# How do I draft a legal discovery request with AI in 2026?

legalpdf.io · August 21, 2026

> Drafting a discovery request with AI in 2026 is a two-stage process: first, you use the AI to build a structured discovery plan (issues, custodians...

Drafting a discovery request with AI in 2026 is a two-stage process: first, you use the AI to build a structured discovery plan (issues, custodians, date ranges, request categories), and second, you use it to generate and refine the actual document — interrogatories, requests for production, and requests for admission — before a human attorney performs a mandatory line-by-line review. AI can compress what used to take 10 to 20 hours of drafting into 2 to 4 hours, but it cannot be trusted to produce a filing-ready document on its own. Courts in 2026 have made clear that attorneys who sign AI-assisted filings own every word of them, and discovery requests are no exception. Below is the definitive, step-by-step method, the tools worth comparing, the mistakes that get draft requests quashed or objected into oblivion, and the cost picture as of August 2026.

## What AI Can and Cannot Do in Discovery Drafting

**Also worth reading:** [What should I do when a discovery request is excessively broad and hard to respond to?](https://legalpdf.io/knowledge/what_should_i_do_when_a_discovery_request_is_excessively_broad_and_hard_to_respond_to.php) · [How do legal teams implement defensible AI discovery protocols in 2026?](https://legalpdf.io/knowledge/how_do_legal_teams_implement_defensible_ai_discovery_protocols_in_2026.php) · [What is the process for filing a complaint asking for discovery in a legal case?](https://legalpdf.io/knowledge/what_is_the_process_for_filing_a_complaint_asking_for_discovery_in_a_legal_case.php)

Generative AI is genuinely useful at three discovery tasks. First, it can map the factual issues in a complaint and answer to the categories of evidence that would prove or disprove each issue, which is the intellectual core of any discovery plan. Second, it can convert a rough issue list into properly formatted requests that track the conventions of your jurisdiction — numbered requests, defined terms, instructions, and signature blocks. Third, it can stress-test your own draft, flagging requests that are likely to draw boilerplate objections, that are overbroad under Rule 26(b)(1) proportionality standards, or that duplicate one another.

What AI cannot do is exercise legal judgment about your specific case. It does not know the judge's standing orders, the local rules on discovery limits, the opposing counsel's litigation history, or which custodians actually hold relevant material. It also hallucinates: AI models can fabricate case citations, misstate rule numbers, and invent procedural requirements with complete confidence. Thomson Reuters and Harvey have both published guidance in 2025 and 2026 emphasizing that AI-assisted discovery workflows succeed only when a licensed attorney validates every citation, every legal standard, and every factual assertion before the document leaves the office. Treat the AI output as a very fast first-year associate whose work product must be checked — not as a co-signer.

## Step 1: Build the Discovery Plan Before Touching the Draft

The single biggest predictor of a good AI-assisted discovery request is the quality of the prompt inputs, and the best input is a written discovery plan. Before generating anything, document the following: the causes of action and affirmative defenses at issue; the elements of each claim that require proof; the likely custodians on the opposing side (by role, not just name); the relevant date range, which for most commercial disputes runs from at least two years before the first disputed event through the present; and the categories of documents you expect to exist (emails, chat messages, CRM records, financial systems, text messages, cloud drives).

Feed this plan into the AI and ask it to produce an issue-to-evidence matrix: for each element of each claim, what documents or testimony would prove it, and which request category would capture it. This matrix becomes your outline. Practitioners using agentic AI tools — the category Thomson Reuters described in its 2026 workflow guidance — can automate parts of this mapping, but a disciplined manual outline works nearly as well and costs nothing. Skipping this step and asking the AI to "draft discovery requests for a breach of contract case" is the most common failure mode, because the output will be generic boilerplate that opposing counsel will object to wholesale.

## Step 2: Prompt the AI to Generate the Request Set

With your matrix in hand, prompt the AI in layers rather than all at once. A workable sequence looks like this: first, ask for defined terms and instructions tailored to your case (define "Document," "Communication," "E-SI," the transaction at issue, the relevant time period); second, ask for requests for production organized by your issue categories, aiming for 15 to 30 requests for a mid-size case; third, ask for interrogatories, remembering that under Federal Rule of Civil Procedure 33(a)(1) a party is limited to 25 interrogatories including subparts unless the court permits more; fourth, ask for requests for admission, which are best used narrowly to nail down undisputed facts like contract execution, receipt of notices, or authenticity of key documents.

