The most common advice about legal AI is also the most misleading: give the model more context, let it handle the research, and use it to move faster across the whole matter. That sounds efficient. In practice, it creates exactly the kind of workflow that gets lawyers into trouble.
Why most lawyers are using AI wrong has less to do with the existence of the tools than with the assumptions behind them. Many firms are treating general AI as if it were a research platform, a cite-checker, and a junior lawyer rolled into one. It isn't. It can draft. It can summarize. It can help frame issues. But it does not know what is true in the way legal work requires truth to be known.
Managing partners don't need another abstract debate about whether AI matters. It does. The practical question is narrower: where does it belong in the workflow, and where does it absolutely not belong without controls? Firms that answer that well will get real efficiency. Firms that answer it badly will spend their savings on cleanup, rework, and risk.
The Hidden Risks of the Legal AI Gold Rush
The legal market has moved past experimentation. The risk now comes from confident misuse.
A widely repeated assumption says adoption itself is the competitive advantage. It isn't. Unstructured adoption is often just unstructured liability. More than 300 cases of AI-driven legal hallucinations have been identified since mid-2023, with at least 200 in 2025 alone, and courts have sanctioned lawyers for filing briefs with fabricated citations, as discussed in this analysis of AI failures in legal practice. That same discussion notes that 74% of legal professionals use AI for legal research and 59% for drafting briefs.
That combination matters. High adoption in low-risk administrative work is one thing. High adoption in research and citation-heavy drafting is something else entirely.
Where firms are getting exposed
The pattern isn't mysterious. A lawyer asks a general model for cases, the output looks polished, the draft moves forward, and nobody checks every proposition against primary authority. The problem isn't only the bad output. The problem is the workflow that lets bad output survive long enough to reach a filing, a client, or a negotiation.
Practical rule: The moment AI output includes a case, statute, quote, or jurisdiction-specific proposition, a human has to verify it against a trusted authority before it leaves the firm.
Managing partners should also think beyond courtroom embarrassment. Public filings create the cleanest examples, but internal consequences show up earlier. Matter teams lose time rechecking shaky drafts. Associates absorb the wrong lesson about what “finished” work looks like. Clients start asking uncomfortable questions about supervision, confidentiality, and accountability. Those concerns overlap with the broader AI liability risks businesses should know, especially when AI use outpaces policy.
The real issue is structural
This isn't a one-off misuse by careless lawyers. It is a structural mismatch between how general AI generates language and how legal work establishes truth.
When a firm installs AI into the wrong stage of the workflow, risk stops being accidental. It becomes predictable. That is why the gold rush framing is so unhelpful. Speed is only a benefit if the process preserves legal judgment, source traceability, and review discipline.
The Core Misunderstanding AI Is Not a Search Engine
Most legal AI mistakes begin with the same category error. Lawyers use a general model as though it were Westlaw, Lexis, or a specialized legal research platform. It isn't.
A general-purpose large language model is a probabilistic predictor, not a legal reasoning engine. As the New York State Bar Association discussion of AI's limits in legal work explains, these systems optimize for linguistic plausibility. That means they can produce fluent, persuasive, and completely false legal content, including invented case law that sounds credible.
A better mental model
Think of a general AI tool as an eloquent intern with extraordinary speed and zero built-in authority checking.
It can write in the style you want. It can organize arguments. It can turn rough notes into readable prose. But it doesn't know which cases are binding, which statute controls, whether the jurisdiction matters, or whether a quote exists unless that information is grounded and then verified by someone else.
That distinction changes everything.
A search engine or legal database retrieves sources. A general model generates text. Those are not the same function, and legal workflows break when lawyers collapse them into one.
Why hallucinations are normal behavior
In legal circles, the word hallucination gets used as if it means a rare malfunction. It doesn't. In a general language model, hallucination is a byproduct of the architecture. The system is trying to predict the most likely next words, not certify legal truth.
