It is the Wednesday after a partners’ retreat and three people want three different AI projects funded from one budget line. One wants a client-facing chatbot. One wants a drafting assistant. One wants help with the data rooms that keep associates at their desks past midnight before every signing. The firm picks one, spends S$30,000, and six months later the tool sits unused because nobody trusts it with the judgment calls it was built for.
That failure was avoidable before a line of the build was written. Singapore’s Ministry of Law published a scoring method for exactly this decision, and first-pass document review automation clears that test in a way a drafting tool or a client-facing chatbot usually does not. This piece works through why, using the guide’s own scoring, not a vendor’s opinion of what looks impressive in a demo.
First-pass document review automation is the right first AI system for a Singapore law firm because it is the one candidate that clears all four axes MinLaw’s guide uses to score a use case: the confidentiality of the client data, the risk level and oversight the task demands, the cost-benefit case, and the firm’s readiness for the change. A drafting tool or a client-facing chatbot usually fails at least one of the four on a first project, verified against MinLaw’s Guide for Using Generative AI in the Legal Sector, published 6 March 2026 on mlaw.gov.sg.
The four axes a first AI project has to clear
The guide’s second step, diagnosing and analysing needs, does not ask a firm to pick the AI project that sounds most impressive. It asks the firm to score every candidate workflow on four things before spending a dollar: how confidential the data is, how much risk the task carries and what oversight it needs, whether the cost-benefit case holds up, and whether the firm is actually ready for the change. A use case that scores well on three axes and fails the fourth is not a good first project. It is a second or third one, once the firm has a working system and a trained team behind it.
Run a drafting tool, a client-facing chatbot, and first-pass document review through that test side by side and the pattern is not close.
| Axis | Contract drafting | Client-facing chatbot | First-pass document review |
|---|---|---|---|
| Confidentiality | Client matter, but the judgment calls sit with a partner reading the whole draft anyway | Public-facing by design, the hardest axis to score well on a first project | Client matter, contained to the firm’s own tenant, never exposed outside the review team |
| Risk and oversight | High. “Market” terms vary by deal, so the check is a full re-read, not a spot check | High and hard to bound. A chatbot answers questions nobody scripted | High but familiar. A lawyer already checks every clause an associate flags |
| Cost-benefit | Real, but the senior review step does not shrink much in year one | Unclear. Few firms can size the deflected-enquiry value before launch | Concrete. The hours are already logged per deal and per document |
| Readiness | Depends on a firm-wide clause playbook existing and being current | Needs new content, escalation rules, and a governance layer before it works at all | The checklist already exists. It is what associates use today |
Document review is not risk-free. It carries the same duty of care as any client work. What sets it apart is that the firm already runs the exact oversight step the guide asks for, a lawyer checking every finding before it goes to a client, so building the AI system does not require inventing a new control. It slots underneath one that already exists.
axes MinLaw's guide scores every AI use case against before a firm builds anything: confidentiality, risk and oversight, cost-benefit, and readiness
MinLaw, Guide for Using Generative AI in the Legal Sector, 6 March 2026, Section 4.2
Where document review wins on each axis
Confidentiality is the axis firms worry about first, and it is the one document review handles most cleanly. The data room or the contract file is already confidential client information under Rule 6 of the Legal Profession (Professional Conduct) Rules 2015. Nothing changes about who sees it: the system runs on the firm’s own tenant, the provider is contractually barred from training on the documents, and the same associates who read the files today are the only people who see the AI’s output.
Risk and oversight look high on paper. A data room in an active deal has real consequences if something is missed. But the oversight model is not new. A senior associate or partner already checks every clause a junior associate flags before it reaches the issues list. Wiring an AI system underneath that same review step, so a lawyer signs off on every AI-flagged clause and every stated gap, adds no control the firm has never run before. A client-facing chatbot has no equivalent check waiting for it. Someone has to build the escalation logic, the content boundaries, and the review process from nothing, on a channel that talks to the public.
The cost-benefit case is concrete because the hours already have a number attached. A firm can pull how many associate-hours went into data-room triage last quarter from timesheets, today, without hiring anyone. Multiply the hours by the loaded rate and the arithmetic is not a projection. A drafting tool’s benefit is real but softer to size in year one, because the senior lawyer still reads the whole document. A chatbot’s benefit depends on enquiry volumes and deflection rates nobody has measured yet.