For each request, instruct the AI to include the scope limitation and the format demand. Under Rule 34(b), a request must describe the categories with reasonable particularity and may specify the form of production — native format with metadata for spreadsheets and databases, TIFF or PDF with load files for email, and so on. Ask the AI to flag any request that seeks privileged categories (attorney-client communications, work product) so you can either carve them out or anticipate a privilege log demand. Finally, require the AI to justify each request in one sentence tied to your issue matrix; if it cannot justify a request, cut it. Requests without a defensible tie to the claims invite proportionality objections under Rule 26(b)(1), which requires that discovery be proportional to the needs of the case considering the amount in controversy, the parties' relative access to information, and the burden versus benefit.

## Step 3: Human Review — the Non-Negotiable Stage

No AI-drafted discovery request should be served without a licensed attorney reading every request against four filters. Filter one is legal accuracy: verify every rule citation, every local rule reference, and any case the AI cited — hallucinated citations have led to sanctions in multiple reported 2024 through 2026 decisions, and Baker Donelson's framework on AI discoverability underscores that courts are increasingly scrutinizing how AI was used in litigation. Filter two is proportionality: cut or narrow any request a court would likely find unduly burdensome relative to the stakes. Filter three is privilege: make sure you are not demanding materials that will trigger a fight you cannot win, and that your own requests do not waive positions you will need later. Filter four is strategy: confirm the request set supports your trial themes rather than simply being exhaustive.

This review typically takes 1 to 3 hours for a 25-request set, which is still a fraction of the 8 to 15 hours a from-scratch draft requires. Document the review. As courts and commentators — including the Law.com analysis of whether opposing parties can see your AI prompts — have noted, your prompts and the AI's outputs may themselves be discoverable in some circumstances, particularly if you feed confidential case material into a consumer AI tool. Use tools with enterprise-grade confidentiality commitments, and never paste privileged material into a tool that trains on user inputs.

## Comparing Your Tool Options in 2026

The market has split into three tiers, and the right choice depends on your volume, budget, and risk tolerance. The table below summarizes the realistic options as of August 2026.

| Feature | General-purpose AI (ChatGPT, Claude, Gemini) | Legal-specific drafting tools (Harvey, CoCounsel, Thomson Reuters, Spellbook) | Full eDiscovery platforms with AI (Relativity, Everlaw, Logikcull) |
| --- | --- | --- | --- |
| Typical cost | $20–$200/month per user | $100–$500+/month per seat, often enterprise contracts | $0.005–$0.03 per document reviewed, or $500–$5,000+/matter |
| Drafting quality | Good prose, generic legal knowledge | Trained on legal corpora, better rule awareness, citation tools | Drafting is secondary; strength is review and production |
| Confidentiality | Varies; consumer tiers may train on inputs | Enterprise agreements, no-training commitments common | Strong; built for litigation data handling |
| Hallucination risk | High for citations | Moderate; still requires verification | Lower for factual data; drafting still needs review |
| Best for | Solo practitioners drafting occasionally | Firms with regular litigation drafting needs | Cases with large ESI volumes needing review plus production |

G2 and Techloy's 2026 roundups of legal AI assistants consistently place Harvey, Thomson Reuters products, and a handful of newer entrants at the top for drafting, while Rev's and Thomson Reuters' eDiscovery guidance points to the platform tier for document-intensive matters. A solo lawyer handling two or three disputes a year can get 80 percent of the value from a general-purpose model plus a strict review protocol. A firm running dozens of matters should invest in a legal-specific tool for the confidentiality terms alone. Nobody should skip the review stage regardless of tier.

## Common Mistakes That Sink AI-Drafted Discovery Requests

The first mistake is overbreadth. AI models default to sweeping language — "all documents relating to" — which opposing counsel will object to as overbroad and not proportional. Narrow every request with date ranges, custodian limits, and subject-matter limits. The second mistake is duplicative requests: AI often generates five variations of the same demand, which inflates your request count and gives the other side a boilerplate-objection narrative. The third is ignoring numeric limits: 25 interrogatories under Rule 33, and any state-court equivalents, are hard ceilings unless you get leave of court. The fourth is format vagueness: failing to specify production form under Rule 34(b) invites the producing party to dump unsearchable PDFs on you. The fifth, and most dangerous, is trusting AI citations without checking them; several 2025 and 2026 court opinions have sanctioned lawyers for fabricated authorities, and a discovery dispute is a terrible place to establish that your filings cannot be trusted. The sixth is confidentiality leakage — pasting privileged strategy into a consumer chatbot, which Law.com reporting suggests could expose those prompts in later discovery about your own AI use.

