Deposition Review Is Your Most Expensive Black Box

Banner artwork by / AI

Most legal departments can tell you what they paid outside counsel. Fewer can tell you what happened between the deposition and the motion. That gap is more than an operational nuisance; it is one of the least visible and least governed parts of litigation. Once the transcript leaves the court reporter, the review process often becomes invisible to the people who are ultimately paying for it. In-house teams may see the invoice; they may even see the final work product. But what they often do not see is how much time was spent reviewing the testimony, what the reviewing lawyer prioritized, or whether the transcript was analyzed in a way that was consistent, efficient, and defensible. 

That is a problem for legal ops, but it is also a problem for litigation quality. A deposition review process that cannot be seen, measured, or repeated is not just inefficient. It is a governance issue. AI creates an opportunity to make that process visible, but it also creates a risk: legal teams can replace an opaque human workflow with an opaque technological one. The standard should be whether AI-assisted review makes the work more traceable, accountable, and measurable. 

Deposition review is a hidden cost center 

A deposition is a factual architecture. Buried across hundreds of pages are the admissions, contradictions, omissions, and credibility markers that drive motion practice, settlement posture, and trial strategy. Yet, the traditional review process is still largely manual. An associate reads page after page, highlights what seems important, drafts a summary, and then hands that work up the chain. In a complex case with multiple deponents, the process can consume days of billed time yet produce outputs that vary widely from lawyer to lawyer. 

That variability is difficult for the client to see. Two capable lawyers may read the same transcript and identify different facts, assign different significance to the same testimony, or make different connections across the record. The problem is not necessarily diligence. Legal analysis is contextual. A sentence can matter differently depending on who said it, when it was said, how it fits with another witness’s account, and whether it can survive a challenge on the record. That is why any AI system used in litigation must be treated less like a novelty and more like part of the evidentiary workflow. It should accelerate the lawyer’s access to the record, not obscure it. 

We have spent much of our careers in settings where resources were limited and the consequences of missing something important were real. There is no luxury of endless review time. You learn quickly that the value is not in reading more pages for the sake of it. The value is in finding the precise testimony that changes the shape of the case. But even capable and experienced lawyers do that unevenly.  

That is the hidden cost of deposition review. The client is not simply purchasing the hours required to read a transcript or prepare a summary. The client is accepting an unmeasured risk that a material admission, inconsistency, or factual connection will not be identified. That risk does not appear as a separate line item on the invoice, and it may not become visible until the case team discovers the omission later, after strategy has already formed around an incomplete understanding of the record. 

This is where AI can add practical value. The question is not whether technology can replace a lawyer’s judgment. It cannot. The better question is whether it can more consistently surface the facts, patterns, and contradictions that lawyer judgment needs to work. The AI should identify candidate findings. The lawyer should determine their legal and strategic significance. Used that way, AI does not eliminate the need for contextual review. It gives the lawyer a more systematic way to navigate the record and reduces the likelihood that an important connection depends entirely on one reviewer noticing it while moving through hundreds of pages. Technology should improve access to the evidence, not obscure it. And it should make the review process more visible to the client, not replace one black box with another. 

For in-house legal teams, that variability is a problem. If you are responsible for outside counsel oversight, vendor management, and cost control, you need more than a bill. You need visibility. You need a way to understand how the work is being done, not just what it costs. 

AI is useful only if it is verifiable 

AI only adds value when the lawyer can verify what the system found and why it matters. Its value should also be measured by whether it can be trusted inside a professional workflow. In litigation, trust has a specific meaning. It means the output can be traced back to the record. If a transcript tool says a witness contradicted herself, the lawyer should be able to see exactly where that contradiction came from. That means a summary is not a black box. It means the technology is helping the lawyer review the evidence, not silently substituting for the lawyer’s responsibility to do so. 