Readiness is the axis that decides whether a pilot survives contact with a real file. Document review already has the thing an AI system needs to learn from: a checklist. Every corporate team has an informal or formal list of clauses it checks on every file, change of control, assignment, exclusivity, termination, indemnity caps, and a sense of what a departure from standard form looks like. That checklist is the specification. Drafting judgment and client-facing answers do not have an equivalent artifact sitting in a shared drive already.
Running the scoring exercise on your own workflows
A partner group can do a rough version of this scoring exercise in an afternoon, before calling anyone.
List every candidate workflow: document review, contract drafting, legal research, client intake, billing narratives, whatever comes up in the retreat conversation. For each one, rate the four axes high, medium, or low: how confidential is the data, how much oversight does the risk demand and does that oversight already exist, how concretely can the benefit be sized from data the firm already has, and how ready is the team, meaning does a checklist, template, or standard already exist to encode. The workflow with the fewest low ratings, not the most exciting demo, is the one to build first.
This is not a shortcut around the real diagnostic. It is the same test MinLaw’s guide sets out at Section 4.2, run informally instead of formally. It will usually point the same direction a formal scoring exercise does, toward the workflow with an existing checklist and an existing review step, which for most Singapore corporate and disputes teams is first-pass document review.
How Kept scores this in the Diagnostic
Kept’s delivery follows the MinLaw five-step procedure end to end, and the scoring exercise above is Step 2 done properly: staff interviews, a walk of the real workflows, a use-case register scored on all four axes, and a prioritised roadmap that names what gets built first and what waits. If no use case clears a conservative payback bar, the Diagnostic says so and stops, rather than building something because the budget line exists.
For most Singapore law firms that roadmap points at first-pass document review, which is why our document review workflow guide covers what a properly built system does to the hours in a data room, with representative arithmetic. The Diagnostic itself runs two to three weeks at S$5,000 to S$15,000 by firm size. The Build & Train phase that follows is six to twelve weeks at S$25,000 to S$45,000, fixed scope, and can qualify for EDG support of up to 50% for a qualifying SME, subject to EnterpriseSG approval, verified July 2026 on enterprisesg.gov.sg.
The reason the sequencing matters beyond the first project: a firm whose first AI system works and is trusted has staff who now understand how these systems get built, checked, and maintained. The second project, whatever the roadmap ranks next, gets easier and cheaper because the firm is no longer starting from zero. The full argument for why that ownership matters more than the tool itself is in why AI projects fail, and the complete five-step framework this scoring exercise sits inside is in what Singapore’s GenAI guide asks of law firms.
Pick the workflow that clears all four axes, not the one that sounds best in a partners’ meeting. For most firms that is the data room, not the chatbot.
Common questions
Why should a Singapore law firm automate document review before drafting or a client-facing tool?
Because it is the one candidate that clears all four axes MinLaw's guide uses to score an AI use case: the data stays confidential inside the firm's own tenant, the risk is contained by an oversight step the firm already runs, the cost-benefit case is concrete because the hours are already tracked per deal, and the firm has years of institutional readiness in the checklist itself. A drafting tool or a client-facing chatbot usually fails at least one of the four on a first project.
What is the four-axis scoring framework in the MinLaw guide?
Confidentiality of the data involved, the risk level and human oversight the task demands, the cost-benefit case, and the firm's readiness for the change. MinLaw's Guide for Using Generative AI in the Legal Sector (6 March 2026) sets these out at Section 4.2 as the test for scoring any AI use case before it reaches a build.
Does starting with document review mean the firm never automates anything else?
No. It is the first project, not the only one. A scored use-case register and roadmap, the actual deliverable from this exercise, ranks every other candidate, drafting, research, client intake, so the firm knows what comes second and third once the first system is running and trusted.
What does a first AI project like this cost, and does the EDG grant apply?
A diagnostic that scores the firm's use cases runs S$5,000 to S$15,000 by firm size, over two to three weeks. The Build & Train phase that follows is S$25,000 to S$45,000, fixed scope, and can qualify for EDG support of up to 50%, subject to EnterpriseSG approval, verified July 2026.
