The AI Paralegal

Before and after: the manual process vs the AI paralegal

Who this is for

You're a managing partner or ops lead at a 5-50 attorney firm. You've watched Allen & Overy roll out Harvey to 3,500+ lawyers and thought "we can't afford an enterprise legal-AI deal but we have the same paralegal-time problem." This guide is for you.

If you're at an Am Law 100 firm with a procurement process, you're probably evaluating Harvey, Spellbook, or EvenUp directly. This guide is the boutique / mid-market alternative — same shape, cost-scaled, your data, your audit trail.

What the AI paralegal actually does

A day in the life of your AI paralegal

Concrete: an associate forwards an NDA to nina@yourfirm.com at 9:14am. Here's what happens over the next 8 minutes — no API calls, no engineer involvement, just the agent doing its job:

1. 9:14: Email arrives at Nina's inbox. The agent wakes (her sandbox auto-resumes).
2. 9:14: Reads the NDA from the email attachment. Classifies it: mutual NDA, software vendor.
3. 9:15: Pulls up the firm's standard mutual NDA from the document store.
4. 9:16: Diffs the two. Identifies 4 material differences: jurisdiction (NY vs DE), term (5 years vs 2), survival clause, definition of "Confidential Information."
5. 9:18: Drafts a 1-page memo: "Vendor NDA flagged for partner review. Material differences listed below. Recommended redlines attached."
6. 9:21: Posts to #nda-triage Slack channel: "Memo for Acme NDA ready. Partner: please review the term-length clause specifically — it's longer than our standard."
7. 9:22: Attorney @-mentions Nina in the thread: "Thanks Nina, can you also check if Acme has any pending litigation in DE?"
8. 9:22: Nina reads the thread, runs a court-records search, posts back: "No pending DE litigation in past 5 years."
9. 9:23: Sandbox auto-stops (no more events to process).

Total elapsed: 9 minutes. The associate would have spent 45-60 minutes on this. The partner reviews a one-page memo instead of reading the whole NDA.

Scenario 2: Research assignment (litigation)

A partner is preparing for summary judgment. They type in #litigation-research: "Nina — can you find cases where courts in the SDNY have granted summary judgment on statute of limitations grounds in breach of contract cases from the past 5 years? I need cases with strong language on the accrual date issue."

1. 10:02: Nina reads the Slack message, identifies: jurisdiction (SDNY), issue (SoL + accrual), time frame (5 years), procedural posture (summary judgment granted).
2. 10:03: Runs Westlaw search with constructed query. Retrieves 34 potentially relevant cases.
3. 10:08: Filters to 8 most directly on point. Reads each. Extracts the language on accrual dates. Cross-verifies citations.
4. 10:22: Drafts a 3-page research memo: 8 cases with brief summaries, key language in quotation blocks, Bluebook citations verified.
5. 10:22: Posts to Slack: "Research memo ready — 8 SDNY cases with strong accrual-date language. 3 particularly strong for your position. See the memo attached."

Total elapsed: 20 minutes. The partner would have spent 3-4 hours, or asked an associate to spend 3-4 hours. The difference: the partner sees the 8 best cases with extracted language, ready to paste into the brief.

Scenario 3: Discovery document review

A commercial litigation case. Opposing counsel produced 12,000 documents in a rolling production. The trial team has 6 weeks before the discovery cutoff. At $35/hour for contract reviewers, first-pass review of 12,000 documents is ~$42,000. Or: Nina.

The supervising associate uploads the document batch to the shared review workspace and posts to Nina's channel: "Nina — first-pass review on the Acme production. We're looking for documents responsive to RFPs 4, 7, and 12. Flag anything related to the March 15 board meeting. Mark privilege using our standard privilege log format."

1. Day 1, 9am: Nina starts reading. She processes 800-1,200 documents per hour (text-based; slower for scanned PDFs that need OCR).
2. Day 1, 5pm: Nina posts progress to the channel: "4,600 documents reviewed. 312 responsive to RFP 4, 89 responsive to RFP 7, 47 responsive to RFP 12. 23 privilege-flagged. 14 hot documents flagged for your immediate attention — mostly email chains referencing 'the March 14 call.' See the hot-doc summary."
3. Day 2, 9am: Associate reviews the hot-doc summary over morning coffee. The March 14 emails are significant. She posts: "Good catch on the March 14 chain. Add a search specifically for 'financial projections' in the March 10-20 window."
4. Day 3, 2pm: Nina completes the production. Final report: 12,000 documents reviewed, 748 responsive total, 67 privilege-flagged, 31 hot documents with summaries.

