Legal Agents: Contract Review at Machine Speed

Table of Contents

• What a Legal Agent Actually Does
• Why Contract Review Was the First Domino
• The Workflow, Step by Step
• The Pricing Shift: From Billable Hours to Per-Contract
• Where Legal Agents Break
• The Hallucination Tax
• The Playbook Problem
• Build, Buy, or Subscribe
• What the Buyers Should Actually Measure
• Insights Most People Overlook
• References

What a Legal Agent Actually Does

Start with what it is not. A legal agent is not a chatbot that answers "what does indemnification mean?" That is a search box with manners. An agent, in the sense the market now uses the word, is a system that takes a goal, "review this MSA against our standard playbook and return a redline", and executes the multi-step chain needed to reach it without a human holding its hand at each step.

In practice that means the agent ingests the document, identifies the contract type, extracts the clauses that matter, measures each one against a reference standard, drafts proposed language for the deviations, writes a plain-English summary of the risk, and decides whether the result is clean enough to hand back or needs a human reviewer. The "agentic" part is the decision-making and tool-use between those steps, not any single one of them.

This is the same architecture showing up across every vertical in the agentic AI-as-a-service cluster, the difference is that contracts are unusually well-suited to it. A contract is a closed, structured artifact. It has a known shape, a finite vocabulary of clause types, and an organization's own historical positions to measure against. Compare that to, say, a sales-development agent working an open-ended outbound territory, and you can see why legal contract review produces cleaner results sooner.

Why Contract Review Was the First Domino

Three things made contract review the beachhead for legal agents rather than litigation, advisory work, or courtroom strategy.

First, the work is high-volume and repetitive. A mid-sized company runs thousands of NDAs, vendor agreements, and order forms a year, and the overwhelming majority are routine. A 2023 study of in-house legal teams found contract review consistently topping the list of tasks lawyers most wanted to offload, not because it is hard, but because it is endless. Goldman Sachs's widely cited analysis of generative AI's labor impact singled out legal as one of the most exposed white-collar categories, with a large share of tasks amenable to automation. You can read the framing in Goldman Sachs's report on generative AI and the global economy.

Second, the ground truth exists. Every company that signs contracts has accumulated a corpus of its own prior positions, what it accepts, what it negotiates, what it rejects. That history is the training signal. A horizontal model with no access to your fallback positions is guessing; a vertical legal agent wired into your contract management system is checking.

Third, the error has a natural backstop. In most contract workflows there is already a human who signs off. That means an agent can be wrong occasionally without the wheels coming off, as long as it is honest about its confidence and the human review step stays in place. That backstop is exactly what makes "machine speed" tolerable here and intolerable in, say, an unsupervised clinical setting.

The Workflow, Step by Step

Strip away the marketing and a production legal agent runs roughly this loop:

1. Intake and classification. The agent receives a document, often straight from an inbox, a contract lifecycle management (CLM) platform, or a shared drive, and determines what it is. An NDA gets a different playbook than a master services agreement.

2. Clause extraction. It segments the contract into its constituent provisions: limitation of liability, indemnification, governing law, termination, IP assignment, payment terms, data processing. Modern systems do this with high reliability because clause taxonomy is well understood.

3. Playbook comparison. Each clause is measured against the organization's standard. This is the heart of the value. The agent isn't asking "is this clause good?", an unanswerable question in the abstract, it's asking "does this deviate from what we accept, and by how much?"

4. Redlining and drafting. For each deviation, the agent proposes alternative language, ideally drawn from the company's own approved fallback positions rather than invented from scratch. Invented language is where risk creeps in.

5. Risk summary and triage. It produces a human-readable summary, "this MSA caps liability at fees paid in the prior 3 months; our floor is 12 months", and assigns a disposition: auto-approve, route to a junior reviewer, or escalate to counsel.

6. Human-in-the-loop checkpoint. A person reviews the flagged items. Crucially, the good systems surface only what needs human eyes, instead of dumping the whole document back on a lawyer's desk.

