Legal Services AI 2026: faster work, defensible work

31%

Personal use

Share reporting personal generative-AI use at work in 2024 responses from more than 2,800 legal professionals, published in the MyCase 2025 Legal Industry Report on the ABA site.

21%

Firm use

The same report's 2024 measure of firm use. It is not the same question as personal use, and the figures should not be subtracted as a formal shadow-use rate.

79%

Any AI use

Clio's provider-led 2025 Legal Trends Report says this share of surveyed legal professionals used AI in their firms. A different sample and question from the MyCase survey.

Where the work changes

Start at intake, not at the model. A client question creates a matter, a conflict check, documents, factual assertions, research, advice, and often a filing. AI may accelerate document sorting, chronology, issue spotting, and first drafts. It does not erase the lawyer's obligation to know which record and rule support an answer. The practical unit is a reviewed work product: a memo with links to authorities, a contract comparison tied to clauses, or an intake summary checked against source documents.

Example: A commercial team compares 60 supplier agreements. The model extracts renewal clauses into a table; a reviewer opens every cited clause in the original document before the summary goes to deal counsel.

Choose the first matter carefully

The safest first job is not necessarily the smallest document. Choose a frequent task with an established reviewer, a known source set, and an error that can be corrected before a client or court acts. Research summaries can be useful when the tool retrieves from a controlled library and the lawyer follows each citation. Client emails can be useful when the factual record and promised action are checked. Avoid starting with autonomous advice, settlement terms, filing, or privilege decisions.

Example: A litigation team pilots chronology drafting on closed matters with known outcomes. It scores missing events, wrong dates, unsupported assertions, and time to a partner-approved timeline before using live matters.

Confidentiality is a workflow decision

A product's security page is not a matter-specific data policy. Identify what enters prompts, what is retained, who may retrieve it, whether vendor personnel or subprocessors can access it, and how a client restriction is enforced. Separate public research from privileged material. Give staff an approved alternative; a policy that only says 'do not use AI' may push work into unmanaged accounts.

Example: A practice group allows public-law research in an approved assistant but requires a separate workspace and explicit matter approval for client documents. Templates remind users to remove unnecessary identifiers.

The source-verification chain

For each legal proposition, require a source that exists, matches the jurisdiction and date, remains good law, and actually supports the sentence. For factual propositions, point to the record, not to a model's paraphrase. The reviewer should be able to move from output to authority in one or two clicks. Test the chain with deliberately tricky cases: superseded statutes, negative treatment, similarly named parties, and quoted text with a missing qualifier.

Example: Before a brief is filed, a paralegal checks citations against the court record and primary authority; the signing lawyer reviews the proposition and current validity.

Client economics and pricing

Under an hourly engagement, bill for actual work and review the ABA's fee discussion before passing through AI costs. Under fixed fees, faster delivery may improve margin or allow a better client price, but only if quality and scope stay stable. Do not count every saved drafting minute as cash. First separate released capacity from lower external cost or truly avoided hiring; then include licensing, configuration, review, correction, and client communication.

Example: A firm handles 200 routine agreement reviews yearly. A trial reduces average preparation by 25 minutes but adds 10 minutes of verification. The net 15 minutes is capacity, not automatic cash savings; realization depends on whether the team uses that time productively.

A 90-day operating trial

Days 1-20: choose one matter type, document the baseline, set data rules and exclusions. Days 21-45: run on past matters and compare with known work, with a source-by-source error log. Days 46-70: allow reviewed live use for a small cohort; sample outputs weekly and record client-facing corrections. Days 71-90: decide by matter type, not by aggregate enthusiasm. Keep a rollback path and tell staff where work goes when the tool is unavailable.

Example: The pilot committee includes a practice owner, knowledge lawyer, information-security lead, billing owner, and one skeptical reviewer who sees the original source set.