A breakthrough in legal research?

The AI company Anthropic made a splash in January when it unveiled a new tool called Cowork for the paid version of Claude, one of the leading language models and a main competitor to ChatGPT. 

Flying under the radar is a new capability that gives lawyers in Canada a very powerful tool for legal research that is much cheaper and, in many ways, better than dedicated legal AI tools like LexisNexis’ Protégé.

Cowork does something I’ve been eagerly awaiting since chatbots gained the ability to search the web in 2023. Claude’s new app can run queries on websites like CanLii, read the results, and produce clear, accurate summaries of what it finds.

This capability isn’t advertised anywhere, and setting it up takes a few minutes. But it is by no means complicated. And it costs only about $30 CAD a month — a fraction of what a subscription to Lexis or Westlaw’s AI will run you. 

However, it’s not perfect — it will miss some relevant sources. But running the same three queries on Cowork and Protégé persuaded me that Cowork with CanLii is the better tool for legal research.

What is Cowork and how does it work?

Cowork is a feature in the paid version of Claude that carries out tasks using files or data on your computer or on the web.  It is currently the most capable and effective form of “agentic” AI out there.

You access Cowork by running the desktop version of the Claude app on a Mac or PC. But to get it to do its magic with CanLii, you need to take one further step and install the ‘Claude in Chrome’ extension in your Chrome browser.

Once you do this, anything you ask Cowork to do that involves surfing the web will be done using Chrome. The fun part is that you can watch it navigate Chrome in real time right in front of you.

Of course, there are risks to having AI visit websites on your behalf, and for this reason, Cowork will seek your permission before doing anything on a website for the first time.

The Cowork breakthrough

Let me spell out the breakthrough here. You type a query into Cowork — something like “Go to CanLii.org and find me three recent cases in Ontario about X, Y, and Z” — and it will do just that. It will narrate every step and then package the results into roughly 200-word summaries for each case.  I found these to be accurate and informative, with a link to the case itself in each heading.

At the moment, there is no other tool that can do this. CanLii itself has recently unveiled its CanLii Search+ AI feature, which is helpful but more limited. It uses a language model to help formulate a query and rank the results by relevance. But it goes no further.

Cowork formulates the query, sifts through the results for you, and summarizes the best results in a convenient, accessible format. We have not, until now, had AI capable of carrying out all steps of a search for us on a publicly accessible database.

My Cowork - Protégé test run

I conducted three queries using both tools. Briefly, here's what I did and what I found: 

(1) I asked Cowork to go to CanLii and find three decisions from an appellate court in Canada in the last 10 years involving a challenge to a strip search by police. It spent roughly 10 minutes combing through cases before producing summaries of three important court of appeal cases on point, from three provinces, between 2016 and 2020. 

I then asked Protégé to do the same. It came up with three cases almost instantly, and two were the same ones that Cowork found. Protégé’s summaries were much shorter and less helpful. But each result had a link to the case, as all results on Protégé do.

What wasn’t clear to me was why Cowork surfaced a case from 2016, and Protégé surfaced one from 2017. 

I then narrowed the search, asking each tool to find me two cases on point in the last three years. Cowork surfaced the two most important, both from 2025. Protégé failed this test by retrieving only cases from 2022 and earlier.

(2) I tried a more specific search: “Find me the three most recent cases from any province in Canada involving the sentencing of a first-time offender — someone with no criminal record — for possession of child pornography in which house arrest was imposed.”

Protégé dropped the ball on this, giving me one wrong case (the offender got jail, despite the summary saying it was a conditional sentence), and misdescribed what happened in another one.

Cowork’s results were all within the last year and involved higher courts, including the Ontario Court of Appeal. The summaries were vastly superior and all accurate.

(3) I conducted a search on both platforms of scholarship involving an issue I know well — the use of surreptitious recordings as evidence in civil litigation — having written an article on the topic recently. Here, the disparity was more pronounced.

Protégé provided links to only two results; one was my article in the Manitoba Law Journal, and the other was misattributed to the wrong author (but at least linked to the full text). Cowork surfaced eight articles, including mine, along with all of the notable scholarship I drew upon in writing it, as well as three pieces I had missed and would have appreciated at the time. All were summarized very helpfully and concisely.

Takeaways

I was thoroughly impressed with the quality of Cowork’s search results and its summaries. The only reservation I came away with was uncertainty about how and why it missed cases that Protégé surfaced (and vice versa). 

While I don’t have an answer to that question, I did notice that when Cowork finished running a query on CanLii, the webpage remained accessible so that I could take a closer look myself. 

In one instance, I went through the top 15 results and found only one additional source I might have flagged. The searches otherwise seemed thorough and effective.

Cowork will no doubt continue to evolve, and very soon, there will be other tools to compare it to. But at the moment, using an agentic tool like this with CanLii strikes me as the most accessible and powerful way to do legal research using AI. Given the cost, it is a tool well worth using.