How-to

The AI-native legal team - redesigning contract review

Christopher Kliebenstein · July 15, 2026

A GC we talked to rolled out an AI contract reviewer that flagged the risky clauses in an NDA in about a minute, work that used to eat a first-year associate's afternoon. A year later the average NDA still took the better part of two weeks to sign. The AI read the contract fast. It then landed in the same queue. Same three approvals as always, and nobody had rethought them.

How a general counsel rebuilds the contract workflow so the speed the AI buys actually reaches the business.

Short answer: An AI-native legal team redraws who decides on a contract, from the first read through to sign-off. Generative AI use among general counsel nearly doubled to 87% in 2026, yet only 47% of departments run it department-wide, and just 17% of C-suite leaders say legal's AI use moves the business.

The tool got faster. The chain of sign-offs behind it did not, which is why the speed shows up in the demo and disappears in the cycle-time report. We rebuild operational workflows for operators, and the legal ones stall for the same reason the finance and hiring ones do: someone automated the read and left the decision path exactly where it was.

Why doesn't faster contract review fix legal's bottleneck?

Because the read was never the slow part. A contract waits on the sign-off chain, the escalation path, and the outside-counsel handoff. None of that moves because a model parsed the clauses in seconds.

Look at where the time goes. LegalOn's 2026 survey of 452 legal professionals found teams spend an average of 3.1 hours reviewing a single contract, and 79% report AI has cut the time they spend on routine tasks. The ACC's 2026 survey of 1,049 chief legal officers found 81% name greater speed as the top benefit they get from generative AI. The gain is real, and it is concentrated on the first read. The contract still routes through the same approvers on the same cadence once the read is done.

This is the pattern behind the broader miss. MIT's Project NANDA found that roughly 95% of generative AI pilots deliver no measurable P&L impact. A faster read in front of an unchanged sign-off chain is what a stalled pilot looks like inside a legal department. We made the same argument for the finance close in how to rebuild a finance function for AI and for the hiring shortlist in why hiring is the workflow every AI-native redesign skips. Contract review is the legal version.

What is an AI-native legal team versus AI tools bolted onto legal?

Bolting AI on keeps the intake-to-signature sequence and the same people deciding, and speeds up one box inside it. An AI-native legal team changes who decides what and when. It moves whole spans of the review to the machine and redraws where a lawyer actually has to weigh in.

Adoption has already gone mainstream, and it is still mostly the bolted-on kind. FTI Consulting and Relativity's General Counsel Report found 87% of GCs now report AI use, up from 44% a year earlier, with 53% now running a formalized technology roadmap. Even the largest rollouts are tool deployments at scale rather than redrawn decision rights. A&O Shearman put Harvey in front of about 4,000 staff across 43 jurisdictions, and the parallel Harvey and PwC alliance extends access to more than 4,000 legal professionals across over 100 countries. Thousands of lawyers now have a faster first read. That is a large purchase of speed, and it does not on its own redraw who signs.

The disconnect shows up when you ask whether any of it reaches the business. Thomson Reuters found 86% of GCs believe legal is a significant business contributor, while only 17% of C-suite executives agree, and just 47% of departments have adopted AI department-wide. The gap between what legal feels it delivers and what the C-suite sees is the gap between a faster read and a redrawn workflow. We mapped the general version of this across every function in the AI-native operating model, function by function.

Isn't AI already more accurate than lawyers at reviewing contracts?

On a narrow, well-defined task, the early evidence says it can hold its own. That strengthens the case for a redesign, because accuracy on the read was never what held the contract up.

The foundational data point is old and worth reading as a starting marker rather than a current benchmark. In a 2018 LawGeex study, an AI system hit 94% average accuracy identifying risks across five previously unseen NDAs against an 85% average for 20 experienced lawyers, and finished in 26 seconds against a 92-minute human average. The task was bounded: known clause types, a single document, a clear risk checklist. Systems have moved on since, and so has the surface area they are asked to cover.

Take that at face value and the conclusion still points the same way. If a machine can clear the first read of a standard agreement quickly and accurately, the value of a redesign sits in what happens after the read. The seconds it saves were never the constraint. A GC who keeps optimizing the read is polishing the part of the process that was already fast.

What does an AI-native contract redesign actually change?

It moves the first pass to whoever owns the need, and it puts the whole contract on one shared system so the handoffs stop hiding the delay. Two examples come closest, and neither is a full teardown.

