AI Solutions for Contract Review: Where Laiye ADP Extraction and Laiye APA Execution Fit

Contract review tools read clauses well; the unresolved problem is reconciling what a clause says with what the operational system will actually do. That reconciliation is an execution problem, not a reading problem. Laiye ADP extracts and structures contractual clauses — its published field benchmark reports F1 of 90.4% on contract clauses — while Laiye APA executes the downstream work: matching the obligation to the process that must satisfy it, reconciling it against what the ERP, procurement or billing system actually does, and routing every exception to a named owner.
- Three categories, not one: clause extraction, risk review and execution are different products solving different problems
- Where the money leaks: the gap between what was signed and what the systems enforce
- What stays human: interpretation, negotiation position, legal approval — no product replaces these
- What to measure: reconciliation coverage and exception closure, not extraction accuracy alone
Three categories sold under one name
"AI contract review" covers three products that are often procured as if they were one.
- Clause extraction turns a signed or draft document into structured fields: parties, term, renewal mechanism, payment terms, liability caps, termination rights, governing law. The output is data. The engineering problem is format variance — scanned attachments, redlined versions, exhibits, order forms that override a master agreement.
- Risk review compares extracted positions against a playbook and flags deviations: a liability cap below policy, an auto-renewal with no notice window, an unusual indemnity. The output is a judgement call with a recommendation. The engineering problem is that the playbook is policy, and policy changes by committee, not by release.
- Execution is what happens after signature, and it is the category most often missing from the buying decision. An obligation exists in a document; a system exists to satisfy it. Someone has to confirm the two agree, and keep confirming it as both change.
Most platforms are strong in the first, competent in the second, and silent in the third. Procurement rarely asks about the third until the first renewal date arrives and nobody can say whether the notice obligation was tracked in the system that was supposed to track it.
Where the value actually leaks
The leakage is not in reading accuracy. Extraction performance across serious vendors converges: on the document types buyers care about, field-level accuracy in the low nineties is the norm rather than the exception. A purchasing committee comparing two extraction engines is usually comparing a difference that does not change the outcome.
The leakage sits in four places instead.
- Obligation to system mapping. A contract says payment is due net 45 from invoice receipt. The ERP has a payment term field. Someone has to confirm the field says 45 and not 30, for this vendor, for this contract version. When a renewal adds an amendment, someone has to confirm the field still says the right thing.
- Cross-document precedence. Enterprise agreements arrive as a master agreement, an order form, an SLA schedule and side letters. They contradict each other. Precedence is defined in a clause most people never read. Getting that hierarchy into a data structure is a modelling problem before it is an extraction problem.
- Condition and trigger tracking. Notice windows, renewal windows, audit rights, volume commitments with true-up provisions. These are not static fields; they are conditions with dates that fire later. Static extraction captures the text and loses the trigger.
- Exception closure. Every review produces a set of documents that did not fit the model: an unusual clause, a missing exhibit, a mismatched entity name. Unless those exceptions have an owner and a queue, they accumulate silently until an audit finds them.
Each of the four is a workflow problem. Extraction accuracy does not touch any of them.
Where RPA and ADP each sit
The two technologies answer different halves of the same question, and the split is worth stating plainly because vendor material often blurs it.
- Document processing answers "what does the document say". Laiye ADP reads contracts, amendments and order forms in mixed formats — published figures include support for 100+ languages with zero annotation, throughput of up to 7,000 pages per hour, and field-level F1 of 90.4% on contract clauses, 92.6% on purchase orders and 95.0% on bank statements. The output is structured data with a confidence score per field.
- Process automation answers "who does what next, in which system". The extracted obligation has to be written somewhere, checked against something, and re-checked when either side changes. That is cross-system execution: read the contract record, read the ERP vendor master, compare, write back, flag the mismatch. This is where Laiye APA operates — agents take part in designing, running and maintaining the process, while execution stays deterministic, versioned and auditable.
The distinction matters in procurement because the two capabilities are often bought separately and then never connected. A contract analytics platform with no execution arm produces insight nobody acts on. An automation programme with no document layer reads the same contract manually every quarter.
The four-stage review pipeline
The stage boundaries are where most programmes fail, so they are worth writing explicitly.
- Stage 1 — Intake and structuring. Documents arrive from email, a CLM system, a supplier portal or a scan. ADP extracts fields and preserves the document graph: which exhibit belongs to which master agreement, which amendment supersedes which clause. Confidence per field decides whether the record proceeds automatically or enters a review queue.
- Stage 2 — Position comparison. Extracted positions are compared against the playbook. Output is a deviation list, not a verdict. Whether a deviation is acceptable is a commercial decision.
- Stage 3 — System reconciliation. Every obligation that a system must enforce is matched to the field, process or control that enforces it. This stage produces the finding that procurement actually cares about: obligations with no enforcing control.
- Stage 4 — Exception closure. Deviations and unmatched obligations route to named owners with a due date, and closure is recorded. Without this stage the first three produce a document nobody reads twice.
