Can AI tell us what changed between two versions of a document?

Yes, within limits. Ordinary comparison software is better than AI at finding every textual difference between two versions; what AI adds is an explanation of what those differences do, which obligations they move and which ones a fee earner should look at first. Treat the AI output as a reading of the change, not as the authoritative list of changes, and check the points that matter against the documents themselves.

Start with the redline, not with the AI

A comparison tool built for the job does something narrow and does it reliably: it marks every insertion, deletion and move between two files. That is a mechanical task with a right answer. If your firm already has comparison software, keep using it as the source of truth for what has changed.

The problem with a redline is not accuracy. It is volume. A document that has been through three rounds can come back with hundreds of marks, most of them numbering, formatting, defined-term tidying and the other side's house style. The fee earner still has to work out which five changes actually alter the deal, the risk or the position.

That sorting job is where AI earns its place. You are not asking it to find the differences. You are asking it to tell you what the differences amount to.

Ask about effect, not just difference

The useful questions are the ones a supervising partner would ask after glancing at the redline:

• Which changes shift a liability, a cost or a risk from one party to the other?
• Have any time limits, notice periods or cure periods changed length?
• Have any caps, thresholds, percentages or monetary figures moved?
• Has anything been deleted entirely, and what did it do?
• Have any definitions changed, and which clauses rely on those definitions?
• Are there changes that look cosmetic but alter the meaning, for example "may" to "shall", "reasonable" to "absolute", "including" to "including only"?

The last two are where a human reader most often slips. A changed definition shows up in one place on the redline and takes effect in fifteen. A single modal verb is a one-character mark that nobody's eye catches at five o'clock. Asking the question explicitly is more reliable than hoping to notice.

Framing the request so the answer is checkable

Give the tool both versions in full rather than extracts, and say which is the earlier one. If you give it a redline as well, say so, because a redline read as a plain document can be confusing: struck-through text is still text on the page.

Ask for the answer in a form you can audit. A list of changes, each one naming the clause and the page it sits on, with a sentence on what the change does, is far more useful than a flowing summary. You want to be able to put your finger on each item in the document in a few seconds.

Ask it to say what it could not determine. If a change refers to a schedule you have not supplied, you want to be told that, not to receive a confident guess about what the schedule says.

It also helps to say what the client cares about. "Flag anything that affects payment timing or termination" produces a sharper answer than "summarise the changes", because it tells the tool what counts as important on this matter.

What to check before you rely on it

Read every flagged change in the document itself. The AI's description of a clause is not the clause; the clause is what binds the client.

Then run two checks the output will not do for you.

First, completeness. AI summaries are a filter, and filters drop things. Run the redline alongside and satisfy yourself that nothing material has been left out of the summary. Pay particular attention to deletions, which are easier to miss than insertions because there is nothing new to read.

Second, consequences. If a defined term has changed, search for that term throughout and look at every clause that uses it. If a number has changed, check whether any other provision refers back to it. Knock-on effects are exactly the sort of thing that is obvious once pointed out and invisible until then.

Finally, keep a note on the file of what you asked and what you checked, in the same way you would note a review by a colleague.

Where this is most and least useful

It is most useful on long documents with many rounds of amendment, on documents drafted by someone else in an unfamiliar style, and when a matter lands on your desk at a late stage and you need to understand how the current draft got there.

It is least useful on short documents you can read twice in ten minutes, and on documents where the change is one sentence you have already been told about. There is no merit in putting a tool in front of work that is quicker done directly.

And it is no substitute for judgement about whether a change is acceptable. The tool can tell you that the indemnity cap has doubled. Whether to accept that is a question for the fee earner and the client.

Where Alesis fits

Alesis answers questions about a matter from the matter's own papers and names the page each answer came from; if the papers do not say, it says so. Documents are read page by page, so citations point at pages, and any page it could not read is flagged rather than skipped. When it cannot find support for a point it says what is missing instead of guessing. It assists qualified professionals and does not replace them, and it does not provide legal advice.