ChamberOne

Defined terms and cross-references: using AI to proofread long agreements, and where it still misses errors

By the ChamberOne team 7 minute read

Most drafting errors that embarrass a firm after signing are not errors of law. They are a clause that refers to "clause 14.3" when the provision moved to 15.3, a term capitalised in one schedule and not in another, or a definition that survived after the clause that used it was deleted. These are mechanical faults, and a long agreement built up over weeks of negotiation produces them in large numbers. AI tools are now being used to catch them, alongside two older aids: Word's own cross-reference fields and dedicated proofing add-ins. Each catches different things and each misses different things. This article compares them and explains where a language model still fails. It ends with a short routine a small firm can run before a document goes out for signature. It is general information about tools and method, not advice on any particular document.

Key points

  • Word cross-reference fields keep clause numbers current after you update fields, but they do not check that the target clause is the one the drafter meant, and hand-typed references are not covered.
  • Rule-based proofing add-ins over-report: a published review found most flagged items were not errors, so someone must clear the list.
  • A language model can judge whether a referenced clause makes sense, but it checks only the version and text it was given, and a late edit makes its result stale.
  • Long files and separately supplied schedules are weak points; ask the tool to list the terms and references it checked so you can spot-check them.
  • Freeze the text, run the checks, verify risk-bearing references by hand, and compare the signing copy with the checked copy.

The errors worth looking for

Four kinds of fault account for most of the mechanical damage in a long agreement. The first is the defined term that is used but never defined, or defined twice with different meanings. The second is the cross-reference that points to the wrong clause, or to a clause that no longer exists. The third is numbering drift, where inserting or deleting a clause shifts everything after it and typed references do not follow. The fourth is the schedule mismatch: the body refers to Schedule 3 as a list of premises, but Schedule 3 in the file is the fee table.

None of these needs legal judgment to spot, which is why they are a reasonable job for software. All of them need judgment to fix, because the right repair depends on what the parties meant.

What Word's cross-reference fields do

Word's built-in cross-reference is the baseline. Microsoft's support page explains that a cross-reference is inserted as a field, and that the point of a field is that it can be refreshed when the document changes. For a numbered clause, Word tracks the target with a hidden bookmark. After edits, the page's instruction is to select the whole document and press F9 to update all fields.

The same page flags two practical traps. If you press Enter at the beginning of a referenced numbered item to add a new item before it, the bookmark can expand to include the new item, and the reference may then display or link to more content than intended. Microsoft's advice is to add the new item from the end of the preceding one. The page also notes that a cross-reference to a separate document is not a live field reference unless you use a hyperlink or combine the files into a master document. That matters for schedules kept as separate files.

Fields only help if they were used in the first place. A reference typed by hand as plain text, such as "subject to clause 9.2", is invisible to Word's update. The fields also do not tell you whether the clause now at 9.2 is the one the drafter meant. They keep numbers current. They do not check sense.

What dedicated proofing tools add

Dedicated proofing add-ins sit inside Word and check the structure of the document. Legaltech Hub's description of Definely Proof, a Word add-in, gives a fair picture of the category. It lists checks for terms used before they are defined, inconsistent punctuation or styling in definitions, defined terms not capitalised correctly, cross-references that no longer match their target, misalignment between numbered items and their text, and a list of leftover placeholders, brackets, comments and highlights. Other products in this category advertise similar checks.

A 2017 LawNext review of one such tool, Contract Tools, is useful because it reports the false alarms honestly. The tool found defined terms by looking for quotation marks, with the option to switch to bold, underline or italics. It also listed the annexes, appendices and schedules the document refers to. On the first contract the reviewer tested, it raised 173 issues, including 53 unknown cross-references and eight defined terms never used, and most turned out not to be mistakes. Many of the unknown references were to external documents or statutes. On two further contracts from the SEC's EDGAR database, every flagged issue was a false alarm. The reviewer still preferred the over-cautious tool. That is a single review of one product from some years ago, but the pattern is the one to expect: rule-based checkers over-report, and a person has to clear the list.

These tools work from the document's own mechanics. If the drafter used quotation marks for definitions, they work well. If definitions are in a table, or a term is defined inline in a clause, results depend on how the tool is set up. Test any tool on a few of your own agreements before relying on it.

What a general AI model adds

A general language model does something different. It reads the text and can reason about it. Asked to list every capitalised term and where each is defined, it can find terms that appear in the definitions clause under a slightly different name from the one used in clause 12. Asked whether clause 8.4's reference to "the Termination Notice under clause 19" makes sense, it can look at clause 19 and say that it deals with assignment. A pattern checker cannot make that second kind of observation, because it only sees that clause 19 exists.

