Why a legal AI needs a workflow engine, not a chatbot

When we started building LEA, the temptation was the same one everyone in this space faces, wrap a strong language model in a friendly chat UI, point it at a corpus of legal documents, and call it a product. We almost did it. We're glad we didn't.

After three months of running real matters end-to-end, trademarks, contracts, employment disputes, the unglamorous middle of legal work, the lesson is uncomfortably clear: the model is the easy part. The product is everything around it.

What "chatbot legal AI" actually looks like in production

A chatbot is a stateless conversation. You ask, it answers. That model breaks the moment a real legal task touches reality:

• A trademark application isn't a question, it's a 14-step process with conditional branches, regulator deadlines, and a human attorney who has to sign off before anything leaves the building.
• A contract review isn't a single answer, it's a redline plus a memo plus a follow-up call plus a version that needs to survive an audit two years later.
• "Did the model hallucinate?" isn't the right question. The right question is: who is on the hook when it does?

A chatbot has no answer for any of that. A workflow engine does.

What we actually built

Underneath the chat surface, every LEA interaction compiles down to a workflow graph. Nodes are skills, narrowly scoped, individually validated capabilities like "draft a mutual NDA" or "screen a mark against famous marks in class 9". Edges are routing decisions: which skill runs next, when a human reviewer is required, when to pause for a client confirmation.

That sounds like over-engineering for a chat product. It isn't. It's the difference between:

"Here's a draft NDA, hope it's right."

and

"Here's the draft. It was generated by skill v4.2, reviewed by counsel Priya S. on June 11, and the three clauses you flagged last time are unchanged. Audit trail attached."

Only one of those is something an in-house legal team can actually use.

The three things a chatbot can't do

1. Route to the right human. Legal work isn't uniform. A founder asking about a co-founder dispute needs a different lawyer than a finance team negotiating a master services agreement. The workflow engine knows.

2. Validate before it ships. Every skill has explicit pre- and post-conditions. If a generated contract is missing a governing-law clause, the workflow doesn't ask the model nicely to try again, it refuses to advance until the gap is filled.

3. Survive an audit. Two years from now, when someone asks "why did we sign this?", the answer can't be "the AI said so." It has to be a replayable, signed, time-stamped chain of decisions. Workflow engines do this natively. Chatbots don't.

The takeaway

The model matters. Of course it does. But if you're building anything that real lawyers will stake their licenses on, the chat UI is just the front door. The product is the building behind it, and the building is a workflow engine.

We'll keep writing about the specifics: how skills are scoped, how validation actually works, how we handle counsel-in-the-loop without making it a bottleneck. If any of this resonates, we'd love to hear from you.