AI lease reading amendments is not one model reading one file. A single tenant relationship is rarely a single document. It is a base lease, a first, second, and third amendment, an estoppel, and one or more side letters, and the number that belongs in the rent roll is whichever version of a term survived the last document to touch it. The hard part is not optical character recognition. OCR was solved years ago. The hard part is reconciliation: deciding which clause controls when four documents say four different things about the same rent.
Key Takeaways
A commercial lease runs 50 to 200 pages, and its abstract compresses to a structured set of roughly 30 to 80 fields. The compression is where the judgment lives.
The engineering problem in reading a long lease is not extraction but reconciliation: an amendment or side letter can supersede the base lease, and the controlling value is the last one standing.
LLM accuracy on long documents follows a U-shaped curve. Liu et al. (2023) found retrieval accuracy degrades by more than 30% when the relevant passage sits in the middle of a long context rather than at the edges.
A side letter that is not indexed to the lease it modifies is invisible to the model. If the document is not in the set, no amount of accuracy on the base lease recovers the term it changed.
The reliable architecture reads each document as a dated layer, resolves conflicts by recency and specificity, and cites every final value back to the exact page and document that governs it.
What Actually Makes a 200-Page Lease Hard to Read?
The difficulty is not the page count. It is that a lease is a stack of documents written at different times that overwrite each other in ways the pages do not announce. A base lease sets a rent. A second amendment changes it. A side letter quietly waives it for eighteen months. The correct answer is the survivor, and finding it means reading all four and knowing which one wins.
Modern language models handle raw length. Context windows now run to hundreds of thousands of tokens, enough to hold a 200-page lease and its amendments at once, per IBM's technical overview. Fitting the document in memory was the old constraint and it is gone. The current constraint is reasoning across the set: resolving a defined term introduced on page 12 and used on page 140, reconciling a base rent against three amendments, and recognizing that a side letter executed the same day as the lease can override a clause the lease itself asserts.
This is why an abstract is compression, not transcription. A lease of 50 to 200 pages typically collapses to roughly 30 to 80 structured fields. Every one of those fields is a decision about which version of a term is live. Transcription copies. Abstraction decides.
How Does AI Reconcile a Base Lease Against Its Amendments?
Reconciliation works by treating each document as a dated layer and resolving conflicts by recency and specificity. The base lease is the foundation. Each lease amendment is a diff applied on its execution date. When two documents address the same term, the later and more specific one controls, and the model must carry that resolved value forward rather than reporting whichever it read last.
The failure mode is order sensitivity. If a model reads the base lease, sees a $30 per square foot rent, and reports it without checking whether Amendment Three reset it to $34, the abstract is confidently wrong. The document was in the context window. The model simply did not reconcile. Robust systems make the amendment chain explicit: they establish document dates, sort the stack, and apply each modification in sequence so the final field reflects the last governing instrument.
The table below shows how a single rent term moves through a document chain, and why only the last layer belongs in the rent roll.
Document | Date | States base rent | Controls? |
Base lease | Jan 2019 | $30.00 / sf | No, superseded |
First amendment | Mar 2021 | Silent on rent | No |
Second amendment | Jun 2022 | $34.00 / sf | No, superseded |
Side letter | Jun 2022 | Abates rent 6 months | Partially, term-limited |
Third amendment | Feb 2024 | $36.50 / sf | Yes, controlling |
The rent roll should show $36.50, with a note that a prior side letter abated rent for a defined window. A model that reports $30, $34, or the arithmetic mean of the four has failed reconciliation, not extraction. Every intermediate figure was read correctly. Only the resolution was wrong.
Why Do Side Letters Break Naive Lease Reading?
Side letters break naive reading for two reasons: they are physically separate documents that are easy to omit from the set, and they often modify a term without restating it, so the model must connect a standalone letter to the specific clause it changes. A side letter that is not fed to the model, or not linked to its lease, is simply invisible.