Associate attorney reviews: the privilege log (final determination is human), the 31 hot documents (already summarized), and a 10% random sample of non-responsive classifications. Total attorney time: ~4 hours over 3 days instead of 6 weeks of contract-reviewer management. Cost: $200 platform + ~$180 model inference vs $42,000 contract review.

What goes in Nina's CLAUDE.md

The CLAUDE.md is the job description. Here's a working template for a litigation-focused firm — adapt for your practice:

Adjust for your practice areas, your jurisdictions, your escalation thresholds. The more specific, the less calibration noise in the supervised trial period.

Practice area customization

One AI paralegal CLAUDE.md doesn't fit all practice areas. Here's what changes by specialty:

Multi-practice firms: one Nina or several?

For firms with distinct practice groups: one AI paralegal per group works better than a single generalist. Litigation Nina has deep familiarity with your court e-filing portals and your motion templates. Corporate Nina knows your deal room setup and your standard NDA tier classification. The CLAUDE.md context that makes each effective is incompatible across groups.

The cost: $200/month per AI paralegal. A firm with 3 practice groups runs $600/month for 3 specialized AI paralegals — still a fraction of one human paralegal's cost, and materially more useful than a single generalist AI.

The 30-day supervised trial

This is the most important investment you make. Skip it and you'll pay for it at month 3 when accumulated calibration errors surface all at once.

After the supervised trial: the supervising associate moves to digest supervision — Nina sends a daily summary of what she processed and what she flagged. The associate reviews in 15 minutes, approves or redirects. Partners see only escalations and periodic spot-check reports.

Cost math your CFO will ask about

Compare to a paralegal: $35,000-$60,000/year fully loaded ($2,900-$5,000/month). If the AI handles 60-70% of the volume, you free up most of your paralegal's hours for the high-judgment work — or cap the role at one paralegal supervising 3 AI versions instead of hiring two more humans.

The bigger ROI shift: your senior associate's billable mix. Hours that used to be research-and-citation now become review-and-strategy. Same hours billed, materially higher per-hour-rate.

ROI table: 10-attorney litigation firm

These are conservative estimates assuming only the research-time saving, not the NDA and discovery savings. Full ROI including discovery review and e-filing is higher.

Common mistakes from the field

The firms that struggled in the first 90 days made these mistakes:

Skipping the citation-verification constraint

Removing the "verify every citation against a database" rule from CLAUDE.md because it seemed slow. Result: one fabricated case citation that made it into a brief. The associate caught it in review — but the near-miss scared the partner off the whole deployment for 3 months. The verification constraint is non-negotiable. The performance cost is worth it.

Not assigning a named supervisor

Three partners each assuming another partner was reviewing Nina's output. Nobody was. By month 2, Nina had developed subtle escalation avoidance (she'd hedge instead of escalating clearly). Nobody caught it because nobody was reading the digest. One named supervisor; one daily digest review; non-negotiable.

Starting with the most complex work

A litigation firm deployed Nina for complex multi-party discovery review in week one — before she'd been calibrated on simpler NDA triage. She misclassified 15% of documents as non-responsive. The firm lost confidence and almost canceled. Should have started with NDA triage (well-defined, verifiable, lower stakes), built confidence, then moved to discovery review after calibration.

Using a generic CLAUDE.md

The template CLAUDE.md is a starting point, not a finish line. "This firm specializes in pharmaceutical patent litigation in the Federal Circuit" is the kind of context that determines whether Nina's research is useful or generic. A firm that spent 2 hours customizing CLAUDE.md saw 3× fewer calibration issues in week 1 than a firm that deployed with the template unchanged.

Treating AI ethics guidance as optional

Not consulting ethics counsel before deploying. Two issues that surfaced: (1) a jurisdiction that requires client disclosure of AI use — clients weren't told and a sophisticated client asked; (2) using a non-enterprise model tier that didn't have the confidentiality guarantees the firm needed. Both were fixable but created unnecessary friction. Consult ethics counsel first; it's a 2-hour call, not a project.

Onboarding (30 minutes, one-time)

That's it. Nina's productive. The senior associate watches her output for the first week, then drops to spot-checks.

