Legal teams engaged in litigation and investigations are operating in an environment defined by the rapid expansion of generative AI in document review, privilege analysis, and fact discovery. The current discussion extends beyond whether AI should be used. The more consequential question concerns how AI is used responsibly, defensibly, and in a manner consistent with attorney’s professional obligations.
Responsible use requires the right tools, disciplined workflows, and a clear understanding of who remains accountable for legal judgment. In this environment, attorneys remain responsible for defining relevance, interpreting facts, assessing risk, and owning legal determinations. Technology supports those functions, but it does not replace them.
BDO’s approach to AI‑enabled discovery and investigations is grounded in this principle. In partnership with Relativity, we deploy AI intentionally, combining our experience in Review Center in RelativityOne, managed document review, and Relativity aiR with the goal of providing structured oversight and supporting attorney judgment.
Attorney Responsibility in AI‑Enabled e‑Discovery
Attorneys should approach AI-supported outputs with a deliberate professional mindset grounded in judgment, skepticism, and accountability. Rather than relying on these outputs at face value, they are expected to critically assess their accuracy, relevance, and validity, applying legal expertise to determine what can be trusted, what requires further verification, and what should be rejected.
The framework – which I call Human‑as‑the‑KRINO – reflects five responsibilities that remain with the attorney.
- Know what the technology evaluated, including the data sources, instructions, and inherent limitations involved in producing the output. Attorneys must be able to explain what the system did, what it did not do, and where uncertainty exists. Outputs that cannot be explained with clarity warrant additional scrutiny.
- Example: For an aiR for Review project, attorneys should understand the prompt criteria used, the population selected, and how unsupported documents (i.e. documents with no text or very large document with too much text) are handled.
- Review the output independently through substantive evaluation rather than procedural confirmation. Review methods include sampling, citation verification, confirmation against source documents, and documentation of acceptance and rejection decisions.
- Example: Perform search and quality checks on the rationale and consideration values provided by aiR for Review and aiR for Privilege to understand the output. Consider generating conceptual analysis workflows based on this information to confirm that the AI output aligns with legal relevance. Perform quality control set sampling to measure accuracy.
- Interrogate the methodology, assumptions, confidence indicators, and omissions associated with the output. Attorneys examine confidence indicators, surface assumptions, and identify areas requiring deeper analysis or supplementation.
- Example: Examine whether the AI responses include confidence indicators or qualifiers (e.g., “likely”, “appears to”) and investigate why uncertainty might exist. Question performance metrics from validation rounds and human quality control (such as stability, overturn rates, and richness estimates).
- Normalize your output against legal experience, matter context, custodial behavior, privilege posture, and strategic considerations. Outputs are evaluated in light of case strategy, evidentiary context, and expectations formed through legal practice.
- Example: With aiR issue tagging, compare outputs against the case themes and elements of claims/defenses rather than focusing on surface-level relevance. For aiR for Privilege, validate the outputs against firm-specific privilege interpretation standards and not just generic definitions.
- Own the determination by accepting responsibility for both the process and outcome through professional licensing and ethical obligations. Determinations remain attributable to the attorney, supported by documentation that demonstrates thoughtful engagement and responsible reliance.
- Example: Review, evaluate, and understand the built-in auditing and reporting within aiR for Review that record version control, showing how prompts evolved over time, and preserve information on your sampling methodologies, validation results, and AI-to-human agreement rates. Familiarize yourself with Relativity auditing to capture information about who reviewed or accepted AI outputs.
As agentic and generative capabilities become more accessible, the distinguishing value provided by legal professionals increasingly resides in judgment rather than production. Drafting, classification, and summarization support efficiency. Trust is established through defensibility, consistency, and transparent reasoning.
Effective adoption, therefore, requires more than deploying advanced tools. It requires a governance model that defines how outputs are evaluated, challenged, refined, and relied upon.
The Art and Science of Prompting Legal AI Workflows
AI outputs are only as reliable as the instructions provided. Prompting quality directly influences relevance identification, issue detection, and narrative analysis. Vague instructions produce inconsistent results, whereas careful instructions aligned to legal objectives produce more reliable insight.
