How to Use AI for Law Firms: A Practical Walkthrough

Why Law Firms Are Racing to Adopt AI (and Where the Numbers Disagree)

Ask ten sources how many law firms actually use AI, and you'll get wildly different answers. One vendor-published report claims adoption leapt from 19% in 2023 to 79% in 2024, while the ABA's survey of 2,800+ legal professionals found firm-level generative AI use actually fell from 24% to 21% over the same period. Both can't be right — and the gap matters if you're deciding how fast to move.

The ABA data carries more weight, but it hints at something important: individual lawyers are adopting AI faster than their firms are. Personal AI use at work rose from 27% to 31%, even as firm-level adoption dipped. Lawyers are quietly experimenting on their own, often with general chatbots that weren't built for legal work.

Where firms do adopt, the use cases are remarkably consistent:

• Document review — used by 77% of AI-using legal professionals
• Legal research and document summary — 74% each
• Brief and memo drafting — 59%
• Contract drafting — 58%

Firm size also shapes the picture. Firms with 51 or more lawyers reported 39% generative AI adoption, roughly double the ~20% rate at firms with 50 or fewer lawyers. AmLaw100 firms interviewed by Harvard Law School's Center on the Legal Profession described AI investments at a scale mid-sized firms can't easily match — one partner shrugged off a $10 million spend as "not really that much money."

The stakes explain the urgency. One AmLaw100 firm's AI-powered complaint response system cut associate time from 16 hours to 3–4 minutes — a productivity gain exceeding 100x. Partners described trying to invert the classic 80/20 problem, where lawyers spend 80% of their time collecting information and only 20% on strategic analysis.

But the data also counsels caution. Independent benchmarking shows AI beats lawyers at question answering, summarization, and extraction, yet loses on redlining and EDGAR research. And with bar authorities from the ABA to state bars requiring human verification of AI output, firms that adopt blindly face ethical exposure, not just efficiency.

The realistic takeaway: AI adoption is real, accelerating unevenly, and concentrated in document-heavy work — but "adopting AI" means very different things depending on who's counting. Firms that verify claims against independent benchmarks, rather than vendor marketing, will make better tool decisions. That's the same principle behind SkillzBench's approach to agent skills: benchmark every capability in a sandbox, publish the lift, and let evidence — not hype — drive adoption.

What AI Actually Does in Document Review and Legal Research

The most useful mental model for legal AI is a pipeline, not a magic answer machine. A document enters, gets parsed into individual clauses, each clause gets classified by type, operative terms like cap amounts and notice periods get extracted, and deviations from your standard get surfaced as flags — a prompt to look, not a verdict. As one practical walkthrough puts it, "a flag is a prompt to look, and the lawyer who signs the document decides what it means."

This is why document review (77%) and legal research (74%) top the list of AI use cases among legal professionals, per Thomson Reuters data. But the pipeline only works end-to-end with legal-specific tools. A general chatbot can summarize a contract and raise plausible issues, but it "cannot tell you whether a clause departs from the position your team accepts, because it has no access to that position" — no playbook, no firm precedent, no matter context.

Independent benchmarks show exactly where the machine wins and loses. The Vals Legal AI Report, cited in Spellbook's analysis, scored AI against a lawyer control group:

• AI wins on document question answering (94.8% vs. 70.1%), summarization (77.2% vs. 50.3%), data extraction (75.1% vs. 71.1%), and transcript analysis (77.8% vs. 53.7%)
• Lawyers win on redlining (79.7% vs. AI scores of 65.0% and 53.6%) and EDGAR research (70.1% vs. 55.2%)
• Chronology generation ended in a dead tie: 80.2% each

The pattern is clear: AI excels at reading and organizing, humans still own judgment-heavy negotiation and source verification. Match the tool to the job accordingly. For contract review and redlining against playbooks, look at Spellbook or Harvey; for eDiscovery culling and privilege flagging, Thomson Reuters CoCounsel; for research, Lexis+ AI, Westlaw Edge, vLex (covering case law from over 17 countries), or Lex Machina for litigation analytics. A platform tuned for eDiscovery won't propose a redline, and one tuned for contracts won't cull a production set.

The same "match and verify" logic applies to the AI agents increasingly running these workflows. SkillzBench is building a free directory that benchmarks agent skills with and without each skill installed, publishing the measured Lift — the same task-defined, control-group methodology legal pilots should demand. If you're wiring AI into document workflows, browse and test free skills there before committing one to production.

The Ethics and Confidentiality Layer You Cannot Skip

Every AI deployment in a law firm runs on top of an ethics framework that predates the technology. Get this layer wrong, and the productivity gains from document review and research become malpractice exposure.

The anchor is ABA Formal Opinion 512, issued July 29, 2024. As summarized by Spellbook, it requires informed client consent before entering client confidences into a generative AI tool — and boilerplate consent language in an engagement letter does not count. That means a real conversation, documented, before any client document touches a third-party system.

State bars have layered on their own requirements. Justia's 50-state survey finds Arizona stating that generative AI "does not alter or lessen the ethical obligations of attorneys," while California requires lawyers to critically review, validate, and correct both AI input and output. Florida's Opinion 24-1, released January 2024, requires informing clients of the actual cost of AI use. Courts are moving too: proposed Tenth Circuit Rule 46.5 revisions, published September 18, 2026, would require human verification of cited authorities, quotations, and analysis before filing — with the duty attaching to the signing lawyer.

Before any document goes into a tool, ask three pre-upload confidentiality questions:

• Does the vendor retain the document after processing?
• Is the content used to train the vendor's models?
• Which subprocessors receive the data?

