Choosing Legal AI: Match the Tool to the Work

Write the job before the shortlist

Describe the bottleneck in one sentence. “We need AI” leaves too much open. “We need to turn a document packet into a factual brief that a partner can check against sources” gives the evaluation a useful boundary.

Record the following:

• Inputs: the records, authorities, agreements or instructions the task uses.
• Output: the actual work the reviewer needs, including its format.
• Scope: the jurisdiction, date range, matter or document set.
• Review: who checks it and which errors would matter.
• Reuse: whether another lawyer needs the same context for later work.
• Change: what happens when instructions or source material change.

Choose a task that occurs often enough to evaluate. Include the time needed to organise inputs and check the result. If the problem is locating a fact and its source, a polished draft alone does not demonstrate that the problem is solved.

Match the tool's purpose to the job

Use these categories to organise questions, while recognising that products may cover several of them.

Legal research

For questions about the law, check jurisdiction coverage and whether the authorities are current. Inspect cited sources and how the product identifies later judgments that affect them. Have the reviewer check the reasoning that applies the law.

Contract review and drafting

For agreement work, check playbook support, comparison of versions, proposed wording and the document format the lawyer uses. The Law Society's contract-review discussion recommends evaluating the use case and the client's need before choosing the technology. 1

Discovery

For preservation, review and production, assess the specific discovery controls required by the matter. EDRM's current model also recognises analysis across that work. A product with analysis features may still need to be evaluated for the production workflow you require. 2

Matter work

For questions about what happened, test how the product connects documents, factual assertions, people and issues. Check whether another lawyer can inspect the support and whether earlier work remains understandable after the packet changes.

General assistance

A general assistant may help prepare wording or organise an approved packet. Check its permissions, data terms and capabilities for the specific task before supplying client material.

Firm operations

Practice-management tools may support intake, billing and other operational work. Evaluate those features against the firm's administrative process, including its approval and recordkeeping requirements.

Ask for a demonstration with difficult material

Use a public, synthetic or otherwise expressly approved packet. Include a few complications that resemble the firm's work: a poor scan, an attachment, an inconsistent date, a contrary account and a source that only supports part of a proposition.

Ask each shortlisted supplier to demonstrate the same task. Keep the brief and expected output consistent. Give the reviewer access to the complete packet so omissions can be identified.

The review should answer concrete questions. Did the tool preserve who made a statement? Can the reviewer open the relevant passage? Did it identify missing material? Does the output distinguish a factual account from a legal assessment?

Use the demonstration to choose what deserves a pilot. Broader use needs evidence from the document types and tasks the firm intends to cover.

Book a Grella demo

Inspect context, corrections and change

Ask what carries into the next task. Some products retain conversations, documents, preferences or structured matter information. Harvey's August 2026 announcement, for example, describes agents using matter context and memory. That is a vendor description, not a comparative accuracy result. 3

Have the supplier correct one factual assertion, then show the next piece of work. Ask whether the correction affects shared information, only the current answer or a particular draft. Inspect what remains as originally written and what requires rerunning.

Also test a new source version and a removed permission. The reviewer should understand which record the answer used and which information was unavailable. Avoid treating the presence of memory or citations as a guarantee that this maintenance works as the firm expects.

Treat data approval as a decision of its own

Before supplying client material, review the actual service and contract. Questions should cover permitted data, external processors, training use, retention, deletion, access controls, connected systems and the firm's client obligations.

US lawyers can use ABA Formal Opinion 512 as one reference for confidentiality, competence and supervision. It addresses the US Model Rules; firms must check the requirements that apply to their own practice. 4

Ask for evidence behind the supplier's answers. Keep a record of approved uses and outstanding questions. Do not infer approval for privileged material from a product label, a paid subscription or a general security statement.

Compare the whole cost of the work

Record the commercial terms and the effort required to use the product. Include onboarding, preparation of inputs, review, correction, export and administration. Check which features or data sources need a different plan and what happens when the firm leaves.

Have the reviewer assess the usable result, not only the speed of generation. A faster first draft may still be useful, but its value depends on the effort required to make it reliable for the intended task.

Use the one-matter pilot guide for a run log and acceptance criteria. This article helps choose what to test; that guide covers how to conduct and measure the test.

Make a short, written decision

For each serious candidate, keep a decision note with the task, demonstrated result, review findings, data-use position, commercial terms and unresolved questions. End with one decision: proceed to a bounded pilot, resolve a named question or stop evaluating that candidate.

For document-heavy matter work, assess Grella's matter workspace and source-linked answers alongside the firm's existing research, discovery and operational tools.

If the bottleneck sits elsewhere, start with the tools built for that job. The shortlist should follow the work your firm needs to deliver.