Generic AI on contracts hallucinates. Grounded AI cites the source for every insight. The difference is whether your client can act on it.
An attorney's job is to give an answer the client can defend. That's the whole thing. The advice that wins the case, gets the deal signed, keeps the company out of court — it has to be grounded in something specific, verifiable, and yours. So when a contract review tool says "this clause is risky," the only useful follow-up question is: based on what? If the model can't answer that question, it isn't doing the job. It's doing a confident-sounding impression of the job.
The pattern repeated everywhere. A new tool would launch. The marketing said "review contracts in seconds with AI." You'd paste in your MSA. You'd get back a list of "risks" with risk levels. The list looked impressive. Then you'd ask the question that mattered: where is this coming from?
The answer was usually some variation of "general market standards" or "common contract best practices" or, worse, a hallucinated reference to a doctrine that doesn't exist. Sometimes it was a citation to a case the model invented. Sometimes it was just confidence with no source.
This isn't a small problem. It's the entire problem. You can't bring that to a client. You can't bring it to a counterparty. You can't bring it to a partner who's going to ask "why are we pushing back on this?"
"Grounded" is a real technical term, not marketing. It means every output the model produces is tied to a specific piece of source material, and the model can show you what.
For a contract platform, grounding has three layers:
If any of those three are missing, the answer is at best generic. At worst, it's wrong in a way that sounds right.
Two side-by-side examples on the same change to an indemnification clause.
"This clause is unusual. Most MSAs include broader indemnification. We recommend adding a carve-out for IP infringement."
"Section 9.2 removes the IP infringement carve-out. Your firm's playbook (clause 14, IP carve-out standard) requires uncapped indemnity on IP infringement claims for SaaS clients. Suggested counter: '...except for claims arising from third-party allegations of IP infringement, for which liability shall be uncapped.' Citation: source clause snippet attached below."
One is something you'd quote in a marketing brochure. The other is something you'd put in a return-to-client email. The difference is grounding.
An attorney making a mistake is fixable. The attorney can revisit, correct, and learn. AI hallucination is structurally different because the model produces wrong answers with the same confidence it produces right ones. There's no signal in the output telling you which is which.
The classic failure mode in legal AI is the invented citation. The model writes "as established in Cardinal Industries v. Ohio, 312 F.3d 451 (2002)..." and the case doesn't exist. The attorney who copies that into a brief without checking has a malpractice problem.
Contract review hallucination is subtler but works the same way. "Standard market practice requires X" sounds authoritative. If the model invented "standard market practice" — and it might have — the attorney who relies on that is building a recommendation on a fiction.
Grounding is the structural fix. A model that has to cite its source can be wrong, but it can't be confidently wrong about something it never had a source for in the first place.
Grounded AI also enables a quality bar that ungrounded AI can't reach: server-side verification. If every insight has a cited snippet, the system can check that the snippet actually appears in the section the insight is anchored to. If it doesn't, the insight gets dropped or re-anchored before the attorney ever sees it.
This catches a specific failure mode: the model writes an analysis of Section 6 but anchors it to Section 2 because the model lost track of the section numbering halfway through the document. The reviewer reads about Section 6 while looking at Section 2 and gets very confused. Grounding makes this catchable.
You can't do this kind of verification on ungrounded output. You have no way to mechanically check whether "this clause is unusual" is actually about the clause the model claims it's about.
Beyond the immediate review, grounded AI gives you something attorneys and risk teams have always wanted from contract platforms: a defensible record of what the model recommended, what the attorney did with it, and why.
Every insight in a grounded system can be marked as applied (you used the suggestion), ignored (you decided not to), or copied (you used the text but adapted it). Six months later, when somebody asks "why didn't we push back on that indemnity carve-out," you have an answer. It's not "I don't remember." It's "the model flagged it as medium-risk; we accepted because the client wanted to close fast; here's the timestamp."
For an outside counsel attorney working on retainer or hourly, this audit trail also matters for billing visibility. The client can see what work the attorney did, what the AI surfaced, what the attorney chose to apply. The opacity that has historically protected billable work also fed the perception that attorneys were a black box. Grounded AI lets you choose to be transparent.
Building grounded AI on contracts is harder than wiring up a chat interface over a foundation model. You need:
Most AI contract products on the market built the chat interface and shipped. Grounding requires the boring infrastructure underneath, which means it takes longer, costs more, and doesn't demo as well in a 60-second product video.
The trade-off, of course, is that the grounded version is the only one an attorney can actually use. The rest is theater.
See what grounded AI looks like on your actual contracts. Drop a redline in Midly and read the citations yourself.
Attorney-built · Playbook-grounded · 14-day free trial