## When to Use AI and When to Draft Manually

Use AI when the matter is document-heavy, the issues are well-defined, and you have time for review before the discovery deadline — in federal court, initial disclosures under Rule 26(a)(1) come within 14 days of the Rule 26(f) conference, and written discovery typically follows shortly after, so you usually have a 2-to-4-week window. Use AI aggressively for the mechanical work: defined terms, formatting, consistency checks, and objection anticipation. Draft manually, or at least draft the core requests manually, when the case turns on a single dispositive document, when the discovery strategy is unusually aggressive or unusual, or when the judge has standing orders with idiosyncratic requirements the AI will not know. Also consider manual drafting for requests for admission, where precision matters more than volume and a badly worded admission request can be deemed admitted against you if the responding party fails to respond — a double-edged sword that cuts both ways.

Timing matters too. Serve your AI-assisted, human-reviewed requests early in the discovery period. Requests served in week 2 of a 6-month discovery window give you time to meet and confer, move to compel if necessary, and still complete depositions. Requests served in the final month accomplish little.

## Cost, ROI, and the Bottom Line

The economics favor AI-assisted drafting in nearly every scenario. A from-scratch 25-request set at an associate's billable rate of $300 to $600 per hour represents $3,000 to $9,000 in internal cost or fees. An AI-assisted workflow — plan, generate, review — runs 2 to 4 hours, or roughly $600 to $2,400 in attorney time, plus $20 to $500 in monthly tool costs. For in-house teams and legal aid contexts, the savings translate into more matters handled with the same headcount. The National Law Review's 2026 predictions collection reflects broad consensus that AI-assisted drafting is now standard practice, with the competitive edge shifting to firms that verify well rather than firms that draft fast.

The bottom line: AI is now the sensible default for the first 70 percent of discovery drafting — the structure, the formatting, the boilerplate, the objection-proofing pass. The final 30 percent — legal judgment, proportionality calls, privilege strategy, and citation verification — remains irreducibly human. Lawyers who internalize that division of labor are producing better discovery requests in a quarter of the time; lawyers who skip the review stage are accumulating the kind of errors that end up in judicial opinions.

## A Note on the Discoverability of Your Own AI Use

One emerging issue deserves attention before you draft anything with AI: your prompts and outputs may themselves become evidence. Commentators at Baker Donelson and Hinchliffe Law have written in 2025 and 2026 about a developing framework for the discoverability of AI materials — whether a party's AI prompts, drafts, and model outputs are work product, and whether the opposing side can demand them. The safest posture is to use AI for drafting assistance on non-privileged framing, keep privileged strategy out of prompts entirely, use enterprise tools with no-training and confidentiality commitments, and document your human review process. If your AI use is challenged, you want to show a supervised, attorney-driven workflow rather than an unsupervised copy-paste operation. This area of law is still forming, so check the current state of authority in your jurisdiction before relying on any assumption that your prompts are protected.

## Quick answers

### Can opposing counsel see my AI prompts in discovery?

Possibly. Courts and commentators in 2025–2026 are still developing the framework for whether AI prompts and outputs are protected work product. Prompts containing privileged strategy are at higher risk of being ordered produced, so keep confidential material out of prompts and use enterprise tools with confidentiality commitments.

### How many interrogatories can I serve in federal court?

Under Federal Rule of Civil Procedure 33(a)(1), a party may serve no more than 25 written interrogatories, including discrete subparts, on any other party, unless the court permits more or the parties stipulate otherwise. State courts have varying limits, so check your local rules.

### Will AI hallucinate case citations in my discovery requests?

Yes, generative AI can and does fabricate case citations, rule numbers, and procedural requirements. Every citation in an AI-drafted request must be independently verified against an authoritative source before service, as courts have sanctioned lawyers for filing AI-fabricated authorities.

### How much does it cost to draft discovery with AI?

Tool costs range from about $20–$200/month for general-purpose AI to $100–$500+/month per seat for legal-specific tools like Harvey or CoCounsel, with eDiscovery platforms charging per document or per matter. Attorney time drops from roughly 8–15 hours to 2–4 hours per request set.

### Do I still need a lawyer to review AI-drafted discovery requests?

Absolutely. AI cannot assess proportionality under Rule 26(b)(1), anticipate the judge's preferences, or make privilege strategy calls. A licensed attorney must review every request for legal accuracy, scope, and strategy before service, and that review typically takes 1–3 hours.

Canonical: https://legalpdf.io/knowledge/how_do_i_draft_a_legal_discovery_request_with_ai_in_2026.php
Markdown: https://legalpdf.io/knowledge/how_do_i_draft_a_legal_discovery_request_with_ai_in_2026.php/index.md