have been meaning to build for three years, then have a person check every date that carries a deadline, because a break clause missed by a week is worth more than every hour the tool ever saved you.

I read a lot of leases in commercial property before I left to build businesses. The thing I keep having to say to teams is that abstraction and management are two different jobs, and the software industry sells you the first while implying the second.

Abstraction and management are not the same job

Abstraction turns a document into data. Management is doing the right thing with that data at the right time. Vendors sell the first and let you infer the second.

Lease abstraction is a reading task: work through a hundred and twenty pages and produce forty structured fields. It is repetitive, slow, expensive when a person does it, and genuinely well suited to AI.

Lease management is a judgement task. Is the condition attached to this break exercisable in practice? Does time run from service or receipt? Is this a trigger notice that has to be served, or a date that passes on its own? Has anything happened on the ground that changes what the clause means? None of that is in the document as a field. It sits in the interaction between the document, the facts, and a body of law about how these clauses get read when they are argued about.

Software that extracts a break date and puts it on a dashboard has done the reading. It has not done the thinking. Buying it and assuming otherwise is how firms end up worse off than they were with a spreadsheet, because a spreadsheet nobody trusts gets checked, and a dashboard everybody trusts does not.

Where AI genuinely earns its place

On the volume reading nobody has time to do properly.

Building the register from scratch. Most firms have an incomplete lease register and no realistic route to finishing it. Running a portfolio of documents through AI to produce a first-pass structured register is the single biggest win here. Incomplete and unchecked beats non-existent, as long as everyone knows which it is.
Answering questions across a portfolio. "Which of these leases have a tenant break in the next eighteen months?" "Which ones put external repair on us?" Questions that used to mean a week of reading now take an afternoon of checking.
Summarising one lease for someone who needs the gist. A new manager taking over a building, an agent briefing a client, a director who wants the shape of a deal before a meeting.
Comparing a draft against your standard form. AI is reliable at spotting what differs between two documents, which is a narrower and safer task than interpreting either of them.
Drafting the routine correspondence that follows from a lease event, once a person has decided what the event is and what should happen.

Notice these are all reading and drafting. The moment the output is a decision or a deadline you rely on, a person comes back into it.

Why "99% accurate" is the wrong question

Because it hides the only thing you need to know, which is what the errors look like when they happen.

Vendor pages advertise accuracy figures with no stated methodology, no independent benchmark, and no definition of what counts as an error. Treat those numbers as marketing until someone shows you the test. More usefully, the headline rate is beside the point. A tool that is right about ninety-nine fields in a hundred is fine if the misses are spread across low-consequence fields, and dangerous if the misses cluster on conditional break clauses, because that is exactly where a mistake is unrecoverable.

Ask a supplier these instead:

On which field types does it fail most? Anyone who has measured their own product knows the answer. Anyone who has not will change the subject.
Does it flag its own uncertainty? A system that says "I am not sure about this one" is worth several that quietly guess. Confidence scoring per field is the feature that matters and the one nobody markets.
Does it link every extracted field back to the clause it came from? If a checker has to hunt through the document to verify a date, nobody verifies anything after week two.
What happens when the lease has a side letter, a deed of variation or a licence to alter? Real leases come with a pile of accompanying documents that change what the lease says. Extraction from the lease alone gives you a confident, tidy, wrong answer.
Who is liable when it is wrong? Read the contract. The answer is you.

The failure mode to design against

Confident, plausible, wrong. That is the whole risk in one phrase, and it is different from the risk a spreadsheet carries.

When a junior abstracts a lease badly, the output usually looks bad. Fields are blank, notes are hedged, someone senior notices. When AI abstracts a lease badly, the output looks exactly like a good abstraction. Every field is populated, the formatting is clean, the dates are plausible. There is no visible signal of the one field that got read from the wrong clause.

Design around that, and it stops being frightening:

Two-tier the fields. Anything driving a deadline, break dates, review dates, expiry, notice periods, notice service provisions, gets checked against the document by a person, always. Everything else is checked on sample. Trying to verify every field equally means nobody verifies anything.
Keep the source in view. Every critical field should carry a clause reference so a check takes seconds rather than minutes.
Diary from the checked layer only. The unchecked register is a research tool. The diary is a professional obligation. Never let the first feed the second automatically.
Date the check. Who verified this field, and when. Six months on, that is the difference between a register you rely on and one you have to redo.
Re-check on event. Before anything happens that depends on a date, someone opens the actual lease. Not the register. The lease.

What this looks like on a Monday

Start with one building, not the portfolio.

Take a single property with a manageable number of leases. Run the documents through whatever AI your firm has already approved. Produce the register. Then have someone who knows the building check it properly and record how long that took and what it got wrong. You now have two things nobody in your firm has: a real error profile for your own documents, and an honest estimate of the checking cost.

That exercise takes a day and settles arguments that otherwise run for months. It also tells you whether you need to buy anything at all. Plenty of firms find that the general-purpose AI tool their business already pays for handles the reading perfectly well, and that what they lacked was a decision about who owns the register.

Frequently asked questions

Does AI replace lease abstraction teams? It replaces most of the typing and none of the checking. The realistic shape is fewer hours spent producing a first draft of the data and more hours spent verifying the fields that matter, which is a better use of an experienced person than reading page ninety of a lease looking for a rent review clause.

Is it safe to upload leases to an AI tool? With an approved business-tier tool and a clear answer on data handling, yes, and plenty of firms do. With a free consumer account and a client's confidential document, no. Get the tooling decision made properly and in writing before anyone starts, because this is the point at which a well-meaning person creates a problem.

What about older or badly scanned leases? That is where extraction degrades most, and it is rarely mentioned in a demo. Demos use clean documents. Your archive has photocopies of photocopies with handwritten amendments. Test on your worst documents, not your best.

Should the diary run automatically off the extracted dates? No. Let it run off the checked layer, with a named person accountable for the check. Automating from unverified extraction is the one design choice in this whole area with a genuinely bad downside.

Do I need specialist lease software, or does general AI do it? Test the general tool first, on one building. Specialist software earns its money on scale, integration and workflow rather than on reading, and you will negotiate better once you know what the reading alone is worth to you.

What is the biggest mistake firms make here? Trusting the output because it looks tidy. The second biggest is treating the whole thing as an IT project rather than training the people who will live with the register.

If you want your team trained to use this well, and the checking layer designed so the register is worth relying on, book a session.