What Nina doesn't do (the boundaries)

• Strategic case decisions. "Should we settle?" "Is this claim worth pursuing?" "How aggressive should we be on discovery?" These are partner calls informed by judgment, relationship context, and risk tolerance. Nina provides the research and analysis; the attorney decides the strategy.
• Direct client communications. Nina drafts all external correspondence; the attorney signs and sends. A client receiving a letter from "Nina, AI Paralegal" without attorney review is an ethics problem in most jurisdictions. Hard rule: Nina's outputs are internal until an attorney releases them externally.
• Court appearances and depositions. Nina prepares deposition outlines, researches the witness, drafts examination questions. The attorney appears, conducts, and exercises judgment in real time.
• Final privilege determinations. Nina classifies documents and flags privilege candidates; an attorney makes the final call on whether a document is privileged. Privilege waiver from incorrect classification is an attorney-responsibility issue.
• Invented citations. If Nina can't verify a case against a connected database, she says so explicitly and stops. She does not summarize from memory, she does not substitute a similar case, she does not present unverified citations as verified. This is the hardcoded constraint that prevents the Levidow-type sanctions.
• Unauthorized matter access. Nina only acts on matters she's been explicitly assigned to by an attorney. She doesn't browse the document store proactively. She doesn't read emails addressed to other attorneys. Assignment-based access only.
• Jurisdiction-specific prohibited conduct. Some states have specific rules about unauthorized practice of law, fee-splitting with non-lawyers, and AI disclosure. These vary; your ethics counsel has the current version for your jurisdiction.

The escalation triggers (configure these in CLAUDE.md)

Nina should automatically escalate — stop work and request attorney input — when she encounters:

• A filing deadline within 48 hours without completed documents
• A contract or matter with value above your defined threshold (e.g., $5M)
• A jurisdictional question outside her configured jurisdictions
• A potential conflict of interest in research results (opposing party also appears to be a firm client)
• A citation she cannot verify after 3 database attempts
• A document marked PRIVILEGED that she wasn't explicitly assigned to review
• Any client-facing action she was not explicitly instructed to take

Escalation is not failure — it's the expected behavior for a well-calibrated AI paralegal. An escalation rate of 5-15% on novel matters in the first 90 days is healthy. Below 2% probably means she's guessing instead of escalating.

Ethics & bar considerations

Most US state bars have published guidance on AI use as of 2026. The common themes:

• Disclosure to clients: Some jurisdictions require disclosure that AI was used in matter preparation. Most don't, but check your specifics.
• Attorney supervision: The attorney remains responsible for AI output. Reviewing the agent's work is not optional — it's a professional obligation.
• Confidentiality: Make sure the model provider's terms of service align with your duty of confidentiality. Anthropic, OpenAI, and Google all offer enterprise tiers with no-training-on-customer-data guarantees. Use those, not consumer tiers.
• Citation duty: The bar comes down hard on fabricated citations (the famous 2023 Levidow sanction). Nina's CLAUDE.md must enforce database-verification of every cite. The platform supports this with a citation-verifier tool.
• Audit trail: Keep audit logs for the standard discovery / malpractice retention period (varies; 6-7 years is common).

This guide isn't legal advice. Talk to your firm's ethics counsel.

State-by-state snapshot (as of May 2026)

This table is a general reference only. Bar guidance evolves rapidly. Always check current rules for your specific jurisdiction.

Pre-deployment ethics compliance checklist

First 90 days: week by week

Firms that succeed with AI paralegals follow a consistent pattern. Here's the week-by-week breakdown based on what works:

Governance: who owns Nina and what that means

The AI paralegal needs the same governance structure as a human paralegal:

• Named supervising attorney. Every matter Nina works on has a named attorney who is responsible for Nina's output. Not "the firm" — a specific human. This is the same supervision structure the bar expects for human paralegals.
• Matter assignment protocol. Nina only works on matters she's been explicitly assigned to. The firm should establish a simple protocol: the supervising attorney (or associate) posts an assignment to Nina's channel or emails her inbox. Nina doesn't proactively pick up work she wasn't assigned.
• Daily digest accountability. The supervising associate reads Nina's daily digest and acknowledges any flagged items. This creates a documented supervision chain: the bar can see that the attorney reviewed AI output regularly.
• Quarterly performance review. Like a human paralegal, Nina should be reviewed quarterly. Review criteria: error rate per task type, escalation rate trend, quality score, scope of tasks (has she taken on new task types that need CLAUDE.md updates?). Read the Performance Reviews guide for the full framework.

