Legal AI, honestly.
Your firm is being sold an AI platform, or has already bought one. This lesson is the orientation nobody gives you: what these tools genuinely do, where they break, and why the buying decision turns out to be the easy part.
01 Start with the honest version
The legal AI market in 2026 is loud. Vendors publish benchmark wins, firms publish press releases, and everyone talks about transformation. Underneath that noise, the technology does something fairly specific and quite useful: it reads faster than you do, drafts a credible first pass, and finds things in piles of documents that would take a human days.
That is genuinely valuable. It is also not the same thing as practising law. The gap between those two sentences is where every real implementation problem lives.
Legal AI is not a junior associate and it is not a search engine. It is closer to an extremely fast, extremely well-read researcher who has no professional liability and no idea when it is wrong. Everything about how you should use it follows from that description.
02 What Harvey actually is
Harvey is the name that comes up most, so it is worth being precise about what it covers. As of 2026 it is not one product but several, and knowing which surface does which job is most of the literacy:
| Surface | What it is for |
|---|---|
| Assistant | Chat, drafting, and analysis of documents you hand it. Increasingly able to produce Word, PDF, Excel and PowerPoint output directly. |
| Vault | Bulk analysis across large document sets. Ask one question of a thousand documents and get a structured, auditable review table back. Built for due diligence at volume. |
| Knowledge | Research against licensed and public sources — tax guidance, SEC filings, EU and national case law, and a firm's own memoranda. |
| Agents | Multi-step work that runs end to end rather than turn by turn. Firms can build their own without code. |
| Spaces | Shared, governed workspaces — including guest access for clients and outside counsel who are not themselves customers. |
| Command Center | The leadership view: adoption analytics and benchmarking. This one is aimed squarely at whoever has to justify the spend. |
Harvey also runs frontier models from more than one lab — Anthropic's and OpenAI's models both appear in its model selector — which is a useful reminder that the platform is a wrapper of workflow, security, and legal data around models that anyone can license.
03 What it does not do
Three limits matter more than any feature list.
It does not know when it is wrong. Modern legal AI is heavily engineered to cite its sources and link back to the underlying document, and that helps enormously. It does not eliminate the failure. Lawyers in multiple jurisdictions have now been sanctioned for filing briefs containing citations that no AI-checked step caught, because no human opened the case.
It does not carry your professional obligations. Competence, confidentiality, supervision and candour to the tribunal remain entirely yours. A tool cannot hold a duty. This is why every serious firm deployment is accompanied by a policy, and why the policy is usually the part people skim.
It does not create adoption. This is the one that surprises firms. Licences get bought, a launch email goes out, and six months later usage has concentrated in a small group of enthusiasts while everyone else has quietly gone back to what they know.
A firm rolls out a legal AI platform to 800 people. Six months later, what does the usage distribution usually look like?
Have an answer in mind? Open this.
The common shape is a steep curve: a modest fraction of users generating the large majority of activity, a long tail who tried it once or twice, and a meaningful group who never logged in at all. The people at the top would mostly have figured it out on their own. The spend is justified by the middle of that curve, and the middle is exactly the part that training moves.
This is why vendors now ship adoption dashboards. Measuring the problem is not the same as fixing it, but you cannot fix what you cannot see.
04 The question that actually matters
Firms spend months on tool selection and comparatively little on the question that determines the return: who, specifically, is going to get good at this, and how?
Notice that this question has almost nothing to do with which vendor won the bake-off. Two firms can buy the identical platform and get results that differ by an order of magnitude, and the difference is not the software.
It usually comes down to three things: whether people have permission and know it; whether they have been shown what good use looks like on their own work rather than a demo matter; and whether anyone senior visibly uses the thing. That last one is worth more than any training budget.
Before the next lesson
Find out two facts about your own organisation: which legal AI tools you are actually licensed for, and what your written policy says you may put into them. A surprising number of people at firms with sophisticated deployments cannot answer either. If you can't, that is itself the finding.