AI contract review has quietly crossed a threshold. The current generation of tools reads a purchase agreement or a credit facility and produces what a mid-level associate would recognize as a competent first pass: clauses extracted and classified, deviations from a playbook flagged, unusual terms surfaced with an explanation of why they are unusual. The technology question is largely settled. The questions that remain are about judgment, and about boundaries.
What review AI does well now
Used honestly, the tools are good at four things. Issue spotting against a playbook: given a set of preferred positions and fallbacks, the system checks every draft against them and flags where the paper lands outside the range. Clause extraction and comparison: pulling indemnification, limitation of liability, termination, assignment, and change of control provisions out of a stack of agreements and lining them up side by side. Deviation detection: measuring a counterparty draft against the firm's standard and identifying what moved, including the quiet edits in definitions that change the economics of a clause three sections later. And summarization of atypical terms: a readable account of what is strange in this contract, which is often what the client wants to know first.
The practical effect is felt in the first hours of an engagement rather than at the end. A stack of forty ancillary agreements that would have absorbed two associate days of triage comes back mapped before the kickoff call, with the handful of documents that deserve senior attention already identified. Nothing about that removes the attorney from the work. It changes where the attorney's hours land.
What the tools do not do is decide. Whether a nonstandard indemnity cap is acceptable depends on the deal, the client's risk tolerance, the relationship, and what was traded away for it last Tuesday. That call belongs to the attorney, and every workflow described below keeps it there.
The boundary problem
Most contract review products assume a simple architecture: upload the documents to the vendor's cloud, let the models run, read the results. For consumer software that assumption is invisible. For a law firm it is the beginning of a series of uncomfortable conversations.
Outside counsel guidelines increasingly prohibit sharing client materials with third-party processors without prior written consent, and a growing number name AI tools explicitly. Client security teams audit their law firms now, and a vendor processing environment is one more system on the audit. Privilege and confidentiality arguments, while generally defensible when a vendor acts as the firm's agent under proper terms, are simply cleaner when the paper never leaves the firm's control. And cross-border work adds transfer complications: a European target's documents flowing to a US vendor's cloud raises questions nobody on the deal team wants to field during signing week.
None of this is theoretical. It surfaces as consent exercises across dozens of clients, carve-outs written into engagement letters, and matters where the tool simply cannot be used, which are usually the largest ones.
What changes when review runs inside your tenant
Tenant-resident review inverts the architecture. The review engine deploys inside the firm's own cloud boundary, the documents stay in systems the firm already controls, and the models run where the documents are. That one design decision changes three conversations.
The client consent conversation gets simpler. We review your documents with software running inside our own environment, under the same controls as our document management system: that is a sentence a general counsel can approve on a call. We send your documents to a vendor, which sends them to a model provider, under two data processing agreements: that is a meeting, at best, and a refusal often enough.
The audit conversation gets shorter. There is no third-party processing environment to assess because there is no third-party processing. The firm's existing certifications, network controls, and logging cover the runtime.
And the precedent conversation becomes possible at all. A review engine inside the boundary can draw on the firm's own negotiated history: your fallback positions, your prior paper with this counterparty, what market actually looks like across your last fifty deals in the sector, measured from your own documents rather than a generic survey. That knowledge cannot leave the building. With a tenant-resident engine, it does not need to.
Accuracy is a context problem
The accuracy ceiling on generic review tools is not model quality. It is context. A pattern-matched flag says this indemnity clause looks unusual relative to a training distribution. A grounded flag says this clause conflicts with the definition of Losses in this draft, sits outside the position your firm accepted in its last three deals with this sponsor, and reverses a change your own team made in draft four.
The second kind of flag requires the engine to know the defined terms, the deal structure, the drafting history, and the firm's precedent. That is what a Context Graph provides. Reframe structures matters, parties, documents, clauses, and negotiated positions into a per-firm graph the review engine queries as it reads, an approach described in the platform overview. Review grounded this way produces fewer flags and better ones, because it evaluates the clause in the deal rather than the clause in isolation.
The same grounding disciplines the tool's confidence. When the system cannot trace a flag to a specific provision, a defined term, or a prior position, it should say so and route the question to a person, rather than dressing a guess in the costume of an answer. Calibrated uncertainty is worth more to a deal team than fluent overstatement.
A generic playbook tells you what the market does. Your own precedent tells you what your firm does. The second one is what clients are paying for.
A workflow that survives contact with practice
The firms getting durable value from review AI run it inside a disciplined loop rather than as a button:
- Triage tiers. Routine paper, NDAs and standard vendor agreements, gets machine review with attorney spot checks. Negotiated documents get machine-assisted review where the tool prepares and the attorney decides. Bet-the-company paper gets full attorney review with the machine as a second set of eyes.
- Every red flag gets a human. The machine can clear routine clauses into a sampling regime, but nothing it flags as a problem is resolved without an attorney's decision on the record.
- An audit trail by default. What the system flagged, what the attorney accepted, overrode, or escalated, and when. That record is protection in a malpractice conversation and, increasingly, an expectation in client audits and under emerging AI governance frameworks.
- Overrides feed the playbook. A flag that partners override on every deal is a playbook error. The system should learn from it inside the firm's boundary, not inside a vendor's shared model.
Two numbers tell you whether the loop is healthy: the override rate on flags, reviewed monthly by practice group, and the escape rate, meaning issues surfaced later that the machine should have caught. Firms that track both learn quickly where the tool is strong, where the playbook needs tightening, and where the machine should not yet be trusted.
Governance is what makes the loop trustworthy at scale. Review runs as a governed workflow with scoped permissions and logged actions, the discipline behind Reframe's agent harness, with the runtime controls described on our security page.
Where this goes
Review is one segment of a longer workflow, and the boundary logic extends in both directions. The same tenant-resident foundation supports drafting assistance inside Word, where attorneys actually work, and diligence at deal scale, where document volume makes the boundary question unavoidable. The pattern is the same in each case: the models come to the documents. The documents do not go to the model's owner.
Firms evaluating review tools this year should put the boundary question first, because it is the one that cannot be patched later. Accuracy improves with every model release. An architecture that ships client paper to someone else's cloud is permanent until you replace it. Outside counsel guidelines are moving in one direction on this point, and the firms that can answer them with architecture rather than assurances will keep the work.