The clearest change in decision rights is at Bridgewater. In Harvey's own account, Bridgewater moved the first pass of vendor-contract review to procurement itself, cutting average vendor contract review from two days to two hours. Read it as a vendor-published figure. The mechanism is the part that matters: the review no longer starts by landing in the legal queue. The team that needs the contract runs the first pass, and legal weighs in where the risk is real.

The clearest change in visibility is at Hormel. Hormel replaced an all-manual, Excel-tracked process with one shared contract system across legal, procurement, sales, and IT, and cut its contract-processing period by about 75%, from roughly three months to roughly three weeks. No model rewrote a clause there. Putting every function on the same system removed the waiting between them, which was most of the three months.

Be honest about the gap. There is no single flagship example yet with a clean before-and-after on both cycle time and how the roles shifted, the way the finance rebuilds show. Bridgewater redrew who does the first pass. Hormel redrew where the work lives. A full AI-native redesign does both at once, and most departments have done neither.

Rethinking your contract workflow? Sign up for the newsletter to get the next piece.

Who is liable when AI-assisted contract review or legal research gets it wrong?

The lawyer and the firm own the output, and the courts are making that point at speed. Redrawing who reads the contract does nothing for you if you have not redrawn who is accountable for what the machine produced.

The count is climbing fast. Damien Charlotin's AI Hallucination Cases database, which tracks court cases involving AI-fabricated citations, has run into the 1,300 to 1,700-plus range by mid-2026 and is growing by roughly eight new cases a day. Fortune reported in May 2026 on the wave of courtroom sanctions that followed. Speed without redrawn accountability is the failure mode, and the AI is rarely the party the judge sanctions.

The teams doing this well keep a person on the hook by design. LegalOn found 80% of legal teams are exploring AI agents, and they consistently prefer supervised, human-in-the-loop automation over letting an agent run unattended. That instinct is right, and it only holds if the workflow names who supervises which decision. We worked through the general version of this question in who owns the mistake when your AI agent gets it wrong.

What happens to paralegals and junior associates in an AI-native legal team?

The base of the work changes first. The high-volume review and the routine drafting that filled a paralegal's day and a first-year's week are the parts a model takes over soonest.

The labor data already shows the pressure. The U.S. Bureau of Labor Statistics projects little or no change in paralegal and legal-assistant employment through 2034, with about 39,300 annual openings coming mostly from replacement, and it names AI and large-language-model adoption as the reason demand is limited. BLS expects lawyers to be less affected. The judgment work and the accountability for the call stay with people.

The mistake is to bank the freed hours as a headcount cut and stop there. Redeploy the junior time into the work the machine cannot own, the escalations and the judgment calls, and the department gets more capable. Cut it, and you own a cheaper version of the same review process, one paralegal short. We do not have a hard number that says the legal pyramid inverts, so treat this as the direction of travel rather than a measured shift.

Where should a GC start rebuilding the contract workflow?

Start with who signs, before you choose a reviewer. The sequence we run inside a legal function looks like this.

Map who decides, and when. Before touching a tool, mark every point where a human sign-off gates the contract: the first legal review, the risk escalation, the business owner's approval, the outside-counsel handoff. Those gates are the cycle time. The read sits off to the side of them.
Move the first pass to whoever owns the need. Let procurement or the deal team run the first read against an agreed standard, the way Bridgewater did, so the contract does not begin its life waiting in the legal queue.
Route by risk. Send the standard, low-risk agreement down a fast path the machine and the business owner can clear. Reserve the lawyer's time for the terms that actually carry exposure.
Put the whole contract on one system. Hormel's 75% cut came from every function working in one place instead of an inbox and a spreadsheet. Shared visibility is where most of the waiting disappears.
Name a human on every escalation and handoff. Keep a supervising lawyer accountable for the risk calls and for whatever the AI produces. The hallucination sanctions are the price of leaving that seat empty.
Measure cycle time and accountability. Read speed is the vanity metric a reviewer improves while the contract still takes two weeks. Track how long the contract takes end to end and who owns each decision along the way.

The through-line is that the intelligence belongs in the decision structure as much as in the read. Most programs buy a smarter reviewer and leave the same sign-off chain behind it, which is the legal version of the mistake we mapped in the process map every AI-native redesign starts with.

What does the GC actually decide?

An AI-native legal team comes down to four calls, and they belong to the GC or CLO. No vendor makes them for you.