The speed gain comes from collapsing what was a sequential, human-paced chain into something that runs in the time it takes to get coffee. A review that took ninety minutes can land in under five. But notice that the human checkpoint never disappears in any responsible deployment, it just moves to the end and shrinks to the exceptions.

The Pricing Shift: From Billable Hours to Per-Contract

This is where the GaaS thesis gets real, and where contract review quietly threatens the economics of an entire profession.

The traditional unit of legal value is the billable hour. A legal agent has no hours. So vendors and forward-leaning firms are repricing the work itself: per contract reviewed, per clause flagged, or, the most aggressive model, per outcome, where you pay only for contracts that clear without renegotiation surprises. McKinsey's analysis of generative AI's economic potential pegged the technology's value to corporate functions in the trillions annually, with legal and compliance among the functions where the per-task cost of knowledge work could fall by an order of magnitude. The macro framing lives in McKinsey's research on the economic potential of generative AI.

For an in-house team, the math is brutal in the best way. If outside counsel charges $400 an hour and a routine vendor agreement eats ninety minutes, that's $600 a contract. A legal agent that does the first-pass review for a few dollars in compute plus a flat platform fee changes the unit economics by two orders of magnitude, and frees the expensive humans for the deals that actually warrant their judgment.

For law firms, the same shift is an existential question. If clients can buy first-pass review as a metered service, the leverage model that funded associate armies starts to wobble. Some firms are responding by becoming agent operators themselves, the services-to-software flip happening across professional services. That move is its own strategy, and a recurring theme across the vertical-agents beat: the firms that survive may be the ones that stop selling hours and start selling outcomes their own agents produce.

The honest caveat: per-outcome pricing sounds clean and is operationally messy. Defining "outcome" for a contract review is harder than defining it for, say, a resolved support ticket. A contract that sails through review but blows up in a dispute eighteen months later was not a good "outcome," but the agent already got paid. Pricing is racing ahead of the measurement that would make it fair.

Where Legal Agents Break

The demos are seductive. The failure modes are specific, and they are worth naming because they determine whether a deployment is a productivity win or a malpractice claim.

The Hallucination Tax

A language model's defining failure is producing confident, fluent text that is wrong. In contract review this is uniquely dangerous because wrong text looks exactly like right text, a fabricated citation to a non-existent statute, a redline that subtly inverts the meaning of a liability cap, a summary that misses the one carve-out that matters.

The infamous 2023 case where lawyers submitted a brief citing fabricated cases invented by ChatGPT, and got sanctioned for it, is the cautionary tale every legal-tech buyer now knows. It also explains why serious legal agents constrain the model heavily: retrieval over the actual document, citation to specific clause locations, and proposed language drawn from approved libraries rather than free generation. The agents that work treat the LLM as a reasoning engine over verified facts, not as an oracle. This is the same reliability discipline the whole agent-reliability conversation in GaaS revolves around, just with higher stakes because the artifact is legally binding.

The Playbook Problem

The dirty secret of legal agents is that most of their value depends on something the buyer often doesn't have: a clean, codified playbook. Plenty of legal teams carry their standards in senior lawyers' heads, not in a structured document. An agent can only measure deviations against a standard that's been made explicit.

So the real first deliverable of a legal-agent deployment is frequently playbook extraction, turning tacit institutional knowledge into machine-readable policy. That's slow, political (lawyers disagree about positions), and it's where projects stall. Vendors that quietly do this work as part of onboarding succeed; vendors that assume the playbook exists ship a tool nobody can use. This is the "last mile" of domain expertise that separates a working vertical agent from a clever demo.

Build, Buy, or Subscribe

Three paths, and the right one depends almost entirely on volume and sensitivity.

Buy a vertical SaaS agent, a purpose-built contract-review platform, if you have steady volume and standard contract types. You get the playbook tooling, the integrations, and the guardrails out of the box. You give up control over the model and pay a recurring fee.

Subscribe to a generalist agent platform and configure it for contracts if legal is one of several workflows you're automating and you value consistency across functions over depth in any one. You trade some legal-specific polish for flexibility.

