The State of Legal AI in 2026: What Matters for Law Firms

Adoption is growing. Measure the work in your own firm

Thomson Reuters' 2026 AI in Professional Services Report says 40% of respondents reported organisational use of generative AI, compared with 22% in the previous report. Only 18% said their organisation tracked its return on investment. These figures cover several professions, including legal, tax, accounting and government. They are not percentages of all lawyers or evidence of savings in a particular firm. 1

Its separate 2026 law-firm analysis uses 736 responses from law-firm professionals across 46 countries, gathered in March and April. That scope matters when using the findings to inform a local decision. 2

A firm can use this research to understand the market, then measure its own work. If a tool prepares a factual summary quickly but takes longer to correct than the old process, generation time alone is a poor measure of value. Count preparation, source checking, correction and final review. Record what the tool missed as well as what it found.

Legal AI now covers several different jobs

Treat the task as the starting point. A research question about the law needs current authorities in the right jurisdiction. A question about what happened needs the relevant matter records. A contract review needs the agreement, the firm's position and the surrounding deal context.

The practical categories include:

• Legal research: finding and analysing cases, legislation and other legal authorities.
• Contract work: comparing provisions, reviewing against a playbook and preparing proposed wording.
• Discovery: managing collected information, document review and production.
• Matter work: connecting facts, people, issues, chronologies and drafts to the record.
• Firm operations: supporting intake, time recording, billing and other administrative work.

Products can cover several categories. Use these labels to organise questions about the work and the sources it needs. Our tool-selection article explains how to match the next purchase to the work.

Persistent context is becoming a product focus

In August, Harvey announced Harvey II, describing agents that use matter and project context together with remembered working preferences. This is a vendor description of its product, not independent evidence of better legal outcomes. 3

For a firm evaluating any product, the important detail is what that context contains. A saved conversation, a shared factual record and a lawyer-approved conclusion are different things. Ask where each is stored, who can change it and how the next task uses it.

Consider a hypothetical team preparing a witness outline. It has an interview note, a disputed email and a later correction. The team needs to preserve who said what, which document supports each statement and which questions remain open. A fluent account that blends the three records can make review harder.

Test what happens when the correction arrives. Does the product show affected references? Must someone rerun the analysis? Does an earlier draft keep its wording? Have the supplier demonstrate the exact behaviour instead of assuming that retained context means every output stays current.

See the matter workspace

Agents make permissions and review points more important

An AI agent can use tools to carry out several steps towards a task. The permissions vary by product: one may only search and prepare a draft; another may edit records or interact with connected systems.

Ask what the agent can read, what it can change and where it must stop for approval. Test the stopping point as carefully as the finished document. A draft saved for review is a different outcome from a message sent to a client.

Our agent-delegation article sets out a task brief, permission boundaries and review checkpoints.

Discovery and case analysis remain connected

EDRM's September 2026 model treats analysis as work across the discovery lifecycle. That is a useful correction to the idea that a team finishes discovery and only then begins understanding the case. 4

Products are also building connections between these jobs. Harvey and Everlaw announced a partnership in September to bring discovery evidence into Harvey. The announcement said availability was expected in fall 2026; it does not establish that the integration is available to every customer today. 5

For a firm, the question is where the work belongs and how references survive movement between systems. Check document identifiers, versions, access restrictions and links back to the reviewed source. Our eDiscovery and case-analysis article covers that relationship in more detail.

Review still needs a named owner

The ABA's Formal Opinion 512 discusses verification, confidentiality and supervision under the US Model Rules. The review needed depends on the tool and task. Firms elsewhere must check their own professional rules, court directions and client requirements. 6

In day-to-day work, assign someone to check material factual assertions, important omissions and the basis of legal conclusions. A citation is a route to evidence, not a finding that the whole sentence is correct. Open the source and read enough surrounding material to test the wording.

Check data use before supplying client material. Review the actual service terms, access controls, processors and retention arrangements. A product category or security badge does not settle whether a particular matter is suitable for the service.

What to do with this picture of the market

Choose one recurring task with a clear output and a reviewer who knows the work. Use a public, synthetic or otherwise expressly approved packet for the first test. Write down the result you need, the errors that would matter and the time spent on the whole process.

Then test a change: add a missing document, correct a source or restrict an important file. Ask the reviewer whether the result is still understandable and whether the team can identify what needs another look. The one-matter pilot guide provides a practical measurement method.

Grella's focus is working from a connected matter record: documents, facts, people, issues and the work built from them. The decision to use it should follow the task, the review needs and the firm's approval of the proposed data use.