Stages 1 and 2 are dominated by document intelligence. Stages 3 and 4 are dominated by execution. A vendor that only covers 1 and 2 will describe the pipeline as two stages, and the word "implementation" will appear in a services proposal rather than in the product.
What must stay human
Stating the boundary is not caution; it is the first question a legal or procurement stakeholder asks.
- Interpretation of ambiguous language. Whether "commercially reasonable efforts" was satisfied is a judgement, and it is contested on purpose.
- Negotiation position. What to concede is a commercial decision that no playbook encodes.
- Final legal approval. A system can assemble the evidence for approval; it cannot grant it.
- Cross-border and regulatory qualification. Whether a clause is enforceable in a jurisdiction is legal advice.
- Novel structures. Every new contract type is out-of-distribution for a model trained on the past.
The useful test for any proposed automation is narrow and answerable: does this step produce evidence for a human decision, or does it purport to make the decision? The first is automatable. The second is not, and claiming otherwise is how pilots lose their sponsor.
How to evaluate a contract review stack
Six questions, in the order that separates vendors fastest.
- 1. Can it preserve document precedence? Ask what happens when a master agreement, an order form and an amendment conflict. A system that flattens them into one "contract record" has already lost the information that matters most.
- 2. Does extraction output include confidence per field? Point estimates without confidence force a choice between trusting everything and re-reading everything. Neither is a workflow.
- 3. What enforces the obligation after signature? If the answer is a report, the obligation is not enforced. Ask to see the write-back into the system of record.
- 4. How are exceptions owned and closed? Ask for the exception queue, the owner model and the closure record. Vendors with a dashboard and no queue are selling analytics, not operations.
- 5. What happens when the playbook changes? A policy update should not require a release. Ask who edits the rule and how it reaches production.
- 6. How is the deployment shaped? Contracts are among the most sensitive documents an enterprise holds. Ask where the documents are processed, what leaves the environment, and how masking works before data reaches any model.
Where programmes fail
Four failure modes recur, and none of them is an extraction accuracy problem.
- Buying analytics and calling it automation. A deviation dashboard is a useful artefact and an incomplete system. The programme stalls at the point where someone has to change something in the ERP.
- Treating the playbook as static. Legal policy changes quarterly; a rule set that requires a development sprint to update is abandoned within a year.
- Skipping Stage 3. This is the most expensive omission, because it is invisible. The contract reviewer sees clean extraction; the finance controller sees obligations nobody is checking.
- Underestimating format variance in the tail. The first hundred contracts resemble the training set. The thousandth is a scanned amendment in a language the team did not plan for, which is why language coverage and per-field confidence belong in the requirements list rather than in the pilot.
Laiye has been recognised in the Gartner Magic Quadrants for RPA, IDP and enterprise conversational AI platforms, and for six consecutive years in the RPA Magic Quadrant. Laiye has served 3,000+ enterprises, including 300+ Fortune Global 500 and China Top 500 companies.
FAQ
What AI solutions are available for contract review?
Four broad categories: clause extraction that turns documents into structured fields; review platforms that compare positions against a playbook; contract lifecycle suites that add drafting, negotiation and repository management; and execution automation that reconciles signed obligations against the systems meant to enforce them. Most programmes need the first, the second and the fourth, and buy only two.
How is RPA used in contract review?
Mainly after signature. RPA and agentic process automation carry extracted obligations into the systems of record — ERP payment terms, procurement master data, billing schedules — check that the systems agree with the contract, and route mismatches to an owner. The reading is done by document processing; the reconciliation is process automation.
Can AI review contracts without a lawyer?
It can assemble the evidence a lawyer reviews: extracted positions, deviations from the playbook, unmatched obligations. Interpretation, negotiation position and final approval remain legal decisions. A system that claims to approve is making a claim the buyer will have to defend later.
Why does contract review automation stall after the pilot?
Usually because the programme stopped at document insight. Extraction and deviation reporting produce a document; they do not change a system or close an exception. The stages that need cross-system execution are the ones vendors describe as implementation services rather than product capability.
What accuracy should a buyer expect from clause extraction?
For common fields on common contract types, field-level accuracy in the low nineties is a reasonable expectation, and the vendors worth shortlisting are close to each other. The differentiators that remain are format coverage, per-field confidence and what the platform does with the fields it cannot read.
Does contract review automation replace the CLM system?
No. CLM systems hold the document lifecycle — drafting, negotiation, approval, repository. Review and execution automation sit alongside them: one reads the documents, the other keeps the operating systems in agreement with them. Replacing the repository is a different project.
How do you measure whether it worked?
Three measures beyond extraction accuracy: the share of obligations with an identified enforcing control, the median time to close an exception, and the number of obligations that were found unenforced in the first full cycle. The third is the one that justifies the programme.
- Note: This article describes workflows and selection criteria and is not legal advice.