The two approaches are complementary, not rivals, and the trade-offs are worth seeing side by side.

Rule-based checker or Word fields
  • Works from structure: fields, numbering, quotation marks
  • Over-reports, for example references to external documents
  • Confirms that a target exists, not that it makes sense
  • Does not see typed references unless built to look for them
General AI model
  • Can judge whether the clause referred to is the sensible one
  • Sees only the text it was given, in the version it was given
  • Output can vary between runs, so record the one you rely on
  • Can state a clean result when the document was not fully read
Rule-based checkers and AI models compared

Where AI still misses errors

The first failure is the silent late edit. A model checks the file you gave it at the moment you gave it. If a partner changes clause 6 after the check, nothing alerts the model or the reader, and the earlier clean result now describes a document that no longer exists. The same applies to a tracked-changes file where the model may read the accepted text, the original text or a mixture. Confirm which version was checked and keep the check date against the file.

The second is length. Models can only take in a limited amount of text at once, and tools deal with longer files in different ways, such as splitting them into parts. A split document may lose the link between a definition on page 4 and a use on page 140, which is exactly the relationship you are checking. Even where the whole text fits, research on language models published in 2023 (Liu and colleagues, later in Transactions of the Association for Computational Linguistics) found that performance can degrade significantly when relevant information sits in the middle of a long input, even for models built for long contexts. That study tested particular tasks on models of its time, so it does not tell you how a current tool performs on your agreement. It is a reason to test with your own documents and to be wary of a clean result on a very long file.

The third is schedules supplied separately. If the schedules, the order form and the data protection annex are separate files, a check on the main agreement cannot confirm that Schedule 3 says what the body assumes. Put the full set in front of the checker, or check each against the body by hand.

The fourth is false confidence. A model can report "no issues found" fluently whether or not it examined everything. Ask it to list the defined terms it found and the cross-references it checked, with clause numbers. A list you can spot-check is evidence of work done. A bare statement is not.

A final proofread routine before signing

This routine suits a small firm with no dedicated knowledge team. It assumes the text is final, because every step is wasted if edits follow.

  1. 1

    Freeze the text. Accept or reject all tracked changes, remove comments, and save a named version. Record who approved it as final.

  2. 2

    Update fields. In Word, select all and press F9, then switch field codes on and off with Alt+F9 to check none show as raw codes. Search for typed references such as "clause" followed by a number, which fields will not have updated.

  3. 3

    Run the proofing add-in if the firm has one. Clear the list item by item and note which items were false alarms, such as references to statutes.

  4. 4

    Run the AI check on the frozen file, including every schedule. Ask for a table of defined terms with the clause where each is defined and used, a table of every internal cross-reference with the clause it points to and a one-line description of that clause, and a list of terms used but not defined.

  5. 5

    Verify the AI output yourself. Spot-check at least the cross-references in the clauses that carry the most risk, such as termination, liability, payment and governing law. Open each target clause and read it.

  6. 6

    Read the schedules against the body. Confirm that each schedule's title, number and content match how the body describes it, and that blanks and brackets are filled or deliberately left.

  7. 7

    Compare the signing copy with the checked copy. Run a document comparison between the frozen version and the PDF or Word file that will actually be signed, so a late edit cannot slip in.

  8. 8

    Keep a short note on the file: tools used, date and version checked, and who reviewed the output.

The lawyer who signs off remains responsible for the document. The tools only reduce the number of mechanical errors that reach that final read.

A sensible division of labour

For most small firms, the practical order is Word fields for the numbering, an add-in for pattern checks if the volume of long agreements justifies the cost, and an AI model for the questions that need reading, such as whether a reference makes sense. The human read is not replaced by any of them. It is the step that deals with intent, and with the cases where the tools disagree.

Whichever combination you use, treat the first month as a trial. Take three agreements you have already signed, run the checks, and see whether they find mistakes that your own review missed and how many false alarms they raise. If the firm handles client documents with an AI tool, also check what the tool does with the text you send it, and that your client confidentiality obligations and any data protection rules in your jurisdiction are met. Those rules differ between Sri Lanka, the UK, the US, Australia, Singapore and India, so ask a lawyer qualified in your jurisdiction about your own circumstances.

Sources

  1. Microsoft Support: Create a cross-reference
  2. Legaltech Hub: Definely Proof by Definely
  3. LawNext (Robert Ambrogi, 2017): Contract Tools review
  4. arXiv: Lost in the Middle: How Language Models Use Long Contexts

This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.

Read next

AI proofreading of defined terms and cross-references