Do I need a paid CoCounsel, Harvey, or Lexis+ AI account to do this right?

Yes. The free or consumer tier of any AI tool is structurally incapable of being privilege-safe because the standard consumer terms permit the vendor to use inputs for model training and offer no tenant isolation. The serious options for mid-size firms are CoCounsel (Thomson Reuters), Lexis+ AI (LexisNexis), Westlaw Precision AI, and Harvey, all of which sell enterprise tiers with Data Processing Addendums, training data exclusion, and audit logs. Pricing varies by seat count, but mid-size firms typically land in the four to twelve thousand per attorney per year range. Some firms also build internal tools on the OpenAI Enterprise API or Anthropic's enterprise tier, which is a heavier lift but gives more control over the data architecture. Free tools have a place for non-client work like CLE prep or business development, but never for matter work.

Is this setup actually privilege-safe under New York or California ethics opinions?

Yes when configured correctly. The 2024 New York State Bar Association opinion and the 2024 California State Bar guidance both recognize that AI use in legal practice is permissible under the rules of professional conduct provided the lawyer maintains competence, supervision, confidentiality, and verification. The privilege analysis turns on whether the AI vendor is treated as an agent of the firm under a written agreement that excludes the firm's inputs from training and prevents disclosure to third parties. An enterprise contract with a Data Processing Addendum, training exclusion, and tenant isolation meets that bar. A consumer chat tool with default training-on terms does not. Florida, Illinois, and Texas issued similar opinions in 2024 and 2025. The convergent rule across jurisdictions is the same. Get the contract right, document the verification, and the privilege analysis holds.

Will the output sound generic or like every other AI memo?

Only if the prompts and the corpus are generic. The firms with the strongest output do three things. They feed the AI a curated corpus of the firm's prior research memos, briefs, and internal know-how as a voice and structure reference. They write prompts that specify jurisdiction, procedural posture, standard of review, and the firm's preferred citation format. They review and edit the output before it leaves the associate's desk so the model learns the firm's editing patterns over time. A 50-attorney firm running this setup for six months produces research memos that read like the firm's senior associates wrote them, not like a generic legal AI. The voice problem is solved by inputs, not by switching tools.

How do I share AI research output with attorneys and clients who don't have access to the tool?

Through the firm's document management system. Lexis+ AI integrates with both NetDocuments and iManage. CoCounsel exports to Word and PDF directly. Harvey has DMS connectors for the major systems. The pattern: associate runs the research in the AI tool, exports the memo to NetDocuments under the matter folder, partner reviews and edits inside Word, and the final memo goes to the client through the existing secure file-sharing workflow. Avoid emailing AI output as attachments or sharing through consumer file-sharing tools. That breaks the audit trail and complicates the privilege defense. For client-facing deliverables, the AI tool should never be the delivery surface. The DMS is the system of record.

What if the firm has restrictions on AI tools or the executive committee is skeptical?

The skepticism is reasonable and the path forward is structured. Start with a 90-day pilot scoped to a single practice group with a written firm AI policy, a verification protocol, and an opt-in associate group. Document the malpractice insurance disclosure with the firm's broker before the pilot begins. Run the pilot on internal know-how and CLE work first, then on lower-stakes client matters with explicit partner supervision. Present the pilot results to the executive committee with realization-rate data, not abstract benefits. Most committee skepticism collapses when they see verified output and audit logs alongside billable-hour reductions on a finished matter. Do not roll out AI by stealth. That is how firms end up reading about themselves in the ABA Journal for the wrong reasons.

Can paralegals and litigation support staff use the AI research assistant too?

Yes, with role-based access. Paralegals are typically the highest-volume users of legal research AI in well-run mid-size firms because they handle first-pass research and summary work that would otherwise consume associate time. The setup that works: paralegals get access to the tool with a workflow that requires associate review before any output is filed in the matter folder, associates review and edit before partner sign-off, partners verify any citation that will appear in court papers. Each tier reviews the output of the layer below. CoCounsel, Lexis+ AI, and Harvey all support role-based seat tiers at reduced rates for paralegals and litigation support staff. The audit log shows every query, every output, and every reviewer, which is exactly what supervision under the rules of professional conduct requires.

How does AI research interact with contract drafting? Should both run on the same tool?