Over the course of using AI in my practice, I have gathered some tips for building effective prompts. I suggest including the following details:
- Describe a Persona - What is the persona that you want the AI to take on?
- Provide Context - What is the context that you want to tell your AI tool? The more specific you can be, the better.
- Probe - In other words, just as you would carefully question a witness on the stand, you should thoughtfully probe the AI. Examine its logic, test its reasoning, and challenge its conclusions so you can fully understand how it arrived at its responses. You should also invite the tool to ask you clarifying questions, ensuring you are actively engaged throughout the process rather than serving merely as a “human in the loop” – a phrase that can imply a passive role rather than that of a thought leader guiding the outcome.
- State a Goal - What is the goal that you want out of your prompt? What are you hoping to achieve?
- Describe the Deliverable - Be explicit about what you want the tool to do. The more specific your direction, the better positioned the tool is to produce useful, relevant, and well-reasoned results.
Prompting in legal matters requires both art and science. It requires the attorney’s understanding of the case, issues, and evidentiary nuances. It also requires technical discipline to structure instructions in a manner that is clear, scoped, and testable.
This dynamic is often summarized as “garbage in leads to garbage out,” highlighting the importance of structured and precise inputs in legal workflows. At BDO, our team addresses this through an intentional collaboration model between attorneys and project managers.
How BDO Project Managers Support Attorney‑Led AI Prompting
BDO project managers play a critical role in supporting AI workflows while preserving attorney leadership. The relationship is, intentionally, not one of substitution, but of enablement.
Attorney subject matter experts remain responsible for defining matter background, legal issues, relevance criteria, and privilege logic. They describe the case in plain language consistent with how the matter is understood legally.
BDO project managers then support this work by focusing on the technical components of prompting. This includes helping structure instructions for clarity, addressing scope alignment, managing ambiguity, and ensuring consistency across prompt iterations. Project managers translate attorney intent into technically executable configurations while preserving the substance of the attorney’s direction. This collaboration helps ensure that attorneys lead the process and technology follows.
Once an initial prompt is drafted by counsel, BDO project managers coordinate structured testing using curated sample sets. Outputs are reviewed with attorneys to confirm whether results align with legal expectations and matter objectives.
Where refinement is needed, project managers help identify how prompt language influences outputs and recommend technical adjustments. Attorneys retain decision authority over revisions and determine when prompts are acceptable for use at scale.
Prompt versions, changes, and validation results are documented throughout the process. This documentation supports transparency, defensibility, and informed reliance.
This iterative approach reflects Human‑as‑the‑KRINO in practice: attorneys remain accountable, project managers provide technical rigor, and AI supports insight generation.
Applying the Framework Across Relativity aiR
BDO supports clients by embedding these principles into discovery and investigation workflows, including matters that leverage Relativity aiR.
aiR for Review applies generative AI to support relevance assessment, issue identification, and investigative analysis across large document sets. Effective use depends on structured implementation and disciplined oversight.
With aiR for Privilege, attorneys establish privilege logic, definitions, and confirmation standards. BDO project managers support implementation and validation while helping to ensure that final privilege determinations remain attorney‑driven.
With aiR for Case Strategy, attorneys articulate the narrative questions they are trying to answer and the issues they seek to explore. BDO project managers assist by structuring prompts that support fact development, issue synthesis, and early insight.
Outputs are reviewed by attorneys and evaluated against known facts and legal strategy. Insights inform case development while determinations remain grounded in legal judgment.
Engagements typically begin with collaboration between client legal teams and BDO professionals to establish testing sets that support prompt development and validation. Prompts are refined through iterative review cycles until alignment with matter objectives is achieved. Once validated, prompts are applied consistently across broader data sets. The process concludes with quality checks, focused follow-up review, and documentation of how decisions were made.
Attorney Judgment as the Foundation of Responsible AI in e‑Discovery
Artificial intelligence capabilities will continue to evolve. Models and interfaces will change, and expectations for efficiency and insight will remain high.
What should not change, however, is attorneys’ professional responsibility. With the right technology and approach, attorneys can uncover the truth more quickly and defend their conclusions with confidence.
Graphics for this article were created by Kael Rose.
Recommended for you
