A Florida litigator inherits a wage-and-hour matter with eighty-seven thousand documents on a tight production schedule. The natural move is to upload to a familiar cloud eDiscovery platform, run AI-assisted review, and start coding. Two months in, opposing counsel serves a discovery request asking which third-party AI vendors have touched privileged materials, citing Morgan v. V2X. The litigator opens the vendor's data processing addendum and finds a sub-processor list that includes a generative AI partner never disclosed at engagement.
That is not a hypothetical anyone should want to test in front of a federal magistrate judge. The point of this article is to make sure you never have to. After thirty years of litigation practice and the last eighteen months of federal courts tightening the rules on AI in discovery, these are the five questions I ask every eDiscovery vendor before uploading a single privileged document.
1. Where do my documents physically live, and who has access?
Start with the architecture, not the marketing. Ask the vendor to name the data center, the cloud provider underneath it if any, and every employee, contractor, and sub-processor with access to your data at rest and in flight. If the answer is "our secure cloud," press for specifics. A vendor that cannot identify the physical location of your documents cannot give you a credible privilege answer either.
The follow-up matters more than the headline answer. Does the vendor sub-process through a public cloud provider that has its own AI partnerships? Does the vendor's standard data processing addendum permit them to add new sub-processors without notice? Each layer is a potential compliance issue under Morgan and Jeffries, and each layer is a potential subject for opposing counsel's next discovery request.
2. Does any generative AI process my documents, and if so, on whose infrastructure?
This is the question that separates ordinary eDiscovery from the post-Morgan compliance posture courts are now demanding. Many cloud platforms now offer AI-assisted review, predictive coding, privilege detection, or document summarization. Each of those features needs to be examined for one question: does the model run on the vendor's own controlled infrastructure, or does the vendor pass your documents through an external generative AI API?
If the answer involves an external API, you need the vendor's contractual relationship with that provider in writing. Does the provider retain inputs for training? Are there model fine-tuning loops that touch your data? Is there a deletion path that the provider, not the vendor, will certify? The answers to these questions are increasingly required by protective orders, and a vendor who has not built the disclosure architecture cannot give you what the order requires.
3. What does your standard contract say about training, retention, and deletion?
Read the data processing addendum, the master services agreement, and the privacy policy together. The relevant clauses are usually scattered across all three, and the language that controls is rarely the language in the marketing brochure. Look for the explicit prohibition on training. Look for the retention period and what triggers it. Look for the deletion certification, including the form of certification and the time frame.
A common pattern I see: a marketing page says "your data is private," the master services agreement is silent on training, the addendum permits aggregated analytics, and the privacy policy reserves the right to use de-identified data for service improvement. Each clause may be defensible alone, but the combined effect is a contractual environment that no privileged production should sit inside without specific carve-outs. The cost of negotiating those carve-outs is real; the cost of not negotiating them shows up later as a sanctions motion.
4. Can you produce an affidavit that satisfies the protective order in this case?
This is the practical test that exposes most vendor problems before they become privilege problems. Send the vendor the protective order or the proposed protective order and ask, in writing, whether they can sign a sworn affidavit confirming compliance with each material restriction. Pay attention to which restrictions they push back on, which restrictions they want to qualify, and which restrictions they ask to negotiate in side agreements that will not appear in court.
A vendor that cannot affirm the order in plain language is telling you the platform was not built to that standard. That does not mean you cannot use them for non-privileged work. It does mean you should not use them for the privileged production at the center of the case. Save the affidavit step for before the upload, not after the discovery request lands. The affidavit also creates a paper record you can hand the court if the issue is ever raised.
5. If I have to migrate mid-case, what does the exit look like?
Vendor lock-in is a privilege problem dressed up as an operations problem. If the protective order changes, if opposing counsel pushes for new restrictions, or if the vendor changes its own terms mid-engagement, you need an exit path that does not leave your data sitting on infrastructure you no longer control. Ask the vendor specifically: what is the export format, what is the time frame, what is the cost, and how do they certify deletion of every copy after migration?
Three patterns to watch for. First, export formats that strip metadata or coding work, forcing you to redo months of review on the new platform. Second, deletion certifications that exclude backup systems, leaving an indeterminate amount of your data in the vendor's archives. Third, contractual hold periods that let the vendor retain your data for thirty, sixty, or ninety days after termination "for system integrity." Each of these patterns is now a discoverable fact under the post-Morgan standard, and each one is fixable in the engagement contract if you raise it before you sign.
What the post-Morgan landscape actually requires
The federal decisions in Morgan v. V2X and Jeffries v. Harcros have not banned cloud eDiscovery, and they have not banned AI in document review. What they have done is establish a baseline of disclosure, consent, and architectural transparency that most vendors were not engineered against. The compliance gap is not theoretical. It shows up in three places: the data processing addendum, the sub-processor list, and the protective order affidavit. The five questions above are designed to surface each gap before the engagement starts.
For a deeper analysis of the two cases and what the rulings do and do not require, see the firm's published commentary on Morgan v. V2X and Jeffries v. Harcros. For the firm's own on-premises eDiscovery service, which was designed against this compliance standard from the start, see /ediscovery.
When to bring outside help
Solo and small-firm litigators are bearing the brunt of the post-Morgan compliance work. The vendor questions are not difficult, but they are unfamiliar, and the consequences of skipping them concentrate on the lawyer with the smallest support structure. If you are scoping a matter with material privileged content, a brief outside consult before vendor selection is one of the highest-leverage hours you can spend. The five questions take about ninety minutes to run through with a competent vendor representative, and the answers usually decide the matter before the protective order is finalized.
JD Woods Law PLC handles on-premises eDiscovery for litigation, regulatory, and internal-investigation matters at a published flat per-document rate, all-in. The compliance posture is documented in advance, the affidavits are pre-built against the standard restrictions, and migration from another vendor is a defined step in the intake. If the five questions above produced answers you cannot live with, that may be the right time to consider an alternative.