Which reviews can the business run without legal in the room? Draw the line between the standard agreement a deal team can clear against a written standard and the terms that need a lawyer. Draw it too narrow and you keep the bottleneck. Draw it too wide and you own the risk nobody reviewed.
Who supervises the machine's output? Name the human accountable for every AI-assisted read and filing. The courts are already assigning that accountability whether or not you have.
What do you measure? Report end-to-end cycle time and stop rewarding read speed. The metric you pick is the behavior you get.
Who owns the redesign end to end? The contract crosses legal, procurement, sales, and the business owner, and the delay lives in the seams between them. Name one owner for the whole workflow, or no one owns the handoff that stays slow.

Answer those four and cycle time moves because the decisions moved. Skip them and you have a fast reviewer in front of a sign-off chain that was the bottleneck all along.

Frequently asked questions

Will AI replace lawyers, paralegals, or in-house counsel? It replaces tasks and reshapes roles. The BLS projects little or no change in paralegal employment through 2034 and names AI adoption as the reason demand is limited, while expecting lawyers to be less affected. The high-volume review and routine drafting move to the machine. Judgment, negotiation, and accountability stay with people.

What is an AI-native legal team versus AI tools bolted onto legal? Bolting AI on speeds up one step, usually the first read, inside the existing intake-to-signature process. An AI-native team redraws who decides what and when, moving the first pass to the business owner and reserving lawyers for real risk. Only 47% of departments run AI department-wide, so most are still in the bolted-on stage.

Is AI more accurate than lawyers at reviewing contracts? On a narrow task, early evidence says it can match or beat them. In a 2018 LawGeex study, an AI system hit 94% accuracy against an 85% lawyer average on NDA risk review, in 26 seconds against 92 minutes. Read that as a bounded, dated benchmark. Accuracy on the read was never the workflow's bottleneck.

Who is liable when AI-assisted contract review or legal research gets it wrong? The lawyer and the firm. Damien Charlotin's database of AI-fabricated-citation cases has passed well over 1,300 by mid-2026 and grows by roughly eight a day, and courts are issuing sanctions. If a model produces the work, the person who filed it owns the outcome.

Where should a GC start rebuilding contract review? With who signs, before you choose a reviewer. Map every human sign-off in the contract's path, move the first pass to the business owner, route by risk, and put every function on one shared system. Hormel cut its processing period about 75% mainly by doing the last part.

Before you buy your next contract-review tool

Redraw who signs first. We write about how operators rebuild functions to be AI-native, the same operating calls a GC, COO, or CEO actually has to make, without the hype and without the tutorials. Join the newsletter to get the next piece when it is out.

Sources

FTI Consulting and Relativity, "The General Counsel Report" (Part 2, March 2026). 87% of GCs report AI use, up from 44%; 53% run a formalized tech roadmap. globenewswire.com
Thomson Reuters Institute, "2026 State of the Corporate Law Department Report." 47% department-wide AI adoption; 86% of GCs see legal as a significant business contributor versus 17% of C-suite executives. thomsonreuters.com
ACC, "2026 Chief Legal Officers Survey" (1,049 CLOs, 43 countries). 52% actively using generative AI; 81% cite greater speed as the top benefit. acc.com
LegalOn Technologies, "2026 State of AI for In-House Legal" (452 legal professionals). 3.1 hours average per contract review; 79% report reduced time on routine tasks; 80% exploring AI agents, preferring human-in-the-loop. businesswire.com
U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, "Paralegals and Legal Assistants" (2024-2034). Little or no change projected; ~39,300 annual openings; AI cited as limiting demand. bls.gov
LawGeex, "AI vs. Lawyers" NDA-review benchmark (2018, foundational and dated). 94% AI accuracy versus 85% lawyer average; 26 seconds versus 92 minutes. prnewswire.com
Legal Dive, "What Hormel Foods learned from its contract management overhaul." Processing period cut ~75%, from ~3 months to ~3 weeks, on one shared contract system. legaldive.com
Harvey (vendor case study), "How Bridgewater Uses Harvey." Vendor-contract first pass moved to procurement; review cut from 2 days to 2 hours; quote from CLO Tracey Yurko. Vendor-published. harvey.ai
A&O Shearman and Harvey partnership announcement. ~4,000 staff across 43 jurisdictions; parallel Harvey and PwC alliance across 100+ countries. aoshearman.com
Damien Charlotin, AI Hallucination Cases Database, mid-2026. damiencharlotin.com Paired with Fortune, "AI hallucinations and courtroom sanctions" (May 2026). fortune.com
MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025," via Fortune. fortune.com

By Christopher Kliebenstein. We build and run AI-native workflows for operators who need them working in production.