That is why superficial review often fails. The output sounds right. The citation format looks familiar. The legal rule appears balanced. Even experienced lawyers can miss the defect if they review for style before they review for source validity.
The danger isn't clumsy prose. The danger is polished error.
What works instead
The fix is task matching. Use general AI where language generation helps and authority is not being outsourced. Avoid using it as the final source of legal truth.
A simple comparison makes the distinction easier to enforce:
| Task | General AI | Legal-specific research tool |
|---|---|---|
| Brainstorm arguments | Good fit | Useful but not necessary |
| Summarize a draft or transcript | Good fit | Sometimes helpful |
| Generate a first draft structure | Good fit | Sometimes helpful |
| Find binding authority | Poor fit | Required |
| Validate citations and propositions | Poor fit | Required |
| Confirm jurisdiction-specific law | Poor fit | Required |
Firms that understand why most lawyers are using AI wrong usually make one early shift. They stop asking the model, “What is the law?” and start asking, “Help me draft, organize, compare, or stress-test what I will verify elsewhere.”
The Five Common Traps of Everyday Legal AI Use
The firms struggling with AI usually aren't failing in dramatic ways. They're making ordinary decisions that seem harmless in the moment.
Lawyers themselves already recognize the boundaries. In a 2025 survey, 83% said using AI to provide legal advice would be “a step too far,” and non-users cited accuracy at 43% and data security at 37% as their top concerns, according to Thomson Reuters' review of how AI is transforming the legal profession. Those concerns map directly to the five traps below.
Trap one: treating output as finished work
This is the fastest route to bad filings and bad client work.
A lawyer asks for a summary, outline, or draft section. The result looks polished. The team trims a few sentences, adds a caption, and moves on. That shortcut feels small because the prose is clean. But legal work isn't judged by fluency. It's judged by authority, accuracy, and judgment.
The safer posture is simple: AI produces a first pass, never a final product.
Trap two: pasting in confidential client material without controls
Public or loosely governed tools create obvious confidentiality problems. Lawyers know this in theory, then still drop draft agreements, fact chronologies, HR records, or internal emails into tools that haven't been approved by the firm.
This is less a technology problem than a policy failure. If the firm has not defined approved tools, acceptable inputs, and redaction standards, the burden gets pushed onto individual lawyers who are trying to move quickly.
Trap three: using the wrong tool for the task
A general model can help draft a client alert. It is a poor substitute for a legal database. It can help turn messy notes into a memo outline. It should not be trusted to identify controlling authority.
Different tools belong in different lanes:
- General LLMs: Drafting, brainstorming, summarizing, rewriting, issue framing.
- Legal research platforms: Authority retrieval, citator functions, validation, jurisdiction-specific analysis.
- Document review systems: Contract analysis, classification, large-scale review workflows where the platform is designed for auditability and supervision.
A lot of AI disappointment comes from firms buying one capability and expecting five.
Trap four: ignoring jurisdiction and matter context
Generic output often sounds broad and useful because it strips away exactly what lawyers must add back in. Which court. Which state. Which standard. Which procedural posture. Which client risk tolerance.
That creates a subtle quality problem. The draft may not be obviously wrong. It may specifically be wrong for this matter.
A technically competent draft that misses forum, posture, or governing law is still bad legal work.
Trap five: weak prompting that invites weak answers
Vague prompts produce broad prose, and broad prose hides errors.
Compare these two requests:
- “Draft a motion section arguing waiver.”
- “Using only the facts below, draft a first-pass argument section on waiver under the contract terms identified in Exhibit B. List factual gaps, assumptions, and issues requiring jurisdiction-specific validation. Do not invent authority.”
The second prompt doesn't make the model perfect. It makes the model easier to supervise.
A quick diagnostic for managing partners
If your teams are doing any of the following, they are likely using AI wrong:
- Submitting AI text into client work without source verification
- Using unapproved tools with live client information
- Expecting one platform to handle research, drafting, and validation equally well
- Accepting generic language in place of matter-specific analysis
- Training lawyers on prompts before training them on review standards
Those aren't edge cases. They're daily operating mistakes.
