Almost everything written about legal AI is written about law firms, which makes sense, because that’s where the spend is and that’s where the vendors are pointed. In-house teams then inherit the same tools and the same framing, and the framing doesn’t really fit them.
A firm sells legal work, so time saved on drafting turns fairly directly into capacity, and a drafting or research tool has an obvious business case behind it. An in-house team isn’t selling anything. It’s a cost centre answering the rest of the business, and its bottleneck tends to sit somewhere else entirely.
The bottleneck is finding, not writing
The question that comes round again and again in-house is some version of “what did we actually agree with them?” Not what the law says. What our own paperwork says. And the answer to that is scattered across executed agreements, schedules, side letters, an email chain, and two amendments that changed the wording without replacing the document they changed.
Nobody has ever mapped that estate. It lives across a document management system, a few shared drives, someone’s inbox, and whatever the deal team saved locally in 2021. The lawyer who gets asked the question knows roughly where to start looking and then spends the afternoon looking.
A legal research tool doesn’t help with any of that, and it isn’t meant to, because it has no access to your files. A drafting tool doesn’t help either. The thing that helps is retrieval across your own estate with the supporting passages attached, so that the answer arrives with its evidence rather than asking to be trusted.
The review doesn’t go away
The reviewer starts from the evidence instead of hunting for it, which is worth a lot. There’s still no shortcut past the review.
Sell it internally as an answer engine and people will act on the output without checking, and in-house teams have less room for that than firms do. There’s usually no second lawyer sitting between the answer and the business decision. Introduce it as a retrieval tool that shows its working and it fits how the work already gets done, which is also the version that survives a conversation with the general counsel.
The practical test is how long it takes to verify one answer. If checking a citation means opening a 200-page PDF and searching it, people will check the first few and then quietly stop, and at that point it’s a liability.
Permissions are harder in-house than they look
Firms think about information barriers constantly, because the obligation is explicit and someone is accountable for it. In-house teams often assume the problem doesn’t apply to them, on the grounds that everyone works for the same company.
It applies. Employment files, investigation material, board papers, a transaction most of the business doesn’t know about yet. Point a system at the legal team’s document estate and it will index all of it, and the question you need answered is whether it can return a passage from a document a given user has no right to read.
Press on where the enforcement actually happens. Filtering results after retrieval isn’t the same as enforcing access during retrieval, because a generated answer can carry content it doesn’t cite. Then test it rather than taking the answer on trust: revoke someone’s access to a document and ask a question only that document can answer.
What to buy, and in what order
If the pain is that you can’t find what you already agreed, then what you’re buying is document analysis over your own estate, and the legal research products aren’t a substitute for it however good they are. If the pain is genuinely “we need to know what the law says”, buy the research product. If it’s drafting volume, buy the drafting tool.
Most in-house teams I talk to describe the first problem and then get shown the second and third, because that’s what the market is built to sell.
Wherever you land, start with a bounded set of documents rather than the whole estate: one contract portfolio, or one entity’s policies. A bounded set is small enough to evaluate properly, and the only way to learn how it behaves on documents you don’t know is to test it on documents you do.
There’s more on the evaluation approach in the legal document work guide, and on how the categories differ in legal AI tools compared.
