Contract review workflow: 42 min to 8 min per contract.

The firm is a mid-sized NZ legal practice with 18 lawyers across two regional offices. The practice has commercial, property, and employment groups, with the commercial group accounting for roughly 60% of fee income. The commercial group's bread and butter is supplier agreements, service contracts, lease reviews, and shareholder agreements for SME clients (typical contract value under $500k, typical review brief tight and turnaround-driven).

The managing partner had been tracking a metric that worried her: average per-contract review time. Across the commercial group, the average was 42 minutes for a standard SME supplier agreement, with significant variance between junior associates (often over an hour, sometimes much longer for unfamiliar agreement types) and senior partners (frequently under 20 minutes for familiar work). The variance was creating two problems. Clients were getting inconsistent turnaround experiences. And the senior partners were absorbing too many "quick reviews" that weren't actually quick.

A previous internal initiative had tried to standardise the review process via template checklists and a shared playbook. That had improved consistency by maybe 10%, but hadn't moved the underlying time-per-contract number much. The team's instinct was that contract review was a fundamentally human, judgment-driven task, and the practice had largely accepted it as a fixed cost of doing the work.

By 2026, two things had changed. First, contract volume was up 25% year-on-year, mostly driven by a couple of bigger SME clients moving more work to them. Second, a competitor practice in the same region had visibly started talking about AI-assisted work in their marketing. The managing partner didn't want to follow the competitor's lead, but she also didn't want to find out 12 months from now that her practice had been beaten on price, turnaround, or both. So she called us.

The brief was unusually crisp. "I don't want a strategy. I want a workflow. Pick the most repetitive, time-consuming, professionally-mid-tier task in this practice that AI could plausibly help with, and let's see if we can move the needle by a measurable amount. Contract review is the obvious candidate. Show me what's possible."

Phase one: Scoping and baseline (week 1)

We resisted the urge to start with tool selection. Instead, we spent the first week mapping the existing contract review process in granular detail. Three current senior reviewers walked us through 12 recent reviews. Two junior associates walked us through five each. We watched, asked questions, and built a process map with timing on every step.

The current process had eight distinct steps: receive contract from client, classify the agreement type, identify the client's underlying commercial intent, identify deviations from the firm's standard position on common clauses, identify any unusual or out-of-norm clauses, draft a markup with proposed changes, draft a covering note for the client, send for partner review on anything above a junior level. Of those eight, three were genuinely judgment-heavy. Five were mechanical, repeatable, and well-suited to AI assistance with appropriate guardrails.

We also pulled a baseline from their practice management system: 200 most recent commercial contract reviews, with reviewer, contract type, review duration, partner sign-off time, and any quality-control flags that had been raised. That dataset gave us the 42-minute average and a granular sense of where time was actually being spent.

Phase two: Tool decision (week 2)

We considered three approaches: a horizontal AI tool (Claude or ChatGPT configured with bespoke prompts), a legal-specific AI platform (several NZ-relevant options), or a custom internal build. The legal-specific platforms had two advantages (better out-of-box quality on certain agreement types, plus more familiar UI for lawyers) and three significant disadvantages (much higher cost, NZ data residency concerns on two of the three candidates, and lock-in risk if the firm wanted to switch in 18 months).

We ran a one-week comparative test with two candidates on the same 10 anonymised contracts, scored by two senior partners. The horizontal tool with bespoke prompting won on three criteria: output quality on the firm's actual work, cost (roughly one-third the alternative), and flexibility for future workflows. The legal-specific platform was kept on a "revisit in 12 months" list.

Phase three: Workflow build (weeks 3-5)

Three weeks of embedded work. We built a structured prompt library covering the seven most common SME contract types the firm reviews, with each prompt tuned against the firm's actual standard positions on key clauses (taken from their existing playbook and partner interviews). The workflow took an uploaded contract and produced a structured first-pass review: classification, summary of commercial terms, flagged deviations from the firm's standard position, identified unusual clauses, and a draft markup with proposed language.

Critically, every output ended with a verification checklist for the reviewing lawyer. The workflow was explicitly designed as first-pass assistance, not a substitute for the lawyer's judgment. The lawyer reviewed the AI output, accepted or rejected each flagged item, added their own judgment-heavy clauses (anything to do with parties' specific commercial intent), and signed off. The professional responsibility stayed where it belonged: with the human lawyer.

We integrated with their document management system so contracts could be uploaded directly from the client matter file, and the AI-generated draft was saved alongside the human-finalised version with full audit trail. Data residency was confirmed to NZ/Australia for the AI vendor. Client confidentiality was maintained: no client data was used to train the underlying model, and a per-matter prompt isolation pattern prevented cross-contamination.

Phase four: Training and certification (weeks 6-7)

Nine of the firm's lawyers were trained on the workflow: six in the commercial group plus three from other groups who reviewed contracts as part of broader work. Training was structured: a half-day workshop, then four contracts reviewed under shadow supervision, then a certification step where a senior partner reviewed the lawyer's first 10 independent uses for quality and judgment.

The certification step was the firm's idea, and we strongly endorsed it. It meant every lawyer using the workflow had been formally signed off by a partner before they could use it on live client matters. That gave the partners confidence the tool wasn't being used by under-prepared staff, and it gave the firm a defensible audit trail for Law Society professional supervision obligations.

Three months post-certification, measured against the pre-engagement baseline:

Per-contract review time. Average dropped from 42 minutes to 8 minutes, an 81% reduction. The 8 minutes is the lawyer time, not including the workflow's own processing time (typically 60-90 seconds, run in the background while the lawyer reads the contract). Junior associates and senior partners converged on the same average, eliminating the variance the managing partner had originally been worried about.

Error rate. The firm tracked two error metrics pre and post: clauses missed in initial review (caught later by partner signoff or, worse, by the client's other side), and standard-position deviations not flagged. Combined error rate fell 67%, predominantly driven by the AI catching mechanical-but-easy-to-miss items (missing definitions, inconsistent capitalisation of defined terms, cross-references to schedules that didn't match). The lawyer remained responsible for judgment calls, which is where errors of significance had always been concentrated.

Capacity. Senior lawyers' per-week capacity for contract reviews rose 340% based on time-tracking data, holding everything else constant. The firm chose not to use that capacity to take on more contract work at the existing price point. Instead, they used the freed time in two ways: deeper advisory work for existing clients (the higher-value, less-commoditised work the partners had wanted more time for), and a deliberate decision to push back on weekend turnarounds, improving partner work-life balance.

Client experience. Turnaround time on standard SME contracts went from "we'll have it back to you in 2-3 days" to "you'll have it back tomorrow morning, often same-day." The firm hasn't yet decided whether to monetise that explicitly (faster turnaround as a paid tier) or absorb it as a baseline improvement. Some clients have noticed and commented, unprompted. None have complained about AI involvement (the firm discloses AI use in their engagement letters, with the verification framework explained).

Professional supervision. Every AI-assisted review remains a lawyer's review, signed off by the named lawyer, with the AI draft archived as supporting material rather than as the deliverable. The firm has briefed both the Law Society's professional standards team and their PI insurer in advance. Neither raised concerns about the workflow as designed.

What didn't happen. The firm didn't try to extend AI into litigation work, mergers, or anywhere judgment-heavy and unique-per-matter. They didn't reduce headcount. And they didn't change billing structure; standard SME contract reviews remain fee-based and the firm absorbed the time saving rather than passing it through as a price cut.