When the Playbook Starts Running Itself

In our last article, we established three essential elements of discovery: find the facts that shape the matter, produce what must be produced, and protect what should not be exposed. The first element drives strategy, yet for twenty years we have often run it too late because the pipeline left us few better options. Agentic AI changes that sequence. It lets legal teams move more fact development to month zero, before the budget is spent and before case strategy hardens around an incomplete record. 

The next challenge is determining what makes month-zero work dependable. Prompting can be part of that answer, but it is not the whole answer. A prompt is still a human firing off an instruction at a late stage, shaped by what they remembered to ask and limited by the corpus the pipeline already produced. The bigger move is encoding human judgment as a process: the department’s best questions, escalation points, validation steps, and matter-type lessons applied consistently across the discovery lifecycle. Prompting showed what was possible. Agents make it operational. 

The distinction matters in practice. Say a team builds a strong prompt for regulatory investigations. After collection, search, and loading, it finds custodians with decision-making authority and flags communications showing knowledge or approval. It works across matters with similar assumptions about sources, evidence, and escalation. Then an executive conduct investigation arrives. The same prompt returns performance reviews, HR policies, and the termination file. But the most important evidence may sit in a Slack thread the workflow never treated as a primary source, because hostile conduct does not always look like a decision chain. The prompt did what it was asked to do. The workflow failed earlier. It never asked whether this matter needed different sources, patterns, escalation rules, or validation. Prompting was the tip of the iceberg. Agents make the larger system visible: not better one-off questions, but the client’s legal and discovery judgment translated into repeatable process we can execute across matters. 

The process already exists 

Legal departments already solved this problem once, for people. We call it a playbook: what to collect, what relevant evidence looks like, what gets escalated, and how results get validated before anyone relies on them. In one matter type, relevant evidence may look like formal decision chains. In another, it may look like informal patterns across channels, or missing documentation where documentation should exist. A playbook is the department’s accumulated judgment and hard lessons, written down. The weakness was never the idea. The challenge was execution: the rulebook may be sound, but someone still has to remember when and how to use it under pressure. 

How agentic AI turns playbooks into skills 

This is where agentic AI moves beyond prompting. An agent does not just answer a question; it can follow a process. A skill is the way that process gets packaged: the rulebook, the reference materials, the sequence of steps, the checkpoints, and the conditions that require human review. In discovery terms, that means we can give an agent more than a prompt. We can give it the ESI protocol, privilege guidance, escalation matrix, source assumptions, validation standards, and matter-type rules that tell it how the work should be done. 

That is the step change. Old automation could execute only the rules pre-coded in advance. Prompting let a human ask better questions, one at a time, usually after the matter was already underway. Skills let agents carry the broader rulebook forward as repeatable work: check these sources, apply these exceptions, escalate these calls, validate before scaling, and document what changed. Each run records what was tested, what surfaced, what got missed, and what changed the next time. The skill does not make the answers magically smarter. It makes the process visible, reviewable, and capable of improving. 

Client-specific service, not matter-specific prompting 

This is where the service model changes. A prompt usually ends with the matter it was written for. A skill can carry forward how we serve a client. It reflects the client’s playbook, data sources, communication patterns, risk tolerance, and escalation preferences, then becomes part of the way we operate as an extension of that legal team. Matter after matter, the engagement does not start from zero. The process gets sharper, the client context is easier to preserve, and the work reflects what we have already learned together. 

A skill also helps preserve the thing turnover usually erodes: institutional memory. Anyone who has watched a tenured employee give notice knows what walks out with them. The senior paralegal retires, the project manager who knew the client’s quirks takes another job, the deputy GC who carried the last three investigations moves on, and the knowledge goes too, because it lived in heads, inboxes, and habits. Encode it into a skill, and more of that playbook survives the people who helped build it. 

This does not stay confined to review. Collection scoping, search validation, privilege indicators captured during ingestion, PII flagged before the redaction crunch—each used to be a playbook step someone had to remember at the right moment. Each can become part of a skill an agent executes. That is how the same logic reaches all three jobs of discovery: find the facts earlier, produce more consistently, and protect sensitive material before the end of the pipeline. 

The written playbook is not enough 

Now the hard part. A written playbook captures the standard, but it rarely captures everything that makes the standard work in practice: why one source matters more than another, when an exception should be escalated, which matter types need different validation, and where prior matters exposed gaps in the process. Building a useful skill means pulling that practical judgment into the rulebook, not just converting the document into instructions. That takes interviews, debriefs, and time with the people who have actually run the process. Treat it like data entry and you will preserve the words while losing the judgment. 

Skills rot too 

Skills still need owners. Clients reorganize, employees move from email to Teams, regulators change focus, and last year’s skill starts producing last year’s answers. That is the risk. A stale playbook is easier to catch because someone has to read it. A stale skill can keep running. The answer is not to slow down adoption; it is to assign ownership. Each skill needs a review cadence, a testing protocol, and a clear path for updating the rulebook when the facts change. 

What stays human is the judgment that cannot be generalized: who the real custodians are, what the legal theory needs, when this matter does not fit the pattern, and which checkpoints require a person’s call. A skill can know the client and the matter type. It cannot know the matter without people. But it can carry the client’s proven process forward, so our teams spend less time rebuilding the playbook and more time helping exercise judgment where it matters. That is the real promise of agents in discovery: not replacing legal judgment, but scaling the process that lets it show up earlier, more consistently, and with a better record.