AI is no longer a future-state conversation for law firms. It is already embedded in the tools teams use every day, from email and document management to eDiscovery and legal research platforms. The real question is not whether AI will show up inside firm workflows. The question is whether the firm will adopt it intentionally, with clear governance, realistic expectations, and the operational readiness required to avoid preventable risk.
At the K2L Conference breakout session hosted by Cooley LLP, attendees received a practical grounding in what “AI” actually means in a law firm context and how to evaluate AI claims from vendors without getting pulled into hype. The session emphasized technical fundamentals, the structural realities of law firm decision-making, and pragmatic use cases that balance speed with professional responsibility.
1) Start by defining what “AI” means
A recurring theme was that “AI” is an umbrella term, and not all AI behaves the same way. For practical decision-making, it helps to separate two model families that show up most often in legal technology.
Discriminative systems classify. They are typically trained to do a specific task such as categorizing documents in eDiscovery or identifying clauses. These systems tend to be more predictable and can be evaluated with standard measures such as accuracy and recall.
Generative systems create. Large language models fall into this category. Outputs are non-deterministic, meaning the same prompt can produce different results. These systems can be excellent for drafting and rewriting, but they require governance because they generate plausible language rather than guaranteed truth.
A simple evaluation improvement comes from a simple question: is the tool classifying within a bounded workflow, or generating new text probabilistically?
2) Treat generative AI like fast pattern recognition, not factual reasoning
Generative AI behaves like a probabilistic prediction engine. It estimates the next token based on patterns learned during training, combined with the context provided in the prompt. The important operational implication is that generative AI should not be treated as a search engine or a factual authority.
This is why the session challenged the common framing of “hallucinations.” The issue is not that a small portion of an otherwise factual answer is occasionally wrong. The issue is that generative output is always probabilistic, and reliability depends on context, constraints, and verification.
For law firms, this distinction matters most in research and citation-sensitive workflows. Unbounded prompts can produce confident-sounding responses that are not grounded in real authorities, which can quickly escalate from inefficiency into reputational and disciplinary risk.
3) Newer features reduce risk, but they do not eliminate the need for oversight
Vendors frequently position emerging capabilities as a solution to reliability concerns. The session framed these as valuable risk reducers, but not guarantees.
Retrieval augmented generation (RAG) can ground responses in documents a firm supplies and may surface citations. That can improve usefulness, especially in summary-oriented workflows, but it does not make outputs deterministic. Citations still require validation.
Reasoning methods and multi-step prompting can improve performance for structured tasks, but they do not create real judgment. Outputs remain probabilistic and should be governed accordingly.
Agentic workflows can chain multiple steps and tools toward a goal. These are promising, but in legal contexts they can be difficult to supervise because transparency decreases as workflows become more complex.
The consistent takeaway is that capability upgrades change how risk is managed. They do not remove the need for verification, training, and clear boundaries.
4) Law firm structures create predictable friction for AI adoption
The session offered a useful way to understand why technology rollouts often stall inside firms. Law firms tend to operate with governance dynamics that resemble a multi-branch system, where decision-making is distributed across firm leadership, partnership committees, and a culture that defaults to risk-first thinking.
Several structural barriers surfaced repeatedly.
Partnership-driven budgeting often prioritizes short-term horizons, while infrastructure and AI readiness investments may take years to show return.
Decision-making can be fragmented, which makes it hard to define an enterprise AI strategy and even harder to enforce consistent standards.
Technical debt can limit what the firm can deploy securely and at scale, regardless of vendor promises.
A “buy AI” instinct can lead to tool-first decisions before the firm has clarified which problems matter most and what success should look like.
This is why AI adoption tends to be less about picking a tool and more about aligning governance, infrastructure, and operational change.
5) Professional responsibility remains the same, but the operational burden shifts
A key point from the discussion was that professional obligations have not changed in response to AI. What changes is how those obligations show up in day-to-day work when probabilistic tools are involved.
