Who's liable when in-house counsel signs off on an AI-drafted contract that turns out wrong?

The shift

Drafting and reviewing a contract — reading it against precedent, spotting the off-market clause, generating a redline — moves from scarce, hours-metered legal labor to abundant, near-instant, near-free output. What does not move is accountability: the signed contract binds the company, and answerability for it stays attached to a licensed human who cannot delegate liability to a model.

The axioms

1. The GC team's value is what it produces — drafting and review hours on contracts.
2. To handle more contract volume, you add more lawyers; throughput scales with headcount.
3. The person who signs off on a contract is a licensed human who is accountable for it.
4. Deciding whether a contract's terms are acceptable to the business is judgment on legal and commercial risk.
5. The business trusts Legal because someone there owns the downside when a deal goes wrong.
6. Output that goes out under the GC's name has been verified by someone who understands it.

Invalid axioms

1. The GC team's value is the drafting and review hours it produces. Generating a competent first draft and a first-pass redline is now abundant and near-free. The habit-trap: teams still measure and defend their worth by throughput — contracts turned, hours logged, matters closed — and the business still routes work to Legal as if producing the document were the scarce part, when producing it is now the cheap part and judging it is the expensive part.
2. Contract volume is handled by adding lawyers; throughput scales with headcount. Drafting and review capacity no longer tracks the number of people. The habit-trap: budgeting and staffing the function around document throughput, so headcount requests get justified by contract count — a metric a model now moves without a new hire, leaving the team sized for a bottleneck that has shifted from producing contracts to verifying them and owning the risk.

Unchanged axioms

1. The person who signs off is a licensed human who is accountable for it. Liability attaches to a person and a company, not a tool. A model cannot be sued for malpractice, sanctioned, disbarred, or held to have breached a duty of care — when an AI-drafted contract turns out wrong, the accountable party is still the counsel who approved it and the company that signed. This is the load-bearing answer to the question: sign-off does not transfer liability to the vendor or the model, it concentrates it on the human who signed.
2. Deciding whether the terms are acceptable to the business is judgment on legal and commercial risk. Whether an indemnity cap is worth accepting, whether a liability carve-out is defensible against this counterparty, whether a clause the model flagged as standard is actually a problem given this deal — that is judgment under stakes specific to the company, not pattern-matching against what most contracts say. The model can surface the issue; it cannot decide the company's appetite for the risk or answer for the call.
3. The business trusts Legal because someone there owns the downside. The value of a sign-off is that a person with standing and skin in the outcome put their name on it. That trust rests on accountability, and accountability did not get cheaper — a system with no liability exposure cannot hold it, so the relationship stays human even when the drafting does not.

New axioms

1. When drafts are free, verification has to happen at the volume AI produces, and nobody has staffed the verifying seat. The constraint moves from writing the contract to checking it — and the checking load scales with how much the model now generates, which is more than before. The old ratio of one reviewer per drafting bottleneck no longer covers it, and the verifying role is neither priced nor headcounted as the thing that actually gates safe sign-off.
2. Automation bias makes a plausible-wrong contract the characteristic failure. A model's output reads clean, cites the right-sounding clauses, and looks like the last hundred contracts — which is exactly what makes a reviewer skim it and sign. The failure mode is no longer the obvious error a lawyer catches; it is the confident, well-formatted clause that is subtly wrong for this deal and gets approved because nothing looked off. Verification designed to catch sloppy work does not catch this.
3. The liability chain between GC, outside firm, and model vendor is unsettled. Today the answer is clean — it lands on the signing human and the company. But vendors are selling contract AI with reliability claims, firms are deploying it on client matters, and there is no settled allocation of fault when the tool's output is the proximate cause of a bad term. Whether a vendor's terms-of-service disclaimer holds, whether relying on a certified legal AI counts as reasonable care, and where an outside firm's duty sits when it used AI on the file are all open — and this is the fast-moving call, because the first malpractice suit, insurer exclusion, or bar opinion that turns on AI use will reset the default faster than the technology will.

Where it breaks

The team is still measured on throughput — contracts turned, matters closed (invalid) — while the failure that actually creates liability is the plausible-wrong clause that gets signed precisely because a reviewer under throughput pressure trusted output that looked right (new). The metric rewards moving contracts through faster, which is the exact condition under which automation bias signs the wrong one; the function is being paid for speed on the thing that is now cheap and exposed on the thing that is now dangerous.

A second collision: headcount is still justified by document volume (invalid), while the real bottleneck is verification at the scale the model produces and clear ownership of the risk when it is wrong (new). The seat the function needs — someone whose job is to catch the confident-wrong clause and own the sign-off — is not the seat the volume-based budget funds, so the accountable human signs off on more AI output with no more verifying capacity than before, and liability stays fully theirs.

ai-axioms.vercel.app/axioms/legal/legal-gc-liability-ai-axioms