Build only if you have unusual scale, genuinely proprietary contract structures, or regulatory constraints that forbid sending documents to third parties. Building means owning the playbook engine, the retrieval layer, the evaluation harness, and the model updates, a real engineering commitment that most legal departments should not take on. The build-vs-buy calculus here mirrors the broader vertical-agent debate, and the default answer for legal is "buy or subscribe," because the moat in this space is workflow data and integration depth, not the model.

What the Buyers Should Actually Measure

Vendors will quote you accuracy. Accuracy is the wrong headline number, because it hides where errors land. A 95%-accurate agent that misses the occasional liability cap is more dangerous than a 90%-accurate one that flags every cap and over-includes a few false positives. In legal review, false negatives, the risk it didn't catch, are the ones that end careers.

Measure these instead:

• Recall on high-severity clauses. What fraction of genuinely material deviations does it catch? This number, not overall accuracy, is the one that matters.
• Escalation precision. When it routes something to a human, is it usually right that the item needed human eyes? Poor precision here just rebuilds the review bottleneck.
• Calibration. Does the agent's stated confidence track its actual correctness? A well-calibrated agent that says "I'm unsure" at the right moments is worth more than an overconfident one that's marginally more accurate.
• Time-to-redline and rework rate. Speed only counts if humans aren't redoing the work. Track how often a human substantially rewrites the agent's output.

The American Bar Association's guidance on the use of AI in legal practice underscores that the lawyer, not the tool, remains responsible for the work product, which makes these measurement questions a matter of professional duty, not just procurement. The ABA's framing is laid out in its formal opinion on generative AI tools and lawyers' ethical obligations.

Insights Most People Overlook

The agent's best feature is honest uncertainty, not raw accuracy. Buyers fixate on benchmark accuracy scores. But in a workflow with a human backstop, the most valuable trait is calibrated confidence, an agent that reliably knows what it doesn't know lets humans spend attention exactly where it's needed. A slightly less accurate but well-calibrated agent beats a more accurate, overconfident one in any setting where a mistake carries real liability. Vendors don't market calibration because it doesn't fit on a billboard.

Per-outcome pricing in law is structurally premature. The whole GaaS movement is racing toward outcome-based pricing, and legal is held up as a poster child. But a contract's "outcome" only fully resolves over its lifetime, sometimes years later, in a dispute. You can't settle the bill on day one for value that won't be known for eighteen months. Expect the market to retreat from pure per-outcome pricing toward per-contract or hybrid models once the first disputes over what counts as a good outcome hit.

The real moat is the playbook, and clients don't realize they're building it for the vendor. Every contract a legal agent reviews refines the codified playbook, and that accumulated, structured policy data is far stickier than the model. Switching vendors means re-extracting and re-tuning that institutional knowledge. The smartest vendors treat playbook capture as the product. The smartest buyers negotiate ownership and exportability of that playbook up front, because it's the asset that locks them in.

Contract review will hollow out the junior tier before it touches the partners. The conventional fear is that AI replaces lawyers. The accurate near-term read is that it removes the apprenticeship rungs, the routine first-pass review that junior associates learned the craft on. That creates a training gap nobody has solved: if agents do the work juniors used to cut their teeth on, where does the next generation of senior judgment come from? This is a quiet structural problem that the efficiency narrative completely ignores.

Speed creates a new failure mode: nobody reads the green ones. When an agent clears 90% of contracts as clean, the human reviewers stop scrutinizing the clean pile, that's the entire point. But automation complacency means the rare bad contract hiding in the "auto-approve" bucket gets signed unread. The faster and more trusted the agent, the more dangerous its rare misses become, because the human attention that would have caught them has been deliberately withdrawn.

References

• Goldman Sachs: Generative AI Could Raise Global GDP by 7%
• McKinsey: The Economic Potential of Generative AI, The Next Productivity Frontier
• American Bar Association: ABA Issues First Ethics Guidance on AI Tools