That matters in litigation because the cost of error is asymmetric. A summary that is wrong wastes time and can send a case strategy in the wrong direction. A contradiction with no citation behind it is worse than incomplete because a lawyer cannot defend what they cannot trace. So, the right test for AI here is not how much time it saves. It is whether it strengthens the judgment of the person using it. Every material conclusion should be citable, every comparison should be tied to the source transcript, and every summary should preserve the path back to the underlying testimony. That is more than a feature — it is the baseline. 

Govern the workflow, not merely the ool 

If a team is going to deploy AI for transcript review, someone has to define the rules before the output becomes part of case strategy. A general disclosure that AI was used is not enough. The client should understand what the technology did, what the lawyer did, and how the final work product was validated. That responsibility naturally belongs with legal operations. Legal operations is not simply a cost-control function. It is a governance function, particularly when technology changes how substantive legal work is performed, reviewed, and billed. The goal is not to dictate how lawyers analyze a case. It is to ensure that the process supporting that analysis is traceable, accountable, and measurable. 

The work should be traceable: Every material conclusion should preserve a path back to the source transcript. If an AI-assisted review identifies an admission, contradiction, omission, or credibility issue, the reviewing lawyer should be able to see the relevant page-and-line citations and evaluate the testimony in context. 

A material conclusion is any assertion expected to influence case theory, motion practice, settlement valuation, witness strategy, credibility assessment, or a recommendation to the client. For conclusions at that level, a general summary is not enough. The record itself must remain accessible. The workflow should also define what happens when an output is unsupported, ambiguous, or inconsistent with other testimony. Those results should not quietly pass into the final work product. They should be flagged for additional review, corrected, or excluded. A conclusion that cannot be traced to the record should not be relied upon. 

The work should be accountable: “Human in the loop” is not a meaningful control unless the human role is defined. The workflow should identify who reviews the AI-generated output, what that person checks against the transcript, when the review occurs, and who is responsible for final sign-off. 

The AI may identify candidate findings. A lawyer must determine whether those findings are accurate, legally relevant, and strategically significant. The level of review should also reflect the consequence of an error. An AI-generated topic index may require a different validation process than a contradiction that will shape a dispositive motion or be used to impeach a witness. The more consequential the conclusion, the more rigorous the attorney review should be. Accountability should remain with a named lawyer, not with the technology provider, the litigation support team, or an undefined group of reviewers. The tool can assist with the work. It cannot assume professional responsibility for the result. 

The work should be measurable: If AI changes how deposition review is performed, the effect should be visible in the matter budget, staffing model, and invoice. Reduced review time should not disappear into a general outside counsel budget. Nor should a technology fee be added without showing whether it replaced attorney hours, reduced leverage, accelerated the work, or improved the resulting analysis. 

Before adopting an AI-assisted workflow, the legal department should establish a baseline for the traditional process. Useful measures could include: 

No single measure will establish success. A faster first draft is not an efficiency gain if it requires substantial attorney rework. Likewise, a less expensive review is not an improvement if it misses testimony that later changes the case strategy. The relevant question is whether the workflow produces a more reliable result with a justifiable use of time and resources. Outside counsel guidelines can translate those expectations into concrete requirements that address the following:  

The purpose of these requirements is not to prohibit innovation or prescribe one method of transcript review. It is to ensure that the client and counsel share the same understanding of how the work will be performed, verified, and billed. 

The payoff is visibility. If deposition review takes less time with AI, that time should not vanish into the general budget of outside counsel. It should show up as a cost signal that reshapes billing expectations, vendor requirements, and internal policy. 

Deposition review has been a black box for as long as it has been billable. Work goes in, a summary comes out, and the people paying for it rarely get to see what happened in between. AI does not automatically change that, but it can. A workflow built on citations and human checkpoints leaves a trail. It lets a legal team show how the testimony was analyzed, how long it took, and what a person actually verified. Teams that build that trail are managing their litigation. Teams that cannot are still just paying for it. That is the standard the profession should expect.