Can Agent Skills Handle AI for Lawyers? What to Expect in 2026

The Adoption Gap: Lawyers Are Using Agent Skills Faster Than Firms Can Govern Them

Generative AI has more than doubled its footprint in legal work in a single year. According to the 2026 Legal Industry Report published by the ABA's Law Practice Magazine, 69% of legal professionals now personally use GenAI tools for work, up from 31% the year before. Individual lawyers have made their decision — the question is whether their firms can keep up.

The uncomfortable answer, so far, is no. The same report frames the problem bluntly: usage is outpacing structure. While lawyers draft correspondence (58% report doing so with AI), run general research (58%), and summarize documents (47%), the institutional scaffolding around them — policies, training, verification standards — has not kept pace.

The training gap is the starkest example. A full 54% of respondents say their firm has provided no training on responsible GenAI use — and has no plans to do so, per the same ABA-published survey. That means most lawyers are learning AI on their own, with no shared standards for what "verified" output looks like. BCG Attorney Search calls this gap "one of the defining management challenges of this moment."

The hesitation, notably, is not about money. The top barriers lawyers cite are:

• Data security (46%)
• Ethical issues (42%)
• Privilege concerns (39%)
• Lack of trust in results (39%)

Every one of those concerns maps directly to governance, not cost. And agent skills — reusable instruction packages that give AI agents like Claude, Codex, and Cursor their legal capabilities — sharpen the problem. Anthropic itself warns that Skills can contain executable code or third-party packages, and published research has found roughly 26% of skills carry at least one vulnerability. A lawyer installing an unvetted skill isn't just adopting a tool; they may be introducing unreviewed code into a privilege-sensitive environment.

This is why 2026's defining question for legal AI is not "should we use agent skills?" — that ship has sailed — but "which skills are safe and useful, and how do we prove it?" Firms need a way to evaluate skills before they touch a matter, the same way they evaluate a new hire. SkillzBench is being built for exactly this: independently crawling, security-scanning, and sandbox-benchmarking agent skills so firms can see what each one actually does — and whether it measurably helps — before deployment.

The lawyers have moved. The governance hasn't. Closing that gap is the work of 2026.

What Agent Skills Can Actually Do for Legal Work in 2026

The most important shift in legal AI isn't a better model — it's a change in the unit of work. As Harvey's team puts it, the unit of legal AI is moving from the prompt to the task: agents now receive a goal, plan their steps, gather sources, and return a review-ready deliverable.

That distinction matters. A chatbot answers a question; an agent researches the answer. As one conference panelist framed it, "LLMs can write an essay, but agentic AI can research an essay." The lawyer's role moves up the stack to scope, strategy, and judgment. "The agent does the work," Harvey writes. "The lawyer decides what to keep."

Where agents create leverage in 2026

The data shows adoption concentrated in work that is high-volume and structurally consistent across matters. According to industry research, the top use cases are:

• Document review — 77% of professionals using AI for legal work
• Legal research and summarization — 74% each
• Correspondence drafting — 54%
• Contract drafting — 49%

The time savings are equally concrete. Thomson Reuters estimates 190 work-hours saved per lawyer per year through AI-enabled efficiency. Harvey offers a vivid benchmark: a first-year associate who once spent 40 hours marking up a services agreement might now spend 5 hours reviewing an agent's markup. At one firm reported by the Thomson Reuters Institute, a complaint response system cut associate time from 16 hours to 3–4 minutes.

Augment, never replace

Every credible source agrees on the boundary. Thomson Reuters states plainly that agentic AI "can only augment legal professionals—not replace them," and Harvey calls lawyer review of all agent output "non-negotiable." The ABA's industry reporting echoes this: AI may accelerate work, but it doesn't replace professional responsibility.

For firms evaluating agent skills specifically, the same discipline applies. Anthropic warns that skills can contain executable code or third-party packages, and published research found roughly 26% of skills carry at least one vulnerability — which is why SkillzBench is building independent security scanning, version-pinned validation, and benchmarking into its free skills directory, so legal teams can verify what a skill actually does before pointing it at a matter. The agents handle the hours; the judgment stays with the lawyer.

The Risks Nobody Can Afford to Ignore: Hallucinations, Security, and Unverified Skills

If a lawyer representing Anthropic itself can file a court document with a botched citation, no firm is immune. That May 2025 incident — where counsel acknowledged that Claude had produced a real article with an inaccurate title and authors — is the clearest warning yet that hallucination remains a top-tier risk for legal AI adoption.

The stakes are rising precisely because adoption is accelerating. Roughly 69% of legal professionals now personally use generative AI, more than double the prior year, yet 54% of firms have provided no training on responsible AI use — and have no plans to. That gap between enthusiasm and governance is where mistakes happen.

Agent skills add a second, less obvious layer of danger. Anthropic itself warns that skills can contain executable code or third-party packages, and advises lawyers to install skills only from trusted sources after reviewing their contents. That warning matters because published research has found that roughly 26% of agent skills carry at least one vulnerability — meaning nearly one in four packages could introduce security or privacy exposure into a firm's environment.