How this compares to Harvey, Spellbook, EvenUp

If you're at an Am Law 100 firm: probably Harvey. If you're a 5-50 attorney shop: building on Sandbox Platform is cheaper, more customizable, and your data stays in your AWS account (or ours; your choice). The trade-off: less polished out-of-the-box than Harvey, more configurable.

Deeper comparison

The right choice depends on your firm size, practice areas, and how much you want to configure vs how much you want polished-out-of-the-box. Harvey is the answer if you have enterprise procurement processes and need a vendor to manage. Sandbox Platform is the answer if you want to own the system and configure it for your specific practice.

Frequently asked questions

Do I need to be technical to deploy this?

No. The dashboard's Legal Paralegal template + your IT person handling the OAuth connections is enough. Most managing partners set this up with a 30-minute screen-share from our team.

Can the AI paralegal hallucinate a case citation?

It can — which is why the citation-verification constraint is mandatory. Every cite in Nina's output is cross-referenced against an actual case database. Unverifiable cites get flagged with a warning, never silently included. The CLAUDE.md hard rule: "Never include a citation you cannot verify against a connected database. If you can't find it, say so."

What about client confidentiality?

Two layers. (1) Platform: your sandboxes run in isolation; data never leaves your workspace. (2) Model: use Anthropic enterprise tier which contractually guarantees no training on your data. Both layers required; both configurable.

Does my state bar allow AI paralegal use?

Most US state bars have published guidance by 2026. Common themes: attorney supervision required, some jurisdictions require client disclosure, confidentiality rules apply to model providers. Consult your ethics counsel — this guide is not legal advice.

Can I bill AI paralegal time to clients?

Jurisdiction-dependent. Most ethics opinions allow billing for the supervising attorney's review time, not for AI compute costs. Most firms absorb the AI cost ($200/month) and bill freed-up attorney hours at full rate.

Should I have one Nina for the whole firm or one per practice group?

5-15 attorneys: one firm-wide Nina. 15+ attorneys or distinct practice groups: one per group (litigation-Nina, corporate-Nina) with a specialized CLAUDE.md for each. Avoids context collision between practice areas.

What happens when Nina makes a mistake?

The supervising attorney catches it in review. The audit log captures everything Nina did so you can trace back exactly what happened. For systematic mistakes: update the CLAUDE.md with a corrective rule.

How does Nina handle privileged documents?

Nina only acts on documents specifically assigned by an attorney. Privileged content is tagged "PRIVILEGED — ATTORNEY EYES ONLY" in all outputs and never sent to external channels.

Can Nina interact directly with clients?

No — hard rule in CLAUDE.md. Nina drafts external communications; the attorney sends. All client-facing letters are human-signed and human-sent.

How does Nina compare to Harvey?

Harvey is Am Law 100 enterprise (~$2K-5K per attorney per year). Nina on Sandbox Platform is the mid-market alternative: ~$200/month per agent, your data, more configurable, less polished out of the box. For boutique firms under 50 attorneys, Nina's cost-to-capability ratio is better.

What case-law databases does Nina work with?

Westlaw and LexisNexis (via your existing subscription API), CourtListener (free), Casetext, Bloomberg Law. You configure which databases Nina can access during setup.

How long does the supervised trial period take?

30 minutes of setup, then 30 days of supervised operation. Expect to update CLAUDE.md 3-5 times in the first 30 days. After 30 days, move to digest supervision (15 min/day of reading Nina's summary).

Can Nina handle discovery document review?

First-pass classification: yes. Nina classifies documents as responsive, privileged, or non-responsive and surfaces hot documents for attorney review. She does not make final privilege determinations. Attorney-review load drops ~50% on large document batches.

What if our firm uses a case management system Nina doesn't integrate with?

If your system has an API, MCP can connect it. Common integrations (Clio, MyCase, NetDocuments, iManage, Relativity) are pre-built. Uncommon systems need a custom MCP connector — about half a day of IT work.

Does Nina work 24/7?

Yes — the sandbox auto-starts on events (email arrives, Slack mention) and auto-stops when idle. A firm with normal business-hours volume typically sees Nina active 2-4 hours per day, costing $8-15/month in compute. You pay only for active time.

What's next