Often the same vendor, sometimes different. CoCounsel and Lexis+ AI both handle research and contract review under the same enterprise license. Harvey is increasingly strong on contract drafting and adequate on research. Some firms run two tools: a research-focused product for litigation and a transactional-focused product for corporate. The privilege architecture is the same either way. The same enterprise contract terms, the same DMS integration, the same audit log requirement. The difference is workflow. Research output is consumed by the attorney drafting the brief or memo. Contract drafting output is compared against a playbook and redlined. Both require attorney verification before anything leaves the firm. The decision between one tool or two is mostly about practice mix and seat economics.

Who reviews AI research output before it gets used? And how does that affect partner-vs-associate adoption?

The reviewer hierarchy depends on the output and the matter. For internal research memos, an associate reviews and a partner spot-checks. For citations going into a brief, a paralegal does the cite-check, an associate verifies the substance, and a partner approves the legal positions. For client-facing memos, every layer reviews. Partner adoption tends to lag associate adoption because partners are time-constrained and risk-averse. The pattern that works: associates run the AI tool day-to-day, partners receive the verified output and edit it as they would edit a senior associate's draft, and the firm tracks the verification chain in the audit log. After three to six months, partners notice the time savings on their own matters and adoption accelerates. Forcing partner adoption before the workflow is mature usually backfires. Let the associates prove it first.

How Can a Law Firm Build an AI Research Assistant That's Privilege-Safe?

The managing partner of a 90-attorney firm I worked with last year asked one question that stopped the room: 'If we put a privileged matter into a chatbot, did we just waive privilege?' The IT director did not have a clean answer. The general counsel did not have a clean answer.

This is the question every mid-size firm has to settle before AI legal research becomes part of the workflow. The answer is yes if the tool is wrong, no if the architecture is right. The difference is a contract, a deployment pattern, and a verification protocol that almost no vendor explains in their pitch.

This guide walks through how a 50 to 200 attorney firm builds an AI research assistant that holds up under attorney-client privilege analysis, work product doctrine, the state bar AI opinions, and a malpractice carrier's underwriting questions. It covers the architecture, the contract terms, the DMS integration, and the verification workflow. Read this before your firm signs any AI vendor contract.

Why this matters for mid-size law firms specifically

BigLaw and mid-size firms diverge on this question because of resources. The Am Law 100 has a chief AI officer, privacy counsel, an enterprise architecture team, and a budget for pilots. They can absorb a mistake. A 120-attorney firm with three offices has none of those things. A single privilege issue would carry through the local bar in 48 hours.

What changes when a mid-size firm builds the AI research assistant correctly: associates run circles around the existing research workflow, partners get to strategic questions faster, and the firm starts winning pitches it used to lose to BigLaw on resource arguments. Firms doing this well report 40 to 60 percent reductions in legal research hours. Firms that skip the privilege architecture step end up reading about themselves in bar ethics columns.

What an AI legal research assistant actually does

A properly built AI research assistant for a law firm is three things stitched together: a large language model with strong legal reasoning, a retrieval system grounded in a verified legal corpus, and an enterprise architecture that protects privilege and work product. CoCounsel uses Westlaw as its retrieval source. Lexis+ AI uses the LexisNexis corpus. Westlaw Precision AI is built directly into the Westlaw platform. Harvey can connect to either or run on its own contract corpus.

Three things make a real legal research assistant different from a generic chatbot:

It retrieves from a vetted source, not the open internet. The Mata v. Avianca pattern (hallucinated citations from a consumer chatbot trained on whatever was on the web in 2021) is not the failure mode of a legal-grade tool. The failure mode here is misinterpretation of real cases, not invention of fake ones.
It runs under an enterprise contract that excludes your inputs from model training. This is the privilege firewall. Without it, opposing counsel can argue privilege was waived by disclosure to a third party whose terms permit training on legal documents.
It produces an audit log. Every query, every document, every output, every reviewer, timestamped. This is the supervision documentation that satisfies the rules of professional conduct.

Think of it as a senior associate who has read every reported case, never sleeps, and will produce confidently structured output that requires verification before it gets used. The architecture decides whether that associate is trustworthy.