If the answers are unclear, the default is simple: never input client-sensitive data into free public AI tools. The convenience is not worth the privilege risk.

The duty of candor under Rule 3.3 closes the loop. Submitting AI-generated research without verification can violate it, as Clio's ethics guidance notes — and independent benchmarking shows why verification matters: AI beats lawyers on question answering and summarization but loses on redlining and EDGAR research. A hallucinated citation is not a hypothetical failure mode.

This is why human verification must be structural, not aspirational. The ABA, California, Arizona, Idaho, and Alabama all call for written firm policies, training, and supervision of staff using AI. The Harvard Law School Center on the Legal Profession's interviews with ten AmLaw100 firms reinforce the point: AI flags and drafts, but the lawyer who signs decides.

For firms extending AI into agentic workflows, the same discipline applies to the underlying tooling. SkillzBench is building a free directory that security-scans and benchmarks agent skills — pinning every version to a commit SHA — so firms can evaluate what they deploy before client data is anywhere near it. The ethics layer is not optional overhead; it is what makes the speed usable.

Running a Pilot That Actually Proves Value (and Writing the Policy)

A sales demo that takes 16 hours of associate work and finishes it in 3–4 minutes sounds irresistible — until you realize a demo measures speed, and speed is the easiest thing to fake. That complaint-response result comes from real AmLaw100 interviews, not marketing copy, but it only counts if your firm can reproduce it on its own work.

A practical guide to AI legal document review puts it bluntly: a pilot without a written comparison measures speed and nothing else. Instead, define the task, run the same documents through both AI and a lawyer control group, and score the outputs side by side. Independent benchmarking from the Vals Legal AI Report shows why this matters: AI beat lawyers on question answering (94.8% vs. 70.1%) and summarization (77.2% vs. 50.3%), but lawyers won on redlining and EDAR research — wait, on redlining (79.7% vs. 65.0%) and EDGAR research (70.1% vs. 55.2%). A vendor pitching one number without the task, benchmark, and comparison group behind it is not making a claim about your work.

Document review and legal research are the proven use cases — 77% and 74% of AI-using legal professionals respectively, per industry data. But they are different jobs. A platform tuned for eDiscovery culling will not propose a redline; a contract platform will not cull a production set. A firm that buys litigation review software to speed up NDA turnaround has bought culling it doesn't need and lost the redlining it does.

The Harvard interviews found the billable hour — at least 80% of fee arrangements — survives AI, with clients wanting faster, higher-quality work rather than lower bills. But state bars constrain how you bill it: Arizona and California guidance says lawyers cannot bill hypothetical non-AI time, and Florida requires informing clients of the actual cost of AI use. Your firm-wide policy should cover:

• Informed client consent before entering confidences into AI tools, per ABA Formal Opinion 512 — boilerplate engagement-letter language is not enough
• Human verification of every output, with the signing lawyer accountable for cited authorities and analysis
• Pre-upload confidentiality checks: does the vendor retain documents, train on them, or share them with subprocessors
• Billing rules: actual time worked, never hypothetical time saved

If you extend the pilot into automated workflows, agent skills can carry parts of the load — SkillzBench is building a free directory and MCP connector to help you find and test which ones actually deliver lift before you trust them in a live matter.

Where AI Goes Next in Law Firms and How to Stay Current

The lawyers who got the most from AI in 2025 weren't the ones chasing the flashiest tools — they were the ones who embedded AI into systems their teams already trusted. That principle is about to define the next phase of legal AI entirely.

Industry analysis suggests 2025 was largely about efficiency gains, while 2026 will be the year agentic AI takes on defined tasks — research, drafting, and case management — inside the platforms lawyers already use, rather than in separate chat windows (Darrow). The market reflects this momentum: the global legal tech sector sits at $20.81 billion in 2025 and is projected to reach $65.51 billion by 2034.

Purchase decisions confirm the integration-first mindset. According to the ABA's Legal Industry Report, the top purchase criterion for legal AI tools is integration with trusted existing software, at 43% — well ahead of provider understanding of legal workflows (33%), trust in output (29%), and ethical alignment (26%). Firms aren't asking "what can AI do?" anymore; they're asking "what can AI do inside our document management system?"

Staying current doesn't mean buying every new platform. It means extending what you already have with tested, verified components:

• Prioritize tools that plug into your existing practice management and research stack — the 43% criterion exists because bolt-on tools create confidentiality and workflow risks.
• Run written pilots that compare AI output against lawyer work on defined tasks, not speed demos — independent benchmarks show AI still loses to lawyers on redlining and EDGAR research (Vals Legal AI Report data).
• Track regulatory developments — proposed Tenth Circuit Rule 46.5 revisions would require human verification of AI-assisted filings by January 2027.

For lawyers and legal operations teams building their own agent workflows, the next frontier is agent skills — instruction packages that give coding agents and chat assistants repeatable capabilities. SkillzBench is building a free, independent directory that verifies, security-scans, and benchmarks these skills so firms can adopt tested ones instead of guessing — plus a free MCP connector that links any agent to the skills you choose. Given that published research has found roughly 26% of skills carry at least one vulnerability, that verification layer matters doubly in a profession bound by confidentiality rules.

The firms that thrive in 2026 will treat agentic AI the way the best pilots already treat document review: as a powerful analytical layer that flags, drafts, and accelerates — while a lawyer verifies, decides, and signs. The technology keeps changing; that division of labor is what keeps it safe.