The Rise of AI Assistants: How Automation is Transforming the Legal Landscape

In 2023, "AI legal assistant" meant a specialized research or contract-review tool sold to law firms by a legal tech vendor. In 2026, it means something much broader: general-purpose frontier models doing substantive legal work under attorney supervision, embedded directly into how a practice runs. I have spent the past three years building exactly that into my own solo practice, and this article is my updated account of what actually changed, what did not, and what it means for lawyers and clients.

What Happened to the 2023 Legal AI Landscape

The original version of this article pointed readers to Casetext's CARA and ROSS Intelligence as the leading edge of AI legal assistants. That aged worse than almost anything else I have published, and the history is worth telling properly because it explains the market lawyers face today.

ROSS Intelligence shut down in January 2021, before my original article even ran, a fact I should have caught at the time. ROSS closed while defending a copyright lawsuit brought by Thomson Reuters over the Westlaw content used to build its research tool. The litigation over its training data outlived the company itself and became one of the defining early cases on whether training AI on proprietary legal databases is fair use. The lesson has stayed relevant: in legal AI, the provenance of training data is not an academic question. It can kill a company.

Casetext took the opposite exit. In 2023, shortly after launching its CoCounsel assistant on top of what was then OpenAI's newest model, Casetext was acquired by Thomson Reuters for a reported $650 million and folded into the Thomson Reuters product line. CARA as a standalone product is gone; its descendants live inside CoCounsel.

The pattern generalizes. The 2023 wave of narrow, single-purpose legal AI startups mostly ended in one of three ways: acquired by the incumbents (Casetext), shut down (ROSS and many smaller players), or repositioned as thin workflow layers on top of the frontier models. What almost nobody predicted in 2023 is that the general-purpose models themselves would become good enough at legal reasoning that the specialized middleware stopped being the interesting part.

From Point Tools to Frontier Models

The 2023 article described AI legal assistants as narrow specialists: one tool for research, another for contract review, another for e-discovery. That framing is obsolete. The current generation of frontier models, Anthropic's Claude 5 family (including Opus 4.8), OpenAI's GPT-5.x generation, and Google's Gemini line, are generalists that handle research synthesis, contract analysis, drafting, and document review in a single conversation.

Three capability shifts matter most for legal work:

• Context capacity. In 2023, a few dozen pages was the practical ceiling for most tools. A modern frontier model can hold an entire deal room's worth of documents in a single session: the full contract stack, the correspondence history, and the governing statutes at once. Cross-document analysis that once required a litigation support team is now a single prompt. I deliberately avoid quoting exact token figures here because they keep growing; the durable point is that document volume stopped being the bottleneck.
• Agentic operation. The 2023 tools answered questions. Current models execute multi-step work: read a complaint, extract the causes of action, check each element against the alleged facts, draft the responsive sections, and flag what needs a human decision. They use tools, browse sources, and run code. This is the difference between a search engine and a junior associate, with the crucial caveat that the associate analogy breaks exactly where accountability begins.
• Verifiable citation practice. Hallucinated case citations embarrassed lawyers in well-publicized sanctions cases starting in 2023, and courts responded with standing orders on AI-assisted filings. The models improved, but the professional rule did not change: every citation gets verified against the primary source before it goes into anything filed or sent. I treat citation verification as non-delegable, and I wrote up how that discipline works across my toolset in my practical AI guide for lawyers.

How I Actually Run AI in a Law Practice

Here is the part I could not write in 2023, because I was observing the technology rather than operating it. Today my practice runs on AI as infrastructure, with me as the supervising attorney at every decision point. Concretely:

• A production AI Legal Analyst on this site. Every page of Terms.Law carries an attorney-supervised AI assistant that triages visitor questions, analyzes uploaded documents, and routes matters to me when they need a licensed attorney. It is deliberately branded an AI Legal Analyst, not an "AI lawyer": it gives informational analysis, and I take over where legal advice begins. Building and running it taught me more about the real failure modes of legal AI than any vendor demo could.
• Client workrooms. For active matters I build private, matter-specific workspaces where documents, drafts, deadlines, and analysis live in one place, generated and maintained with AI assistance and reviewed by me. I now build the same kind of AI legal workrooms for other firms and businesses that want their matter files to work this way.
• Document generation at scale. The hundreds of template generators, calculators, and guides on this site are AI-assisted builds that I specify, review, and correct. AI does the assembly; the legal judgment about what a clause should say and when it applies is mine.
• Engineering with AI coding agents. The site itself, more than ten thousand static pages plus the backend that powers the chat and intake systems, is maintained with AI coding tools under my direction. A solo attorney maintaining that surface area was not a realistic proposition in 2023.

