This week's useful signals concern how AI fits into legal work: a specific litigation workflow, a lower processing cost, a local deployment option, and a corporate client's changing production habits. A Florida appellate decision supplies a concrete reminder that the cost of failed verification can reach both counsel and client.
These developments deserve different responses. Some support a bounded experiment, some belong in client conversations, and others sharpen an existing review obligation; none establishes that buying more AI will improve a firm's results.
01Workflow design
Cooley starts with a redaction task and keeps the decision with the lawyer
Cooley announced that it is developing a Google Cloud Gemini Enterprise agent to identify potentially confidential information in court filings and recommend redactions. Lawyers will make the final determinations before filing. The agent remains in development, and the announcement provides no measured accuracy, time savings, or cost results.
Leadership perspective. The useful feature is the choice of work: a recurring, consequential task with a defined point at which professional judgment takes over. A mid-sized firm can borrow that design discipline without reproducing Cooley's development arrangement. For a comparable pilot, measure missed sensitive material, unnecessary redactions, and total attorney review time against the current process. A system that makes a first pass faster but creates more checking may consume capacity rather than release it.
Redaction also exposes the weakness of a generic promise that a lawyer will review the output. The review must identify omissions as well as assess proposed deletions, so the workflow needs a way to inspect what the agent left untouched. Treat this as a useful implementation example, with effectiveness still to be demonstrated.
02AI economics
Cheaper models change the cost of a task, not the value of the finished work
Anthropic released Claude Haiku 5.5 with published API prices of $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Longer prompts carry higher rates. Anthropic reports substantial cost and performance improvements, but those claims do not establish results in a law firm's work.
Leadership perspective. Lower processing prices make narrow, repetitive uses more plausible: classifying intake material, extracting specified document fields, or preparing a summary for review. Firms should evaluate those tasks separately from complex analysis rather than assume every step requires the same premium model. Conversely, a lower API price does not automatically reduce a legal-software subscription, and tokens are the units of text a model processes, not a measure of completed legal work.
The economic test remains the cost of an acceptable result, including setup, checking, corrections, and maintenance. Ask a prospective supplier which model handles each task, how changes are validated, and whether lower underlying costs reach the firm. For most firms, attorney review time is the more consequential variable to measure before expanding a pilot.
03Controlled environments
Local meeting assistance makes deployment location a practical choice
Google launched AI Edge Foresight for Mac, an experimental meeting companion powered by local AI models. Google says it can enhance shorthand notes and retrieve information from transcripts and private files, operating offline with sensitive data remaining on the device. Those are developer claims, not an independently verified assessment of security or legal-work accuracy.
Leadership perspective. This offers a concrete alternative to assuming that useful meeting assistance requires sending every conversation to a cloud service. For firms with suitable Macs, it is a reason to investigate a local processing route, not a reason to authorize unrestricted recording. Where computation occurs and how the resulting information is managed are separate decisions.
A local transcript can still become an unmanaged copy of client information. Before a limited evaluation, examine recording permissions, device access, retention, deletion, backups, and whether matter information stays separated. Assess the accuracy of notes against the actual conversation, because a plausible summary can omit the qualification that changes its meaning. An experimental personal app may prove useful, but firm deployment also requires a workable administration and support path.
04Client relationships
A corporate legal team's AI first pass may change the work sent outside
Harvey announced deployment across TC Energy's legal team, supporting work in Canada, the United States, and Mexico. According to the supplier's account, the pilot encouraged lawyers to use Harvey for an initial pass on their work. The announcement includes the general counsel's perspective, but no independent outcome measurements, and it does not specify when the pilot or deployment occurred.
Leadership perspective. The new evidence is the published account of a buyer changing its own production process. That does not establish a reduction in outside-counsel spending, or a general change across corporate clients. It does suggest that a client's first request may increasingly arrive with an AI-assisted draft, preliminary analysis, or a narrower unresolved question.
A firm should understand what the client has already done and what it expects counsel to add. Reviewing a client-generated first pass can require fresh verification, even when the draft looks complete; pricing and staffing should reflect that work. At the next suitable client discussion, ask how their internal process is changing and which decisions still need outside judgment. The relevant competitive distinction is the firm's contribution after the first pass.
05Professional obligations
A Florida fee award puts client-supplied research inside the verification obligation
In Kid International, LLC v. City National Bank of Florida, No. 4D2025-2599, the court awarded the opposing parties reasonable appellate attorney's fees after an AI-assisted brief contained fabricated authorities, quotations, and mischaracterized holdings. Counsel had also accepted client-supplied research and admitted inadequate verification. Plaintiffs and counsel are jointly and severally liable; the amount will be determined on remand, and the opinion remains subject to timely rehearing.
Leadership perspective. The decision extends earlier sanctions reasoning to an opponent's fee motion under Florida Rule of Appellate Procedure 9.410(b). Its operational lesson travels further than its jurisdiction: checking whether a case exists is only the beginning. The quotation, cited proposition, and actual holding also need examination, including when the material originates with a client.
The attempted amended brief did not resolve every citation problem. For leadership, that makes verification a production step with an accountable owner, not a repair procedure after detection. A defined source-checking record before filing provides a more inspectable control than an instruction to be careful. The ruling concerns defective submitted authority, and does not establish that all AI-assisted legal research is improper.
The work through completion
Measure the process after the first output
The thread across these items is the difference between producing an intermediate output and delivering usable legal work. A suggested redaction, inexpensive extraction, meeting note, or client draft earns its place only when the full process improves. Leadership can make that distinction concrete by assigning a reviewer, defining acceptance criteria, and measuring the work through completion rather than stopping at the AI's response.