Before you start

You need:

An enterprise legal AI vendor relationship. CoCounsel, Lexis+ AI, Westlaw Precision AI, or Harvey are the four serious options for mid-size firms. The contract terms matter more than the marketing.
A document management system. NetDocuments and iManage are the two systems that integrate cleanly with all major legal AI vendors. PracticePanther, MyCase, and Clio integrations are catching up but trail on enterprise features.
A written firm AI policy. Two pages, signed by the executive committee, distributed to all attorneys and staff. This is the first artifact your malpractice carrier will ask for.
A pilot practice group willing to run the workflow for 90 days. Litigation, real estate, or commercial transactions all work. Pick a group with a partner who is willing to invest the time.
About 45 minutes for a partner or senior associate to walk through the first end-to-end research workflow.

One thing to settle before any privileged document touches the tool: the privilege architecture. We have a dedicated section below. It is non-negotiable. The architecture decides whether your AI use is privilege-safe or whether you have created a Mata v. Avianca-style problem at scale across the entire firm.

Task 1: First-pass case law research with grounded citations

The failure pattern: a junior associate runs a Westlaw search using natural-language search, picks the first eight or ten cases that look on point, drafts a research memo, and the partner discovers in deposition prep that two cited cases were distinguished or overruled in 2024 and 2025.

What to ask Westlaw Precision AI or Lexis+ AI for instead:

Research the current state of California law on the apex deposition doctrine for senior corporate executives, focused on the Northern District of California and the California Court of Appeals. I need: (1) the leading authority and current standard, (2) the three most recent published opinions applying the doctrine post-2024, (3) any pending appellate review or rule changes in 2025 or 2026, (4) the standard for whether plaintiff has exhausted lower-level discovery before seeking apex deposition, (5) every cited authority flagged with KeyCite or Shepard's treatment. Output as a research memo in the firm's standard format.

The AI returns research with treatment indicators on every citation. Overruled cases get flagged in red. Distinguished cases get flagged in yellow. The associate verifies the top three to five cases manually before relying on them in any brief or memo. This is the workflow that prevents a Mata v. Avianca outcome. The verification is built into the tool. The associate's job is to confirm the verification, not to do it from scratch.

For unfamiliar jurisdictions or specialized areas (admiralty, ITC, FERC, immigration), the verification matters even more because the associate has less internal sense of what looks wrong. The AI catches the obvious issues. The partner catches the subtle ones.

Task 2: Building a firm know-how corpus

The failure pattern: a senior associate in the New York office spends six hours researching an issue another office's senior associate already researched 14 months ago, because the firm has no way to surface prior work product across offices and practice groups.

What to ask CoCounsel or Harvey for instead:

Search the firm's internal know-how corpus for any prior research, memos, briefs, or analyses on California's apex deposition doctrine. Return the three most relevant prior work products, with the matter name, date, drafting attorney, current treatment status (any cases distinguished or overruled since the original work), and a brief summary of the firm's prior position on each. Identify any gaps where the firm has not addressed an issue that's now relevant.

The AI surfaces prior work product faster than a manual search through NetDocuments or iManage. The associate then evaluates whether the prior research is still good (no overruled authorities), whether it covers the current question, and whether it represents the firm's current position. The senior associate's six-hour research becomes a 30-minute review of prior firm work followed by targeted research on whatever gaps remain.

The setup investment: one quarter of IT and knowledge-management work to load the firm's prior briefs, memos, and research into the AI tool's tagged corpus. This work pays back hard because every future research session benefits from it. Firms that skip this step end up paying for AI seats and getting a fraction of the value.

Task 3: Statutory and regulatory analysis with cross-references

The failure pattern: an associate researching a regulatory question reads the statute, cross-references three regulations, missed a 2024 amendment to a related provision, and produces a memo that's substantively wrong on the operative point.

What to ask Lexis+ AI or Westlaw Precision for instead:

Analyze the current text and recent amendment history of California Business & Professions Code Section 17200 (Unfair Competition Law). I need: (1) the current statutory text as of April 2026, (2) any amendments enacted in 2024 or 2025, (3) the cross-referenced regulations and the most recent amendment dates for each, (4) the three most cited California Supreme Court decisions interpreting the statute since 2020, (5) any pending legislation that would amend the statute, (6) any federal preemption questions raised in recent decisions. Output as a structured memo with hyperlinked citations.

The AI runs the cross-referencing and amendment-tracking faster than a human associate. The associate then reviews the output for accuracy and adds the firm's strategic interpretation, which the AI cannot provide. Three hours of associate time becomes 45 minutes of AI time plus 45 minutes of attorney review.