A side letter is a collateral agreement executed alongside or after the lease that grants a concession the main document does not show: a rent credit, an early termination right, an exclusivity promise, a personal guaranty carve-out. Because it lives outside the lease body, two things go wrong. First, in a diligence pile of scanned PDFs, the side letter is the document most likely to be missed during collection, and no reading system recovers a term from a file it never received. Second, even when present, the letter frequently says "notwithstanding Section 4.2, Tenant shall pay" without repeating the section it overrides, so the model must resolve the cross-reference to know what changed.
The accuracy of a lease abstract is bounded by the completeness of the document set. A model reading 95% of the documents at 99% accuracy is still missing the term the fifth document changed.
This is named entity recognition meeting document-level reasoning. Finding "Tenant shall pay $34.00" is entity extraction. Knowing that this sentence, in a side letter dated after the second amendment, is the reason the abstract shows a temporary abatement is reconciliation. The first is necessary. Only the second is sufficient.
How Accurate Is AI When It Reads a Long Lease With Amendments?
AI reads standard fields in a long lease with high reliability and non-standard, cross-document terms less reliably, and the gap widens with the length and depth of the amendment chain. Industry benchmarks commonly put leading tools at 90 to 97% on standard commercial lease terms, but that average masks lower performance on exactly the reconciled fields that carry risk.
Two research findings explain the shape of the error. First, position matters. Liu et al. (2023), in "Lost in the Middle," found that retrieval accuracy across model families follows a U-shaped curve and degrades by more than 30% when the relevant passage sits in the middle of a long context rather than near the beginning or end. A term defined on page 90 of a 200-page document is in the danger zone. Second, aggregate accuracy hides distribution. A tool that reads parties, dates, and base rent at 97% and amendment-superseded rent at 80% can still report a headline number in the mid-90s, because the easy fields are numerous and the hard ones are few.
Field type | Reading task | Relative reliability |
Tenant, dates, square footage | Single-location extraction | Highest |
Base rent, standard escalation | Labeled-value extraction | High |
Amendment-superseded rent | Cross-document reconciliation | Lower |
Side-letter concession | Linked-document resolution | Lower |
Exhibit-governed value | Cross-reference resolution | Lowest |
The operator lesson is to underwrite against the reliability of the reconciled field, not the average. The most defensible workflow pairs the model's speed with human-in-the-loop review targeted at the low-reliability rows, and it cites every final value to the document and page that governs it, so a disputed rent resolves by opening the third amendment rather than re-abstracting the file.
Frequently Asked Questions
How does AI decide which lease document controls a term? Robust systems treat each document as a dated layer, sort the base lease and its amendments and side letters by execution date, and apply the later and more specific instrument as controlling. The final field reflects the last governing document, not whichever passage the model happened to read last.
Can AI miss a side letter when reading a lease? Yes, and it is the most common structural failure. A side letter is a separate document, so if it is not collected and fed to the model, its concession is invisible regardless of how accurately the base lease is read. Completeness of the document set bounds the accuracy of the abstract.
Why is reconciliation harder than extraction in a long lease? Extraction finds a value in one place. Reconciliation decides which of several conflicting values, spread across a base lease and multiple amendments, actually governs. Liu et al. (2023) showed model accuracy drops more than 30% when the relevant passage sits in the middle of a long context, which is exactly where superseding clauses tend to live.
Conclusion
Reading a 200-page lease is not a scanning problem. It is a reconciliation problem wearing a scanning problem's clothes. The base lease, the amendments, and the side letters each read cleanly on their own. The value that belongs in the rent roll is the one that survived the whole chain, and producing it means knowing which document controls, resolving the cross-references, and refusing to trust a field until the last instrument to touch it has been read.
For the operator, this reframes what to demand from any reading system. Ask which documents were in the set, not just how accurate the read was. Ask whether the amendment chain was resolved by date, not summarized in aggregate. Ask whether every final value cites the page that governs it. Firms that treat the lease as a stack of dated layers get a citable record. Firms that treat it as one long file inherit the term the last document quietly changed.