Training becomes central. Administrators and practice leaders need a shared framework for when AI assistance is appropriate, what requires verification, and what types of matters have additional procedural or regulatory constraints. Domain-specific requirements can vary significantly across practice areas, such as patent prosecution rules, eDiscovery procedures, and sensitive regulatory matters.
The governance-first posture is becoming a best practice. Establishing principles, acceptable use standards, and vendor requirements is often the right starting point, before broad deployments occur.
6) Avoid the “problem problem” by starting with the business need, not the tool
A practical strategy framework emerged clearly: successful adoption starts with the problem being solved.
The sequence looks like this: define the business rationale, confirm practice alignment, prioritize high-impact use cases, evaluate fit between the problem and the solution, execute thoughtfully, and iterate.
This approach prevents a common failure mode where leadership asks for “AI” without defining the outcome, and operations gets pushed into selecting tools that do not map to measurable value.
7) Four use-case patterns that help firms apply AI safely
The breakout session grouped common applications into four patterns that can be used to guide pilots and policies.
Direct model queries for drafting and rewriting
These workflows are strongest when tasks involve drafting internal content, revising language, improving clarity, or adjusting tone and consistency. Results improve materially when prompts include clear intent and sufficient context.
Summary and retrieval with an accuracy versus recall mindset
Legal workflows often require teams to understand the trade-off between accuracy and recall. Accuracy describes how much of what the system returns is correct. Recall describes how much of the relevant universe was actually found.
Generative tools do not solve recall in the way classification tools can. Even when citations are provided, verification effort can reduce time savings. This is why expectation setting and benchmark testing matter, especially when client-facing representations of diligence and completeness are involved.
“Closed universe” research to reduce hallucination risk
A strong pattern for research-adjacent use is to keep AI inside a bounded document set. Traditional methods identify the relevant authorities first. Those materials are then loaded into a system capable of summarizing and answering questions within that curated corpus. This reframes the task from open-ended research to constrained summarization and extraction.
Drafting support using context-rich instructions
Drafting workflows improve when inputs include precedent examples, formatting requirements, reference materials, and a clear audience. There is a practical limit, however. For highly specific clause drafting, time spent writing instructions can exceed the time saved, so these workflows work best where structure and examples already exist.
8) AI is already embedded, so governance becomes the sustainable response
The session’s broader message was that AI is increasingly a foundational layer inside software platforms. A blanket “no AI” posture is difficult to sustain because even basic features may include AI-assisted functionality.
The sustainable approach is to separate deterministic workflows from probabilistic ones, set policies based on risk level, and train teams to use generative tools within controlled, auditable boundaries.
What this means for law firm leaders
For managing partners, executive directors, CIOs, and administrators, the next step is not to rush into an AI purchase. The next step is to make the firm “AI ready” in a way that protects clients and supports the business of law.
At Xantrion, this typically maps to four practical priorities.
- Establish governance before scale. Define acceptable use standards, confidentiality boundaries, matter risk tiers, and verification expectations. Clear rules prevent inconsistent behavior and reduce the likelihood of avoidable incidents.
- Reduce technical debt that blocks safe deployment. Identity, access controls, device management, data classification, and secure collaboration are foundational. Without them, AI tools increase risk faster than they create value.
- Evaluate vendors with operational criteria, not marketing claims. Require clarity on whether systems are deterministic or probabilistic, what data is retained, how outputs are benchmarked, and what auditability exists for supervision and compliance.
- Train attorneys and staff on practical use patterns. The biggest risk is not “AI” in the abstract. It is professionals using probabilistic outputs as if they are authoritative without context, constraints, or verification.
Handled well, AI becomes a productivity accelerator for both front-office and back-office work. Handled casually, it becomes a source of reputational, ethical, and client relationship risk. The takeaway from this breakout is that law firms can adopt AI responsibly, but doing so requires the same discipline firms apply to any other high-impact operational change: governance, readiness, training, and measurable outcomes.
