Guides
Freeholder or leaseholder dispute I'm thinking of buying I'm a director — what am I liable for? I own the freehold of a converted house I want to take control of our building The leasehold map All 60+ guides
LEASE-iQ
LEASE-iQ overview Sign up (free)For commercial leases →
Talk to us
We use LEASE-iQ to do the work for you
More
Pricing About Try LEASE-iQ free →
Last reviewed: 18 April 2026 · Spot something wrong? Tell us and we'll fix it.
“The Lessee shall pay [PAGE BREAK] 10% of...”
What an LLM does when a lease text is cut
The detail

Why generic AI fails on leases.

For directors, managing agents and leaseholders: three structural reasons ChatGPT, Gemini, and Copilot give confident but wrong answers about a lease, and why the tool you actually need is one designed for the job.

Information only, not legal advice. Always take professional advice before acting. England and Wales.

Before you blame the AI

The problem is not that AI is bad. It is that a general AI is the wrong tool for this job.

A lease is not a blog post. It is a formal document with a specific structure, sitting inside 40 years of statute and case law. Asking ChatGPT a lease question is like asking a sommelier to fix your boiler, not because they are stupid, but because the task is outside the training. This page explains the three places where the generic tools fall down, and what to do about it.

1. PDF parsing 2. Lease structure 3. Legal construct The consequence The answer References
Failure mode one

It can't read a PDF accurately. That isn't how LLMs work.

The most basic failure happens before the model has started reasoning. It happens at the point of reading the document.

Large language models do not read text the way you do. They are next-token predictorsA transformer-based LLM generates output one token at a time by predicting the most statistically likely next token given everything before it. It does not "look up" or "copy" text. Even when the input document is fully in context, the output is generated probabilistically. See Liu et al, Lost in the Middle, 2023., they generate the most statistically likely next word given everything that has come before. When that process meets a PDF, three specific things go wrong.

For a lease, that hallucination lands on figures that matter. Apportionment percentagesThe fraction of the total service charge payable by a flat, typically expressed as a percentage or a defined share. One wrong digit on apportionment can change a leaseholder's bill by hundreds or thousands of pounds per year.. Ground rent figures. Unexpired term. The £250 per leaseholderThe consultation threshold under Landlord and Tenant Act 1985, s.20 for qualifying works. If the cost to any one leaseholder exceeds £250, the landlord must consult. Miss the threshold and recovery is capped at £250 per leaseholder. s.20 threshold. The 18-month time barLTA 1985, s.20B. Costs must be demanded within 18 months of being incurred, or notified in writing with a statement that a demand will follow. Miss it and costs become unrecoverable through the service charge.. A single hallucinated digit changes the answer.

The cut-off problem

Leases are scanned. Scans have page breaks, tight margins, and sometimes text that gets clipped. When an LLM meets a sentence that ends mid-phrase, it does not stop, because stopping mid-phrase is statistically unlikely. It fills in the words that the pattern suggests "should" come next. Those words were never in your lease.

And there is "lost in the middle"Liu et al (2023) showed that LLMs perform markedly worse at retrieving specific information from the middle of long documents compared to the start or end. For a 40-page lease, the most commercially consequential clauses are often in the schedules near the end, or in definitions near the start, but variations, assignments, and side-letters are frequently buried in the middle.: the well-documented finding that LLMs reliably miss specific facts buried deep in long documents. A lease is exactly the kind of long document where the clause that matters is buried in Schedule 5 Part II.

None of this is a flaw that "prompt better" can fix. It is how transformer architectures work.

Failure mode two

It doesn't understand the ontology of a lease.

Even if a model could read the document perfectly, the next failure is conceptual. An LLM sees a wall of text. A lease is a structured object.

In computer science, an ontologyA formal representation of the concepts in a domain and the relationships between them. Knowledge graphs (Google's Knowledge Graph, medical ontologies like SNOMED CT, legal ontologies like LKIF) encode this structure so that machines can reason about "A is a type of B" or "C overrides D" rather than treating every sentence as flat text. is a formal structure that encodes the concepts in a domain and how they relate to each other. A lease has a rich ontology: a main body that incorporates schedules, schedules with parts and sub-paragraphs, definitions in clause 1 that govern every later clause, cross-references that mean clause 4.7(b)(iii) only makes sense alongside Schedule 2 paragraph 5, and deeds of variationA deed of variation is a supplementary deed that formally amends the original lease. It can change rent review mechanisms, permitted use, alterations, or add rights. Once executed, it becomes part of the lease for all purposes, but it is a separate document filed alongside. The original lease wording remains, but is overridden in specific places. that override the original wording years later.

A concrete example

Your lease, dated 1985, says at Clause 3.2 that the lessor insures. Schedule 5 Part II, added in a 1998 supplemental deed, says that where the building is converted to flats, the RMC insures. A 2019 deed of variation extends RMC insurance to include terrorism cover. The answer to "who insures?" depends on reading all three, in the right order, and understanding that the later document overrides the earlier one. An LLM reading these as flat text has no model of that hierarchy.

There are other ontological features a generic model has no representation of: the demised premises chainThe demise defines what the leaseholder owns (usually the inside skin of the flat: internal walls, floor coverings, non-load-bearing partitions) and what they don't (structure, roof, foundations, common parts). The boundary between demised and retained parts is the single most contested question in water-leak and repair disputes. Get the demise wrong and you get the repair liability wrong., which defines who owns what and who repairs what; the covenant chainWhen a lease is sub-let, obligations flow through: the head-lessee owes the freeholder, the under-lessee owes the head-lessee. The under-lessee is rarely in direct privity with the freeholder. This matters enormously for enforcement and for Section 20 consultation, which must reach the qualifying person. when a flat is sub-let; the distinction between covenants and regulationsCovenants are binding promises in the lease itself, enforceable via forfeiture or damages. Regulations are rules the landlord can make under a general clause (often "the landlord may make reasonable regulations"). Regulations are weaker, more easily challenged as unreasonable, and cannot extend beyond what the lease permits. A challenge that treats them as equivalent is a frequent AI error.; and reservationsRights reserved to the landlord (or kept back from what was demised), such as rights of access, rights to run services, or rights to build. Reservations are a separate schedule from the demise itself and are easy to miss. They are often the deciding factor in alteration consents..

