Editor's Note: The author would like to thank Relativity colleagues Jeff Gilles, Jill Ragan, and Elise Tropiano for their significant contributions to this article.
Some things are better together. Peanut butter and jelly. Email threading and textual near-duplicate identification. A second monitor and literally anything.
I want to make the case for a new pairing: aiR Review + Assist.
Most teams have met these two separately, and that's fair enough; they often show up at different points in a matter, and they do different work. More specifically, Assist is a conversation with your data. aiR Review classifies documents and extracts information exhaustively against criteria you write. Early case over here, review over there, case strategy over yonder – and never the twain shall meet.
I'd argue they should meet, because the line between them isn't really early case versus review at all. It's the shape of the questions you're asking.
Assist answers questions. It finds the most relevant material in an index and hands you a cited answer, quickly. What it doesn't do, by design, is exhaustive coverage. Ask what the evidence says about a topic and you'll get a real answer with supporting citations you can click into. Ask for every single document mentioning that topic and it'll politely tell you that's a different job.
That different job belongs to aiR Review, where every eligible document or image gets looked at, gets a call, and gets a rationale explaining the call, allowing you to answer your broad and targeted questions at scale
So: questions on one side, coverage on the other. Here's where that pays off.
#1: Let Assist build the sample you iterate against.
Prompt criteria development has a chicken-and-egg problem you might not expect.
You need a small set of documents to test your criteria against. But a good set is one that exercises criteria you haven't written yet. Best practice asks for 50 to 100 documents, diverse across your categories of interest, rich in pertinent material without being unrealistically clean, and coded by humans in advance. Getting there has traditionally meant search terms reports, clustering, date filters, custodian lists – in other words, a heroic amount of clicking. I've seen teams spend the better part of two days on it.
Assist gives you a running start. Take your draft protocol and, instead of translating every criterion into Boolean, just ask: What evidence is there of gifts, incentives, or entertainment offered to procurement contacts?
Now you've got citations you can open, verify, and save into a search. There's your iteration set for that issue.
But the documents might not even be the best part. You also come away knowing how the issue shows up in this data! The vocabulary people used, the nicknames, the abbreviations that mean nothing to anyone outside the company, who was in the room – you get the point.
That adds up to substantive early insight. Our prompt criteria best practices tell you to define internal jargon, aliases, uncommon acronyms, and unfamiliar email addresses (hint: your key custodian has a personal Gmail in there somewhere). Excellent advice. But how do you know what any of that is before you've read the data? This is how.
Two things to keep in mind, though.
First, a sample made entirely of clean, obvious hits will make mediocre criteria look fantastic, and you'll find out the hard way at scale. Go ask for the hard stuff on purpose. The documents that use the right vocabulary in the wrong context. The issue mentioned in a purely administrative email. The borderline calls reviewers will argue about over Slack. Pull a plain random sample alongside all of it too, so the boring, non-responsive material is represented.
Second (and this one's important): when you get to the validate phase, let aiR Review pull that sample itself, randomly, from the population. Don't hand it the documents you found by asking Assist pointed questions.
Why? Because your validation sample exists to tell you how the criteria will perform on documents nobody hand-picked. If you seed it with material you surfaced by asking about the exact issues in your criteria, you've effectively baked your own assumptions into the measurement. You'll get a lovely precision number that describes how well your criteria work on documents that were already obviously about the issue, which is not the question any validation is trying to answer, and it's not a number you want to have to defend later.
Curated samples are for developing; random samples are for proving.
#2: Flip it and reverse it: let aiR Review decide what's worth talking to.
The first pairing sends documents from Assist into aiR Review. This one runs the other direction.
Assist gets sharper as your document set gets tighter; that’s how retrieval behaves. Point a question at everything you collected and you're asking about one specific transaction while competing with calendar invites, newsletter subscriptions, and three years of “sounds good, thanks!” Narrow the set to material that's already been determined to matter, and the answers get more precise. Our indexing guidance says the same thing: build focused indexes, aligned to specific custodians, issues, themes, or time periods.
An aiR Review pass is a great way to get there. Run your criteria, then take what comes out the other side – the responsive set, the key documents, whatever's flagged on a particular issue – and save it as a saved search. Then, build your index on that. Every conversation you have from then on happens against documents that earned their place, with a documented basis for why.
This then opens up the part of the matter where your most expensive people spend their time: building the timeline, figuring out who knew what and when, finding deposition exhibits, and preparing for opposing counsel to ask something pointed about a document you last thought about three weeks ago.
Those are all question-shaped problems, and Assist is good at helping you tackle question-shaped problems more quickly.
#3: Scope your custom analysis before you turn it loose.
Custom analyses then changes the math a little. You can define your own analysis in plain language, up to five per project, and apply it across a whole document set – classification, extraction, evaluative questions, and with the Vision model, image analysis.
The catch is that each insight is compact. Given one title and roughly a thousand characters of instructions, it works best as one question with one answer format. Bundle three asks into a single open-ended prompt and the output may come back inconsistent. It can be a tight space to work in, which makes it worth knowing the shape of your data before you start writing.
Say you want to find the governing law provision out of every agreement in the population. Ask Assist what the corpus says about governing law and jurisdiction clauses first. Five minutes later, you know whether these agreements state it plainly, whether two competing conventions are floating around, and what the edge cases look like. Then you write an insight that accounts for that reality, including instructions for the missing and ambiguous cases (because there are always missing and ambiguous cases).
Fewer rewrites, fewer surprises. Five minutes of asking sure beats three rounds of guessing.
One limit to note: Assist reads extracted text, so if your custom analysis is a Vision analysis hunting for handwriting or stamps, it can't scope that for you. Sampling and human eyeballs for iteration, in that case.
But wait, can't I just ask Assist to write the prompt criteria for me?
Someone always asks. The answer is no – not because we're being precious about it, but because that isn't what Assist is designed to do. It answers questions about indexed documents. “Write me a prompt for aiR Review” isn't a question about your documents, so it isn't a supported ask.
Do the two-step instead. Ask Assist for the facts, the actors, the issues, and the vocabulary. Then, write the criteria yourself with all of that in hand. This is the best way to ensure you understand what you submitted and can explain it to someone else, which tends to come up. You can also create a document with this information and feed it into aiR Review’s prompt kickstarter functionality to help speed things along.
TL;DR
Reach for Assist when you have a question. Reach for aiR Review when you need every document accounted for. Above all, remember: both of these solutions live in a single platform. One system of record, one system of action, and endless possibilities to interrogate, understand, and do something meaningful with your data without ever having to move it around.
Ask questions to find the documents that'll teach you how to write your criteria. Iterate on that curated set, then step back and let aiR Review pull a random sample to prove the criteria hold up. Once the review's done, point Assist at what came out and start asking the questions your case really builds on.
Keep humans at the decision points, verifying citations and pondering each matter. And keep in mind that a conversational answer is a starting point, not a coverage report – both applications show you their reasoning specifically so you can check their work. So check it!
And, if you’re looking for someone to check your workflows, please know we’re here to help. Reach out to our success team any time for guidance on how to make the most of these and all of the AI capabilities embedded throughout Relativity aiR. We’d love to assist!
Graphics for this article were created by Caroline Patterson.
