A language model can tell you that page 3 and Schedule B disagree. It cannot tell you which one is correct, and the difference between those two statements is the whole of the professional judgment involved.
Key takeaways
Yes, for a narrow and well-defined class of problem: places where a document contradicts itself. A model reading a contract, a report or a file bundle will find defined terms that are used but never defined, cross-references pointing at a clause that was renumbered two drafts ago, party names that change spelling between the recitals and the signature block, dates that run backwards, and figures in the body that do not match the figures in the schedule. That is genuinely useful work, and it is work that human reviewers are measurably bad at because it requires holding forty pages in mind at once.
What it cannot do is tell you the document is wrong. A consistency check is a closed operation — it compares the document to itself. If a purchase price is stated as the same wrong number in all four places, every consistency engine ever built will return a clean result. That is not a limitation to be engineered around; it is what the tool is.
The first three are where the return is. Treat the fourth as a list of questions the machine has raised, not a list of defects it has found.
Most consistency checking is generic. Two Canadian requirements are specific enough to encode as a rule and costly enough to be worth encoding.
The yearly-rate statement. Under section 4 of the Interest Act, where interest is made payable by the terms of a written or printed contract at a rate or percentage per day, week or month, or at any rate for any period less than a year, no interest exceeding five per cent per annum is chargeable, payable or recoverable on any part of the principal unless the contract contains an express statement of the equivalent yearly rate. That is a document-internal test: a monthly rate appears, an express equivalent annual rate does not, flag it. It applies to contracts other than mortgages on real property and hypothecs on immovables, which the section carves out and treats separately.
Tax disclosure. section 223 of the Excise Tax Act requires a registrant making a taxable supply other than a zero-rated supply to indicate to the recipient — in the prescribed manner, or in the invoice or receipt issued, or in a written agreement — either the consideration and the tax payable in a manner that clearly indicates the amount of the tax, or that the amount payable includes the tax. Subsection 223(1.1) adds that where the invoice, receipt or agreement shows the tax payable or the rate or rates at which tax is payable, it must show the total tax or the total of the rates. A fee agreement that is silent on tax while the invoices show it is exactly the sort of clash worth catching before the client does.
The single configuration decision that matters is whether the tool is allowed to edit. It should not be. A consistency finding is evidence — it tells you something happened during drafting, and sometimes the something is important. If the tool silently harmonises the schedule to the body, you lose the information that they ever differed, and you may have picked the wrong one.
Require three things of every finding: the two locations, the two verbatim strings, and a one-line statement of why they clash. That format makes review fast and makes a false positive obvious in seconds. It also means the reviewer is reading the document, not the model.
The reason to be careful is that unresolved ambiguity is not a cosmetic defect. Where a genuine ambiguity survives ordinary interpretation and one party drafted the clause, an Ontario court may resolve it against the drafter — Treadstone Law sets out how that works in its note on ambiguous contract terms and contra proferentem and in a shorter answer on how Ontario courts interpret ambiguous language. The downstream cost is set out in its note on damages for breach of contract in Ontario.
For an Ontario lawyer the relevant rule is not an AI rule. Commentary [4A] to rule 3.1-2 of the Law Society of Ontario’s Rules of Professional Conduct says that to maintain the required level of competence a lawyer should develop an understanding of, and ability to use, technology relevant to the nature and area of their practice and responsibilities, and should understand the benefits and risks associated with it, recognising the duty to protect confidential information under section 3.3. Commentary [4B] makes that proportionate: it depends on the practice area, the geography and the requirements of clients.
The review discipline is the older obligation. Under rule 6.1-1 a lawyer must assume complete professional responsibility for their practice and directly supervise non-lawyers to whom tasks are assigned, reviewing the work at frequent intervals to ensure its proper and timely completion. A tool is not a person and the rule does not reach it directly, but the shape of the answer is the same: output that nobody read is not work product.
Accountants, engineers and brokers are under differently worded versions of the same expectation. Whichever profession you are in, the finding is a prompt for a reviewer, not a result.
A six-lawyer Ontario commercial practice runs a consistency pass over a share purchase agreement at second draft. The tool returns nineteen findings. Eleven are renumbering casualties from a heavy edit, two are defined terms that survived a deleted schedule, three are date-order problems in the closing mechanics, and two are arithmetic clashes between the price adjustment formula and the worked example in the schedule.
The nineteenth is the one that mattered: the indemnity cap is stated one way in the body and another way in the definition it depends on. A partner resolves it in four minutes because the finding quoted both strings. Nobody claims the machine understood the indemnity; it noticed that two numbers about the same thing were different.
Measure the thing you can count. The countable metric here is findings per document that survive review as genuine, and the proportion of those found after the document left the firm rather than before. Both are countable before and after. Hours saved is not, which is why it makes a poor business case.
This is a service-delivery use: a firm checking documents it produces for a book of clients. If what you actually want is your own month-end and reconciliation work tightened up, that is a different job and it lives on the accounting automation page. If the question is about front-desk intake and scheduling in a practice, that belongs on the professional practice owners page.
Does this work on scanned documents?
Only as well as the extraction underneath it does. On a scan, the arithmetic checks are the ones to distrust first, because a single misread digit produces a confident and wrong finding. Spot-check extracted figures against the image before relying on a numeric clash.
Can it compare two versions of the same agreement?
Yes, and that is a different and easier task than internal consistency — it is comparison, not inference. The value of a model over a plain redline is that it can group the changes by what they do rather than by where they are.
Should the findings go on the file?
The resolved list should. It is short, it shows the document was reviewed, and it is the record you will want if the clause is ever argued about. Keeping it costs nothing and is consistent with the retention obligations in section 230 of the Income Tax Act for a firm’s own books.
Will it catch a missing clause?
Sometimes, and never reliably. A missing clause is an absence, and absence is only detectable against a template or a checklist you supply. Give it the firm precedent to compare against and it becomes a completeness check, which is a different and useful thing.
A 30-minute call is enough to tell you whether AI pays for itself here.
