AI Deposition Prep 2026: Everlaw vs Relativity aiR

If your litigation budget has a line item for document review, 2026 is the year it changes shape. Everlaw and Relativity both shipped agentic deposition and testimony tooling into general availability this summer, and the pitch is no longer “we find the hot docs faster” — it’s “we draft the exam outline, build the impeachment chart, and cite every line back to the record.” For business owners funding a case, that’s the difference between paying an associate for 40 hours of transcript wrangling and paying for 6 hours of review on a machine-built draft. But AI deposition prep software is landing as judges tighten AI-disclosure standing orders, so picking a platform now means picking one you can defend on the record.

What’s actually new in AI deposition prep software

The 2025 generation of legal AI was retrieval plus summarization. You asked a question, the system found responsive documents and wrote you a paragraph. The 2026 generation is agentic: you give it a witness, a date range, and a theory of the case, and it runs a multi-step plan across the corpus. It pulls every document the witness authored or received, cross-references prior testimony, flags contradictions, and emits a structured outline with exhibit numbers attached to each line of questioning.

Everlaw’s assistant now operates over what it calls the case record rather than a single review set, so transcripts, productions, and your own work product live in one queryable index. Ask it for an impeachment chart on a witness and it returns a table: the statement, the date, the contradicting document, the Bates number. Relativity’s aiR line split into distinct products — aiR for Review for responsiveness and privilege coding, aiR for Privilege for log generation, and newer testimony-oriented capabilities that reason across depositions. The architectural bet differs: Everlaw builds one assistant that spans the matter; Relativity builds specialized models per task and lets you compose them.

Both vendors made the same change under the hood, and it matters most for defensibility. Every assertion now carries a citation to a source document with a retrievable identifier. Relativity surfaces a confidence signal and the specific passage the model relied on for each coding decision. Everlaw links claims to the underlying document with a click-through. Neither system hands you an unsourced conclusion the way a general-purpose chatbot will, which is why general-purpose chatbots keep generating sanctions orders and these tools mostly don’t.

Why it matters

• Associate hours move from production to verification. The cost curve on a mid-size case bends when first-pass deposition outlines take an afternoon instead of a week — but only if you restructure the staffing model rather than letting the saved hours quietly refill.
• Citation grounding is now the buying criterion. In legal AI document review, a tool that can’t point at the source document is a liability. Ask every vendor to show you the citation trail on a live matter, not a demo corpus.
• Court disclosure orders are converging. A growing number of judges require certification that AI-assisted filings were human-verified. Platforms with per-assertion audit logs make that certification a report you export; platforms without one make it an affidavit you sweat over.
• Per-matter pricing beats per-seat for episodic litigants. If you’re a business owner with two active matters, not a firm with two hundred, unit economics decide this. Consumption pricing on AI e-discovery 2026 platforms typically runs per gigabyte ingested plus per AI operation.
• Privilege review is the highest-value automation. Privilege logs are the most expensive line item per page in most productions, and both vendors now target it directly.
• Vendor lock-in is real and load-bearing. Your AI work product — coding decisions, outlines, chronologies — lives in the platform’s format. Ask about export before you sign, not at the end of the matter.

How to use AI deposition prep software today

1. Run a bake-off on a closed matter. Take a case you’ve already tried or settled, load a defined document set into both platforms, and compare output against what your team actually produced. You know the right answer, so you can measure recall and hallucination honestly. Budget one week and a defined gigabyte cap.

2. Write a house prompt template and make everyone use it. Unstructured prompting drives quality variance. Standardize:

   ROLE: Litigation associate preparing a deposition outline.
   MATTER: [matter name] — [one-line theory of the case]
   WITNESS: [name], [title], [employer], [relevant date range]
   
   TASK: Build a deposition outline organized by topic.
   For each topic:
     1. State the fact we need admitted.
     2. List 3-6 foundation questions leading to it.
     3. Attach every supporting document (Bates + one-line description).
     4. Flag any document that CONTRADICTS the fact.
   
   CONSTRAINTS:
   - Cite a document identifier for every factual assertion.
   - If the record does not support an assertion, write
     "NOT IN RECORD" instead of inferring it.
   - Do not characterize testimony you have not been given.
   - Output as a table: Topic | Question | Bates | Contradiction

3. Build the impeachment pass as a separate run. Do not mix outline drafting with contradiction hunting — the models perform better with a narrow objective:

   TASK: Impeachment analysis only.
   INPUT: Deposition transcript of [witness], [date].
   For every statement in the transcript, search the case record
   for documents that contradict, qualify, or postdate it.
   
   OUTPUT TABLE:
   Transcript cite (page:line) | Statement | Contradicting doc (Bates)
   | Nature of conflict | Confidence
   
   Rules:
   - Only report conflicts you can cite. No inference chains.
   - Rank by materiality to the theory of the case.
   - Mark anything ambiguous as REVIEW REQUIRED.