To a general-purpose model, all of this is flat prose. To a director trying to answer a leaseholder's challenge, these distinctions are the case.

What happens next

Why directors and managing agents are seeing more confident-sounding letters.

The three failure modes don't just produce bad answers. They produce confident-sounding bad answers, at speed. That combination has a cost on both sides of the correspondence.

For directors and managing agents

Leaseholders arrive with AI-drafted letters that cite made-up clauses, fabricate apportionment figures, or misapply s.20. The prose sounds authoritative. The statutory references are correctly named but wrongly applied. Hours get burnt rebutting each letter. The relationship sours. The actual underlying issue, if there is one, gets lost in the noise.

See our page on AI-drafted leaseholder disputes for the response pattern.

For leaseholders, the risk runs the other way

Writing to your freeholder with a ChatGPT-drafted challenge that misquotes your own lease can:

  • Weaken your position on costs at tribunal, including on s.20CLTA 1985, s.20C. Allows a leaseholder to apply for an order preventing the landlord from recovering their legal costs through the service charge. The tribunal considers conduct on both sides, and a weak challenge based on fabricated clauses can make s.20C harder to secure. applications
  • Colour the relationship on later discretionary decisions like consents for alterations, sub-letting, or pets
  • Undermine a real grievance you may have had by burying it inside claims that do not survive scrutiny

Before you press send, check the clauses you are citing are actually in your lease. Our Talk-to-us service reads every clause and writes a brief that strengthens a challenge rather than weakens it.

The answer

What a lease-specific system looks like instead.

All three failure modes can be addressed. Not by a better prompt. By a different architecture.

For the PDF parsing problem: open-source OCR combined with LLM multimodal parsing, wrapped in a custom validation layer that double- and triple-checks the extracted text for structure, clause boundaries, and figures during parsing. The reconciled output goes into the knowledge graph; the raw OCR does not. Off-the-shelf models do the reading; the custom layer does the checking.

For the ontology problem: a knowledge graph that encodes clause hierarchy, schedules, cross-references, definitions, and variations as first-class structure, so the system can traverse them in the right order.

For the legal construct problem: UK leasehold statute and the governing case law pre-processed as rules that get applied to each answer, not hoped-for recall from pre-training.

And on top of all three: a cross-check layer that runs multiple independent responses, removes outliers, and flags disagreement honestly instead of resolving it silently.

This is what LEASE-iQ is built to do.

Open-source OCR plus LLM multimodal parsing, wrapped in a custom validation layer that double- and triple-checks during parsing. Knowledge graph of clause hierarchy. Pre-processed UK statute and governing case law. A multi-response cross-check designed to flag disagreement rather than resolve it silently when generating answers, with answers tied back to the specific clause and page in your own lease. Where the system can't reach confidence, it is designed to say so rather than guess.

How LEASE-iQ works →    Try it free →
References

Sources cited on this page.

All links verified at the date of last review shown at the top of this page. If a link has moved, please tell us.

On how LLMs read documents
  1. 1
    Liu et al (2023), Lost in the Middle: How Language Models Use Long Contexts. Stanford NLP. arxiv.org/abs/2307.03172 Peer-reviewed evidence that LLMs miss specific facts buried in long documents.
Statutes
  1. 2
    Landlord and Tenant Act 1985 (LTA 1985), especially ss.18, 19, 20, 20B, 20C, 21, 27A. legislation.gov.uk/ukpga/1985/70
  2. 3
    Leasehold Reform, Housing and Urban Development Act 1993 (LRHUDA 1993). Collective enfranchisement. legislation.gov.uk/ukpga/1993/28
  3. 4
    Commonhold and Leasehold Reform Act 2002 (CLRA 2002). RTM and lease extension. legislation.gov.uk/ukpga/2002/15
  4. 5
    Leasehold and Freehold Reform Act 2024 (LFRA 2024). legislation.gov.uk/ukpga/2024/22
  5. 6
    Building Safety Act 2022, especially Schedule 8 (leaseholder protections). legislation.gov.uk/ukpga/2022/30
Case law
  1. 7
    Daejan Investments Ltd v Benson [2013] UKSC 14. Supreme Court, s.20 dispensation. supremecourt.uk (UKSC-2011-0057)
  2. 8
    Phillips v Francis [2014] EWCA Civ 1395. Court of Appeal, s.20 threshold applied per set of works. bailii.org
Advisory bodies
  1. 9
    LEASE (Leasehold Advisory Service). Government-backed, free initial advice. lease-advice.org
  2. 10
    HMCTS First-tier Tribunal (Property Chamber). The tribunal that hears service charge and leasehold disputes. gov.uk/courts-tribunals/first-tier-tribunal-property-chamber

Ready for a tool that was built for this?

LEASE-iQ is designed around the three failure modes above, not around a better prompt.

How LEASE-iQ works → Try it free →

Stay ahead of leasehold compliance

Plain-English updates on what matters for leaseholders and directors. Written by people running their own building at Hafer Road.

No spam. Unsubscribe anytime.

Next steps

Four ways to take this further.

Free to read on. Free to test against your lease. Free to ask the bot. Or paid, if you want us to write the letter for you.