The Co-Pilot Workflow A Smarter Framework for AI
The practical model for legal AI is supervision, not delegation.
General AI predicts plausible language. Legal work demands controlled reasoning, source discipline, and answers a lawyer can defend. That mismatch is why firms get into trouble when they treat a chatbot like a junior associate with perfect recall. The safer model is a co-pilot workflow in which the lawyer sets the boundaries, the system helps with a defined task, and human review decides what survives.
Step one: define the assignment narrowly
Start with a task that has a clear output, clear inputs, and clear limits.
Good uses include summarizing a deposition for admissions, drafting a factual background from approved notes, revising a client email for tone, or generating counterarguments against a position already developed by counsel. Those tasks benefit from speed and pattern recognition without asking the model to make the final legal call.
Requests such as “analyze this matter” or “tell me the strongest argument” usually fail because they invite the model to fill gaps with probability. That is acceptable in brainstorming. It is dangerous in legal analysis.
Step two: match the tool to the risk
Different tasks need different systems.
Use a legal research platform for authority retrieval and citation checking. Use a general model for language work, issue spotting, or first-pass organization when the source set is controlled. Use a separate operational process for intake, marketing, internal knowledge management, or workflow automation.
Some firms also use internal operations teams or outside support to set guardrails around approved tools, access controls, and review steps. Gorilla is one example of operational support for firms that want to scale AI use safely without outsourcing legal judgment.
Step three: curate the record
File dumping is lazy workflow design.
A smaller, relevant packet usually produces better output than a full matter folder because the model has fewer chances to mix signal with noise. Give it the contract section at issue, the key email chain, the hearing excerpt, or the approved chronology. Leave out duplicate drafts, irrelevant exhibits, and background material that does not affect the assignment.
For firms trying to tighten this process, a short set of prompting standards for legal work helps lawyers define what the model may use, what it must ignore, and what uncertainty it must surface.
Step four: direct the drafting process
A co-pilot workflow depends on staged output.
Ask for a first-pass draft, then force the model to expose its weak points. Require it to separate record facts from inferences, identify assumptions, list missing facts, flag where jurisdiction-specific review is required, and restate the analysis in a more neutral or more adversarial tone. That sequence is more useful than a single prompt asking for “the answer.”
I advise firms to treat AI output as a working memo from an assistant who writes quickly, sounds confident, and cannot be trusted on its own. Teams that understand that trade-off get real efficiency. Teams that skip it create expensive cleanup work.
Step five: verify adversarially
Review should be structured and skeptical.
| Review layer | Human question |
|---|---|
| Facts | Did the model state facts that are actually in the record? |
| Law | Is each legal proposition supported by trusted authority? |
| Scope | Did the draft answer the question asked, and only that question? |
| Jurisdiction | Does the analysis fit the governing court and law? |
| Strategy | Would I sign my name to this as my own judgment? |
That final question matters most. A co-pilot workflow works because it accepts the basic reality of legal AI. The model is probabilistic. The lawyer is accountable. Firms that build around that distinction can get speed without giving up standards.
How to Craft Prompts That Elicit Legal Insight
Prompting is where legal AI becomes either useful or expensive.
A weak prompt asks for an answer. A strong prompt asks for a controlled draft, grounded in defined materials, with visible uncertainty. That is a very different use of the tool.
One of the most useful contrarian lessons in legal AI is that more context isn't always better. A recent legal-industry discussion argues that feeding in too much irrelevant material can make output less reliable because excess text muddies the retrieval and response process. The better approach is selective context engineering with curated documents and narrow task framing, especially for matter-specific work, as explained in this discussion of why less context can produce better legal AI results.
The bad prompt and the better prompt
Here is the kind of instruction that wastes time:
Review this case file and tell me the best arguments.
It is too broad. It invites assumptions. It doesn't specify source boundaries, deliverable format, or what should happen when the model lacks information.