4. Automate the export, not the judgment. Both platforms expose REST APIs. Pull AI output into your own tracking so review status stays auditable outside the vendor:

   curl -X POST "https://api.example-vendor.com/v1/matters/$MATTER_ID/ai/outlines" \
     -H "Authorization: Bearer $LEGAL_AI_TOKEN" \
     -H "Content-Type: application/json" \
     -d '{
           "witness": "J. Alvarez",
           "date_range": ["2023-01-01", "2024-06-30"],
           "include_citations": true,
           "confidence_threshold": 0.75
         }'

   Then persist every run with its prompt, model version, and reviewer:

   {
     "run_id": "dep-alvarez-2026-07-14-001",
     "matter": "Acme v. Northwind",
     "prompt_template": "house-dep-outline-v3",
     "model_version": "vendor-reported-string",
     "assertions_returned": 214,
     "assertions_verified_by_human": 214,
     "assertions_rejected": 11,
     "reviewer": "attorney-of-record",
     "reviewed_at": "2026-07-15T16:20:00Z"
   }

5. Adopt a two-attorney verification rule. Nothing generated goes into a filing or an exam binder until a human has opened the cited document and confirmed it says what the model claims. Log it. When a standing order asks how you verified, the answer is a file, not a memory.

6. Cap and monitor spend per matter. Consumption pricing punishes exploratory prompting. Set a monthly ceiling per matter, review the operation log weekly, and identify which prompts burn the most for the least return.

How it compares

Dimension	Everlaw AI assistant	Relativity aiR	General-purpose LLM
Architecture	One assistant across the full case record	Specialized products (Review, Privilege, testimony) you compose	Single general model, no case index
Citation grounding	Click-through to source document	Per-decision rationale plus confidence signal	None reliable; hallucinated citations documented in sanctions orders
Deposition outline generation	Native, transcript-aware	Native via testimony capabilities	Manual paste, no exhibit linkage
Privilege log automation	Assisted	Dedicated aiR for Privilege product	Not defensible
Best fit	Teams wanting one integrated workspace and a shorter learning curve	Firms with existing RelativityOne infrastructure and heavy volume	Drafting non-record prose only
Pricing model	Platform subscription plus data/AI consumption	Consumption per GB plus per AI operation, often via a partner	Per token, no legal hosting guarantees
Audit trail for court disclosure	Exportable activity history	Detailed per-document decision logging	Effectively none

If your matter already lives in RelativityOne, the switching cost usually decides it, and aiR’s per-document logging is the stronger artifact when a judge asks how a privilege call was made. If you’re starting fresh, or you’re a business owner coordinating outside counsel across a couple of matters, Everlaw’s single-workspace model stands up faster and a non-specialist can audit it. Both beat the third column so decisively that the real comparison is between them and doing nothing.

What’s next

Watch the disclosure orders more closely than the product roadmaps. The federal bench has not converged on a single rule, and the working requirement in most courtrooms is still “a human verified this.” The platforms that win the next 18 months will make verification an exportable artifact rather than a claim. Expect both vendors to ship court-ready audit reports as a first-class feature, because clients will start demanding them in engagement letters.

On capability, the near-term frontier is cross-matter reasoning and true multimodal ingestion. Most litigation AI workflow tooling still treats each matter as an island; the obvious next step is a witness who appears in three of your cases and a system that knows it. Multimodal matters more than it sounds. A meaningful share of the modern record is screenshots, chat exports, and video, and the tool that reasons over a Slack thread and a Zoom recording with the same fidelity as a PDF will take share quickly.

Pricing is the variable most likely to move against you. Consumption models are generous during land-grab phases and rarely stay that way. If you’re negotiating now, push for a multi-year rate lock on AI operations, a defined export format for your work product, and contractual clarity that your matter data is not used for model training. Those three terms are worth more than any feature on the current roadmap.

Frequently Asked Questions

Is AI deposition prep software defensible in court?

The tool is not the issue; unverified output is. Courts have sanctioned lawyers for filing fabricated citations from general-purpose chatbots, not for using e-discovery platforms with citation grounding. Use a system that cites the record, verify every assertion against the source document, log the verification, and check your judge’s standing order before filing.

How much does it cost per matter?

Ask for a modeled quote on a real document volume rather than a list price. Both platforms combine a subscription or hosting component with consumption charges tied to data volume and AI operations, so a 50 GB matter with heavy AI use looks nothing like a 5 GB matter with light use. Get the quote in writing with a stated overage rate.

Does Everlaw AI assistant or Relativity aiR for Review have better accuracy?

Vendor benchmarks favor whoever published them, and neither generalizes to your document set. Run the closed-matter bake-off described above. A week of testing on documents where you know the correct answer produces a more useful number than any published study.

Can a small business use these without a big firm?

Access usually runs through outside counsel or a service provider rather than direct licensing, which is fine. What matters is that you ask your counsel which platform they use, whether they’ll pass through AI efficiency gains in the billing arrangement, and how they document verification. That last question tells you a great deal about the firm.

Will this replace contract attorneys doing document review?

It replaces the first pass, not the judgment. Volume review work is genuinely shrinking, but demand is shifting toward people who can audit AI coding decisions and defend them — a smaller, more skilled, better-paid layer. Plan staffing for verification capacity, not headcount.

What about client confidentiality and model training?

Both enterprise platforms contract against using client data for training and run in controlled hosting environments, but confirm it in your specific agreement, including subprocessors and retention periods. The real exposure is not the vetted platform; it’s an associate pasting a privileged transcript into a consumer chatbot. Write the policy, then enforce it.