The core discipline every source converges on is treating AI output as draft intelligence, not finished truth. Fluent but inaccurate output is the failure mode Clio highlights most sharply: a clause that misstates the legal effect still reads like solid drafting. Harvey's governance framework goes further, calling citation auditability non-negotiable — every claim must trace to a verifiable source, and output without citations should be re-read from scratch.

Before installing any skill, firms should ask:

• Does this skill contain executable code, and who wrote it?
• What files, matter data, or systems can the agent reach once enabled?
• Has the skill been security-scanned and version-pinned to a known commit?
• Does it come from a source the firm would defend in front of a client or a court?

The uncomfortable truth is that no vendor closes this gap for you. Harvey's 500+ agent library and Anthropic's bundled legal plugins are vetted by their creators, not by your firm — and "an agent with undefined access boundaries is a confidentiality incident waiting to happen," as Harvey's own team puts it.

This is the problem SkillzBench is being built to address: an independent directory that security-scans, version-pins, and benchmarks agent skills before you ever install one, with a free MCP connector so you control exactly what each agent sees. Until verification infrastructure like that becomes standard practice, the burden stays where the research puts it — on the lawyer, before anything reaches a filing, a client, or a court.

A Verification-First Playbook for Deploying Legal Agent Skills

The gap between AI enthusiasm and AI discipline is where legal practices get into trouble. With 69% of legal professionals now personally using GenAI — more than double the prior year — but 54% of firms providing no training and having no plans to do so, the structure is lagging the usage. A verification-first playbook closes that gap.

Start low-risk, then scale. Thomson Reuters recommends beginning with low-risk applications like deposition summaries and contract analysis that operate independently of core workflows, progressing to higher-stakes work only after governance muscle develops. Harvey's analysis shows agents create the most leverage where tasks are high-volume and structurally consistent — transactional work, compliance, and high-volume litigation corners.

Before any skill touches a matter, establish written AI governance policies and mandatory training. ABA Formal Opinion 512 confirms that competence, confidentiality, client communication, supervision, and reasonable billing remain central duties when lawyers use generative AI — and governance "cannot be relegated to IT or innovation alone."

Verification must also extend to the skills themselves. Anthropic warns that Skills can contain executable code or third-party packages, and published research found roughly 26% of skills carry at least one vulnerability. Before installation, check:

• Provenance — pin each skill to a specific commit so you know exactly what you're running
• Security — scan for vulnerabilities and executable code before enabling anything
• Licensing — confirm the license permits your firm's commercial use
• Access boundaries — as Harvey puts it, "an agent with undefined access boundaries is a confidentiality incident waiting to happen"

Finally, benchmark before trusting. Run the same task with and without the skill on identical models and measure the Lift in percentage points — a skill that sounds helpful in a README may add nothing measurable. SkillzBench is building exactly this verification layer for legal teams: a free directory where every skill is version-pinned, security-scanned, license-checked, and benchmarked in a sandbox, plus a free MCP connector so each agent only sees the vetted skills you toggle on. The platform is currently pre-launch, with a live waitlist for firms wanting early access.

Treat every output as "draft intelligence, not finished truth" — verify citations from scratch, keep a lawyer in the loop on every deliverable, and scale only as fast as your verification discipline grows.

What 2026 Holds: Business Models, Training, and the Skills That Still Matter

The billable hour, it turns out, is more resilient than the tech press predicted. When Harvard's Center on the Legal Profession interviewed ten AmLaw100 firms, every one of them expected the billable hour model — still an estimated 80%+ of fee arrangements — to survive AI's productivity gains. Firms plan to capture that value through higher rates rather than passing savings to clients, and none anticipate attorney headcount cuts. One interviewee put it plainly: "Even with our AI initiatives, we just brought in the largest associate class in the history of the firm."

Scale, not software, becomes the real competitive moat. The same Harvard research found that the magnitude of AI investment will be difficult for mid-sized firms, creating a genuine threat to second-tier players. Agentic AI demands clean data and difficult engineering — one panelist described it as "not necessarily for the weak." Deep pockets win the build-out; everyone else has to be smarter about what they install.

The harder 2026 question is associate development. Agents shift junior work toward review, judgment, and source verification — a first-year who once spent 40 hours marking up a services agreement might now spend 5 hours reviewing an agent's markup, according to Harvey. Leading firms are treating this as a curriculum problem, pairing deployment with structured training that includes "intentional friction":

• Hand-drafting key documents before comparing them to agent output
• Structured review programs that teach verification, not just approval
• Citation auditing — every claim traced to a verifiable source

The stakes are real. As Harvey warns, firms that deploy agents without rethinking training will produce "associates who can review competently but cannot draft from a blank page when the matter calls for it." And the governance gap compounds the risk: 54% of firms have provided no training on responsible GenAI use and have no plans to start.

That brings the 2026 picture into focus. Skills and agents are easy to install and hard to trust — published research has found roughly 26% of skills carry at least one vulnerability, and Anthropic itself advises lawyers to review skill contents before enabling them. SkillzBench is building exactly this kind of verification layer: security scanning, version pinning, and benchmarking that tests every skill with and without it, so firms can see whether a skill actually helps before it touches a matter. The dividing line in 2026 isn't between firms that adopt AI and firms that don't. It's between firms with skills that are proven and firms with skills that are merely installed — and the second group is where AI gets expensive.