A better version looks like this:
Based only on the attached termination letter, contract clause, and chronology, draft a first-pass list of arguments for and against wrongful termination. Separate record facts from inferences. Identify missing facts that would change the analysis. Do not cite cases. Do not assume facts not provided.
The difference is not cosmetic. The second prompt tells the model what materials count, what output format to use, and what not to do.
Use prompt formulas by task type
Different legal tasks need different prompt designs.
For summarization
Use this structure:
- Define the source such as a deposition excerpt or draft agreement
- State the audience such as partner, client, or litigation team
- Specify the output such as bullet summary, risk list, or chronology
- Require traceability by asking the model to point to the source passage
Example:
Summarize this deposition excerpt for partner review. List key admissions, evasions, and factual inconsistencies. For each point, identify the page or passage from the text provided.
For argument mining
Example:
Read the opposing brief excerpt below. Identify its three strongest arguments, its unstated assumptions, and the factual or legal points most vulnerable to challenge. Limit your response to the text provided and note where additional authority would be required.
For first-draft writing
Example:
Draft a neutral first-pass section for a memorandum on the contractual notice issue using only the contract language and timeline below. Mark any sentence that depends on an assumption. End with a checklist of items a lawyer must verify before using this draft.
Ask the model to expose uncertainty
Lawyers often prompt AI to sound confident, then complain when it does.
A better pattern is to force the model to reveal limits. Ask it to identify missing facts, competing interpretations, and reasons its answer may be incomplete. That alone improves quality because it changes the output from “finished prose” to “supervised work product.”
For teams that want a repeatable method, this guide on how lawyers can write better AI prompts is useful as an operational reference.
The less-is-more principle in practice
Don't upload the whole matter file unless the task requires the whole file. Most of the time, a smaller curated set works better:
- For a contract issue: relevant clauses, amendment history, and disputed correspondence
- For a motion draft: your outline, governing standard, and verified fact section
- For a deposition summary: the relevant witness segment, not every transcript from the case
That discipline improves output quality and lowers review burden. It also trains lawyers to think clearly about the assignment before they ask the machine to write.
Your Firm's AI Adoption Checklist for 2026
The firms that handle AI well don't rely on informal habits. They build policy, training, and review into the operating model.
A managing partner should be able to answer seven questions clearly.
Governance and approved use
- Do we have a written AI policy that defines approved tools, prohibited uses, confidentiality rules, and escalation procedures?
- Have we defined acceptable use cases such as summarization, first drafts, transcript analysis, and internal brainstorming, while restricting citation-heavy or final-advice use without verification?
- Do we have a client communication position on when AI use should be disclosed?
Training and competence
- Are lawyers trained on review, not just prompting? Prompting without verification training creates polished risk.
- Do associates know the difference between drafting assistance and legal research? If they don't, they will use the wrong tool in the wrong place.
- Are partners reviewing AI-assisted work in a way that teaches judgment? Supervision matters more now, not less.
Firms don't need universal AI enthusiasm. They need universal clarity about where the tool stops and the lawyer starts.
Security and workflow controls
- Have we vetted the tools for security and data handling?
- Do we have a safe environment for experimentation so lawyers aren't improvising with public tools?
- Is every AI-assisted work product subject to human sign-off before it goes to a client, court, or counterparty?
A simple leadership test
If a partner cannot tell you which tasks AI may assist with, which system should be used for which task, what data may be entered, and who verifies the output, the firm is not adopting AI. It is merely allowing it.
That distinction will matter even more as firms move into 2026 planning. The market will reward disciplined implementation, not casual experimentation. Firms that want a practical model for safer scaling can review approaches like how law firms use AI safely to scale operations.
Gorilla helps law firms build practical growth systems around content, web infrastructure, SEO, and AI-enabled operations without confusing marketing automation with legal judgment. If your firm wants a clearer plan for using AI safely while improving visibility and intake performance, a conversation with Gorilla is a sensible place to start.