I also do this work for clients: reviewing a firm's AI stack for confidentiality, competence, and ethics-rule compliance is now a defined engagement on my AI implementation services page, and the AI Implementation Legal Hub collects the contracts, policies, and compliance mapping that AI-using businesses need, along with a live demo of the workroom approach.

What Has Not Changed: The Supervision Layer

The 2023 article said AI "complements but cannot wholly replace lawyers." That prediction held, but the reason is sharper than I understood then. It is not primarily that the models lack capability. It is that legal work is an accountability structure, and accountability cannot be delegated to software.

The professional-responsibility framework caught up quickly. The State Bar of California issued practical guidance on generative AI use in late 2023, and the ABA followed with Formal Opinion 512 in 2024. The core duties map cleanly onto AI use: competence (California Rule of Professional Conduct 1.1) requires understanding what the tool can and cannot do; confidentiality (Rule 1.6) governs what client information can touch which systems under what terms; and supervision (Rule 5.3) treats AI output like the work of a nonlawyer assistant, reviewed before it carries the lawyer's name.

In practice, the supervision layer in my workflow means: no AI output reaches a client, court, or opposing party without my review; no client confidence goes into a system whose data terms I have not read; and every legal citation is checked against the primary source. None of this is burdensome once it is built into the workflow. All of it is malpractice exposure if skipped.

The Impact on Legal Teams and Clients

The 2023 article predicted lower costs, new business models, and a leveling effect for small firms. Directionally right; the mechanism was surprising.

The leverage went to individuals, not institutions. Large firms adopted AI, but bureaucracy, billing models, and risk committees slowed them down. The biggest relative gain went to solo and small-firm lawyers, who could rebuild their entire workflow around the technology without asking permission. A solo attorney with a well-built AI stack now operates with the drafting throughput, response time, and matter-tracking discipline that used to require staff. My practice is the case study I know best, and I am far from alone.

Flat fees became easier to offer. When drafting time drops, hourly billing quietly punishes the lawyer for efficiency. AI-heavy practices tend to migrate toward flat-fee productized services, which clients strongly prefer anyway: a known price for a known deliverable.

The access-to-justice gap remains stubborn. The original article quoted an unsourced statistic here; I have removed it. The honest 2026 statement is that the Legal Services Corporation's Justice Gap research has documented for years that most civil legal problems of low-income Americans get little or no professional attention, and AI has so far changed that less than optimists hoped. Consumer-facing AI gives people better information, which is real progress, but information is not representation, and the unauthorized-practice line means the last mile still requires lawyers willing to serve that market.

Practical Adoption Guidance for Firms

The long FAQ section of the 2023 version speculated about adoption frameworks. Having now done the implementations, here is the condensed version I give firms that ask:

• Start with your own confidentiality map, not with tools. Decide which categories of client data may touch which classes of systems (enterprise API with zero-retention terms versus consumer chat apps) before anyone opens an account. Most AI ethics problems in firms are procurement problems in disguise.
• Automate the work you already check. First drafts, summaries, chronology-building, and intake triage are ideal because review is already part of the workflow. Do not start with anything that goes out the door unreviewed.
• Write the verification rule down. "Every citation checked against the primary source, every factual claim traced to a document" should be a written office policy, not a habit that lives in one lawyer's head.
• Expect the gains in turnaround, not headcount. The realistic near-term win is same-day drafts and faster client response, which compounds into reputation and repeat business. Firms that adopt AI to cut staff usually get quality problems; firms that adopt it to raise responsiveness usually get growth.

FAQ

Did AI replace lawyers or junior legal staff?

Neither, in the wholesale way 2023 commentary feared, but it did reshape entry-level work. Tasks that trained junior lawyers (first-pass research, document review, initial drafts) are now largely AI-assisted, so firms have had to become deliberate about how associates build judgment. The lawyers most displaced were not replaced by AI; they were outcompeted by other lawyers using it.

Is it safe to rely on an AI chatbot for legal advice?

For orientation and information, modern models are genuinely useful. For decisions with consequences, no: a chatbot does not know what it does not know about your facts, carries no duty to you, and cannot be held accountable when it is wrong. That is exactly the line my own AI Legal Analyst is built around: it informs, and a licensed attorney advises.

What should a small firm automate first?

Intake triage and first-draft generation. Both sit safely behind attorney review, both produce immediate time savings, and both teach the firm how the models fail before anything client-facing depends on them.

What happened to the specialized legal AI vendors?

Mostly acquisition or shutdown, as described above. The surviving value shifted to two poles: the frontier model providers on one end, and lawyer-supervised implementations tailored to a specific practice on the other. The undifferentiated middle got squeezed.