Designing hybrid workflows that hold up

Most discussions of hybrid AI and traditional legal research workflows settle into one of two framings. In the first, AI does the initial pass and the lawyer verifies. In the second, the lawyer handles the conceptual work and AI manages the mechanical tasks. Both framings produce workflows that function adequately on routine assignments and break down precisely when the work is most consequential.

The first-pass model understates the verification problem it is designed to solve. The errors it catches most reliably are the ones that look like errors: a citation that does not exist, a case name that does not match the proposition it is offered to support. The harder errors are the plausible outputs: the real citation with the wrong holding, the standard applied in the wrong jurisdiction, the characterization of a line of cases that a lawyer who had read them would not recognize. Catching those requires the same subject matter knowledge the lawyer was supposed to save time by not deploying. The mechanical split has a different problem: the boundary between mechanical and judgment-requiring tasks is less stable than the model assumes. Legal research tasks that look mechanical, generating search terms, identifying secondary sources, summarizing a regulatory scheme, often require embedded judgment at every step. Handing them off entirely produces outputs that no longer reflect the lawyer's actual understanding of the problem.

The course's approach to hybrid workflows is built around a different question: not whether to use AI, but which tasks in a research workflow actually benefit from AI involvement, and which tasks get worse when AI is in them. That question does not have a universal answer, and the taxonomy this note describes is not meant to provide one. It is meant to give students a framework for making the decision themselves, task by task, in the specific context of whatever research problem they are working on.

The taxonomy has four categories, and the value is in the habit of applying it before starting a task rather than in having memorized the categories. The first category covers tasks where AI involvement improves speed without meaningfully degrading quality: generating a list of search terms, drafting an initial outline of a familiar legal standard, producing a summary of a document the lawyer has already read and understood. The output is useful, the failure modes are manageable, and the time savings are real. These are the tasks where AI earns its place in a workflow without much controversy.

The second category covers tasks where AI involvement introduces errors that are genuinely difficult to catch: synthesizing case law across jurisdictions, characterizing the holding of a case the lawyer has not read, identifying the elements of a standard in an unfamiliar practice area. These tasks produce outputs that look authoritative and are wrong in ways that require legal knowledge to identify. The deeper problem is that the outputs of AI doing these tasks look similar to the outputs of AI doing them correctly, and the verification burden is correspondingly high.

The third category covers tasks where the cognitive work of doing it yourself is the point: reading a case for the first time, forming an initial judgment about whether a line of authority supports a proposition, deciding whether a research trail has covered the relevant ground. These are tasks where outsourcing to AI does not save time in any meaningful sense, because the time spent is the time building understanding, and building understanding is what the task is for. A student who asks AI to summarize a case they have not read and then works from the summary has not done the task. They have done a different, less useful task while believing they have done the original one.

The fourth category is the most practically important and the hardest to teach: tasks where AI is genuinely useful for orientation but where human judgment has to take over quickly. Getting a preliminary map of an unfamiliar area of law, identifying the vocabulary a field uses to describe a problem, understanding the basic structure of a regulatory scheme. AI does those tasks well enough to get a lawyer into the substance faster than they could get there alone. The failure mode is staying in AI-assisted orientation mode longer than the task warrants, treating the map as a destination rather than a starting point. Students who understand this category have the most useful piece of the taxonomy, because most of the interesting hybrid workflow decisions live there.

A common response to AI unreliability is to tell students to verify every output thoroughly before relying on it. That guidance is professionally correct and practically useless as workflow instruction. If every AI output requires full independent verification, the efficiency rationale for AI involvement disappears entirely, and students are left with a tool they have been told to distrust without a framework for when to trust it. The result tends to be one of two failure modes: students who verify everything and find the workflow too burdensome to sustain, or students who take the instruction as aspirational and verify selectively according to instincts they have not been helped to develop. Neither outcome produces the calibrated professional judgment the course is trying to build.

A functional hybrid workflow requires calibrated trust, which means students need explicit instruction in which task types warrant heavy verification, which warrant lighter verification, and which warrant doing the underlying work yourself rather than verifying an AI output at all. The taxonomy provides that instruction not by specifying the right answer for every task but by giving students a framework for arriving at it. Heavy verification is warranted for second-category tasks, where AI involvement introduces hard-to-catch errors on work that will be relied upon directly. Lighter verification is often sufficient for first-category tasks, where the failure modes are more visible and the outputs are easier to check. Third-category tasks do not warrant verification because the right answer is not to use AI for them at all. Fourth-category tasks warrant calibrated judgment about when the orientation phase has ended and the independent work needs to begin.

This is worth being direct about: the taxonomy came from watching where student workflows actually broke down, not from theory. Across multiple cohorts working on research tasks with AI tools, the failure patterns were consistent enough that they suggested categories. Students who used AI to synthesize case law they had not read produced memos that looked authoritative and were unreliable in ways they could not identify themselves. Students who used AI for initial orientation and then moved into independent research produced better work than either pure AI or pure traditional research would have, but struggled to articulate why or when they had made the transition. Students who used AI for first-category tasks, search terms, outlines, initial structuring, found the tools genuinely useful without most of the failure modes that make AI in legal research a professional liability problem.

Instructors should treat the taxonomy as a starting point that their own students will pressure-test, not as a fixed framework to be handed down. The categories are not exhaustive, the boundaries between them are genuinely fuzzy in practice, and a research task that belongs clearly in one category in a familiar practice area may belong in a different category in an unfamiliar one. What the taxonomy offers is a vocabulary for the conversation, not a resolution of it. Students who leave the course arguing about which category a task belongs in have internalized the framework. Students who leave it reciting the four categories have not.

Instructors who want to use the taxonomy as a design tool rather than just a teaching concept have a few practical options. The most direct is to sequence exercises so students encounter tasks from each category in order, starting with first-category tasks where AI involvement is clearly beneficial and the failure modes are visible, then moving through the categories in a way that builds the habit of asking which category applies before reaching for a tool. The sequencing matters because students who encounter AI at its most useful before they encounter it at its most deceptive are better positioned to distinguish between the two.

A more demanding design option is to build exercises that require students to declare which category a task belongs in before they begin and to argue for their categorization as part of the deliverable. That structure surfaces the judgment the taxonomy is trying to build rather than letting it remain implicit. It also creates a natural point of assessment: the quality of a student's argument for their categorization is often a better indicator of whether they have internalized the framework than the quality of the research itself. Students who categorize correctly but cannot explain why have the right answer for the wrong reason. Students who categorize incorrectly but reason carefully about why often have a more useful understanding of the framework than students who got it right by intuition.

No taxonomy survives contact with a genuinely novel research problem, and instructors should expect students to find the edges. When a task is unfamiliar enough that a student cannot determine which category applies, that uncertainty is exactly where professional judgment belongs, and instructors should name it as such rather than treating it as a gap to be filled. A student who recognizes that they do not know how to categorize a task and slows down accordingly has done something more important than a student who categorizes confidently and proceeds. The taxonomy is a tool for activating professional judgment, not a substitute for it, and the clearest sign that a student has internalized it is that they know when to stop applying it.

Instructors should also expect the taxonomy to age. The four categories reflect the current state of AI capability in legal research, and that state will change. Tasks that belong in the second category today, where AI involvement introduces hard-to-catch errors, may move toward the first category as tools improve and verification processes become better integrated into platforms. Tasks that sit comfortably in the fourth category may shift as AI orientation tools become more reliable and the transition point to independent work becomes easier to identify. The goal is for students to leave the course with the habit of asking the question, not with a fixed answer to it, because the fixed answer will be wrong before they are three years into practice.