
A Second Request review budget can look reasonable until the team works out how much attorney time is needed to validate the results.
That is one of the practical concerns raised by recently reported Federal Trade Commission instructions governing technology-assisted review (TAR) and AI-assisted review. The instructions reach well beyond which technology a party intends to use. They address when training can begin, who can train the model, what performance must be demonstrated, and which materials must be shared with the agency.
For corporate legal departments and outside counsel, those conditions belong in the initial planning conversation. They can affect whether a proposed workflow delivers the savings the team expects.
Recent agency developments should be read together, but they should not be treated as a single new rule.
In September 2026, the FTC amended its Rules of Practice to clarify the factors staff may consider when modifying Second Requests. Those factors include production burdens, how information is maintained, the competitive issues involved, and timing agreement negotiations. The amendment provides a clearer framework for discussing modifications. It does not establish the detailed TAR and AI performance requirements.
Separately, instructions issued in an FTC Civil Investigative Demand and a merger Second Request imposed more rigorous conditions on TAR and AI-assisted review. Those reported instructions are the source of the collection, disclosure, and validation requirements discussed below. Counsel should examine the actual instructions in each matter rather than assume identical terms apply across investigations.
The Department of Justice has taken a separate step. Its targeted Second Request process allows participating parties to make priority productions focused on potentially determinative issues under a timing agreement. That process may narrow or resolve an investigation, but full compliance can still be required.
The operational lesson is to build the plan around the agency, the instructions, and the agreement governing the matter.
The reported FTC instructions condition the start of model training and predictive coding or AI processing on completed custodian collections and agency review and confirmation of required disclosures.
That creates a difficult dependency for teams accustomed to overlapping collection, processing, and review. A delayed collection may affect more than the documents from one custodian. It may hold up the start of the classification workflow.
Before committing to a production calendar, counsel and the discovery team should resolve what collection completion means for the matter. How will additional custodians be handled? What happens when a source requires a second collection? How will refresh collections affect training and validation?
Those questions need explicit answers. A workflow that can technically accommodate incoming data is not necessarily authorized to do so under the applicable instructions.
The reported requirements also introduce calendar checkpoints: expanded disclosures at least one week before training or review, an anticipated production calendar within one week after training begins, and an elusion sample provided at least two weeks before compliance. Those dependencies should be visible in the project schedule from the outset.
The reported FTC instructions call for at least 75% recall, measured with a margin of error of plus or minus 3% at a 95% confidence level, and at least 75% precision. They also require disclosure of a responsiveness cutoff score no lower than 0.50.
Recall measures how much of the responsive material the process finds. Precision measures how much of the material identified as responsive actually is responsive.
Meeting both targets can be challenging, particularly when responsive documents make up a small share of the collection. A workflow that finds most responsive documents may also capture substantial nonresponsive material. Improving precision without losing responsive documents can require further training, testing, and investigation of disagreements.
For budgeting purposes, document volume alone is therefore insufficient. Teams also need to understand the likely prevalence of responsive material, the proposed sampling design, and who will perform the required coding.
A tighter statistical requirement may translate into substantial additional subject matter expert work. That effort should be estimated before the client approves a review approach, not discovered after training begins.
These requirements give teams a reason to compare review methods again.
TAR 1.0 workflows commonly rely on a defined population and separate training and control sets. Continuous Active Learning, often called TAR 2.0, supports an iterative approach in which reviewer decisions continually inform document prioritization.
CAL may offer operational advantages where collections evolve or the burden of a TAR 1.0 control-set approach becomes significant. It is not, however, an automatic exemption from agency instructions. Counsel still needs agreement on the proposed workflow and its validation.
Smaller populations deserve a separate calculation. Once setup, expert coding, validation, reporting, and potential remediation are included, a conventional search-and-manual-review approach may be more economical.
The comparison should account for the complete path to production. Faster classification does not necessarily mean a faster or less expensive response.
The reported AI instructions require disclosure of prompts or review protocols, along with relevance scores and rationales for produced documents where available.
That makes prompt development a counsel-level planning issue. Instructions may reflect legal judgments about responsiveness, and counsel should assess potential work-product concerns before agreeing to disclosure terms.
The discovery team should establish version control from the beginning: which instructions were used, which population they were applied to, what changed, and how the resulting classifications were tested.
Validation also needs separation from development. If a team changes its prompts in response to errors found in a test set, repeatedly testing against that same set can make the results appear stronger without establishing performance on unseen documents. Independent testing and preserved sample membership help make the process explainable.
The FTC’s clarified modification framework expressly recognizes production burden as a relevant consideration. Discovery teams can help counsel make that discussion concrete.
Rather than simply describing a requirement as burdensome, the team can show:
The reported instructions also restrict pre-culling without FTC agreement and supplemental responsiveness review of documents classified as responsive. Familiar steps such as search-term reduction or a second-pass responsiveness screen should not be assumed permissible.
Any proposed modification remains a matter for counsel and the agency. The provider’s role is to explain the technical options, quantify the consequences, and support an executable plan.
At Contact Discovery Services, that planning connects collection, analytics, review, validation, and production before the team commits to a method. For a Second Request or CID, an early workflow assessment can help counsel identify where the proposed protocol creates avoidable work and what evidence would support a different approach.
Planning a regulatory response? Talk with Contact Discovery Services about the collection dependencies, validation workload, and review options that should inform your next agency discussion.
This article provides general educational information, not legal advice. Requirements and available modifications depend on the agency, the instructions issued, and the agreements governing the matter.