AI Features

This is the feature customers love. It’s our top engaged feature by a wide margin, it’s been live for over a year, and we’ve held a 96 percent accuracy rate the whole time.

Ask AI is a chat that lives in a sidecar, with the option to expand to fullpage. The user asks any HR or employment-law question in plain English. The AI doesn’t reach for the open internet or pull a generic ChatGPT-style response. It uses RAG over our own corpus: the deep library of articles our lawyers have written every time employment law has shifted somewhere in the country. The answer is generated against that corpus, and the response always links back to the source articles the AI used.

So it’s not a wrapper. It’s a grounded answer with sources. That distinction matters in legal, where a confident-sounding wrong answer is worse than no answer at all.

We picked RAG because it was the fastest way to get something out the door, and we never bothered exploring alternatives once it started working. It worked really, really well. Our team of lawyers reviews answers every week, and the 96 percent number has been stable for over a year. When the answer is wrong, our CSMs reach out to the customer directly. That has happened a handful of times in twelve months.

We don’t run “I’m not sure” guardrails on the answers themselves, because the answers are accurate enough not to need them. The only guardrail in the UX is detecting when a question isn’t actually a legal question, in which case the AI politely steps out of the way.

The AI merge is opt-in. The user accepts or rejects what the AI produces. If something does slip through wrong, the next legal update on that section will re-flag it, so the document doesn’t stay uncompliant for long. We’re planning a “scan this chunk for compliance” feature down the road.

The design challenge here is sheer volume. A customer who hasn’t updated their document all year can have hundreds of updates queued up. Pages and pages of legal text to walk through. Before AI, working through that backlog took hours. With AI in the loop, customers get through the same volume in under ten minutes. The win isn’t that AI is doing the legal thinking. The win is that AI removes the busywork so the customer can focus their attention on the chunks that actually need their judgment.

The rules engine has been the spine of SixFifty since the company started six years ago. It’s the backend that takes 5 to 100 questions about a customer’s situation, factors in jurisdiction-specific knowledge, and decides which sections of a legal document apply. It also drives compliance workflows like “hire a new employee in California,” which lays out every step a customer has to complete: offer letter, employment agreement, signatures, I-9, and on down the list.

We hooked AI up to the rules engine. Now a customer kicks off a workflow and the AI walks them through it.

The interaction lives in a fullpage chat, with the workflow’s progress on the right rail. The user can see every step that needs to happen, what has been completed, and what is still pending. If a step requires something the AI can’t generate on its own (like a signature from the new hire), the chat calls it out clearly so the user knows what’s on them.

Most of what the AI auto-builds is either pulling documents the customer already has on file, or running them through the question flow and showing them a preview of the resulting document on the right side. The new hire’s name and email come from the customer’s HRIS, so by the time the user kicks off “hire someone in California,” the AI already has most of the inputs it needs.

For a single new hire, that workflow can generate fifteen to twenty legal documents. Hours of work, or thousands of dollars in lawyer time, collapsed into a single chat session. The customer puts the new hire into their HRIS, and we handle the rest.