For multi-jurisdictional regulatory questions, the same prompt pattern works at scale. Run the AI on each jurisdiction in parallel, then have the associate compare the outputs and identify the variance. The AI is excellent at parallel retrieval. The judgment about which variance matters stays with the lawyer.

Task 4: Internal memos and client advisories at speed

The failure pattern: a partner asks an associate for a five-page memo on a recent regulatory change, the associate produces it in eight hours, and the memo is substantively fine but follows a different structure than the firm's last three memos on similar topics, which makes it harder for the partner to skim.

What to ask Harvey or CoCounsel for instead:

Draft a client advisory memo on the 2025 amendments to California's Privacy Rights Act, specifically the new requirements for AI-driven decisioning notices effective January 2026. Use the firm's standard client advisory format (uploaded sample). Include: (1) executive summary in three sentences, (2) what changed and what's new, (3) who is affected and at what threshold, (4) the compliance steps clients should take in the next 90 days, (5) the firm's perspective on enforcement priorities. Voice should match the uploaded sample. Output as a Word document.

The AI produces a structured first draft that matches the firm's voice and format. The associate edits substance and adds firm-specific strategic perspective. The partner reviews and signs off. Eight hours of associate time becomes 90 minutes of AI work plus 90 minutes of attorney review.

The critical move: feed the AI three to five examples of the firm's prior client advisories. Without samples, the output reads like a generic LexisNexis newsletter. With samples, the output reads like the firm's senior associates wrote it.

For practice-group newsletters and CLE materials, the same pattern works at lower stakes. The AI handles the structural draft. The senior associate edits for substance and adds the firm's strategic positioning.

Task 5: Brief and motion research support

The failure pattern: a partner working on a summary judgment brief under deadline asks an associate to research the standard for tortious interference with contract under New Jersey law, the associate produces a memo, the partner adapts it into the brief, and a clerk later finds that one cited case was actually a New York case the associate confused with a New Jersey case with similar facts.

What to ask Westlaw Precision AI for instead:

I'm drafting a summary judgment brief opposing defendant's motion in [matter]. Research the New Jersey standard for tortious interference with contract, specifically focused on: (1) the elements of the claim, (2) the recent New Jersey Supreme Court treatment of the 'malice' element post-2022, (3) what level of factual specificity is required at the summary judgment stage, (4) the three most analogous cases on similar fact patterns where summary judgment was denied. For each citation, confirm the jurisdiction is New Jersey, confirm the procedural posture matches summary judgment, and provide the KeyCite treatment indicator. Flag any cited authority that originates outside New Jersey.

The AI returns research with explicit jurisdiction and posture flags. Citations from outside New Jersey are flagged automatically. The associate verifies the top cases manually. The partner reads the verified output and decides which authorities go into the brief.

This prompt pattern catches the most common research failure modes at the prompt level rather than relying on the associate to catch them in review. The AI cannot make a New York case look like a New Jersey case if the prompt requires explicit jurisdiction confirmation.

For appellate work, the same pattern extends to standard of review questions, preservation of error analysis, and the procedural posture verification that appellate courts care about more than trial courts do.

Task 6: Conflict-check and matter-history searches

The failure pattern: a new client engagement requires a conflicts check, the conflicts software runs a name-match query and returns a clean report, and three weeks into the matter the firm discovers an indirect conflict involving a former client of one of the lateral partners that the conflicts software did not catch.

What to ask CoCounsel or Lexis+ AI for instead, scoped to the firm's internal corpus:

Run a comprehensive conflicts analysis for prospective client Acme Logistics. Search the firm's matter database, prior engagement letters, and former-client database for: (1) direct representations of Acme Logistics or any subsidiary or affiliate, (2) representations adverse to Acme Logistics in any prior matter, (3) representations of any entity in the same industry where confidential information may overlap, (4) any prior representation by lateral attorneys before they joined the firm (search the laterals' prior firm disclosures), (5) any representation of opposing counsel's clients on related matters. Output as a structured report with each potential conflict, the basis, and a recommendation for the conflicts partner.

The AI does the comprehensive search faster than the existing conflicts software in most mid-size firms. The output is a flagged list for the conflicts partner to review. Five-minute clearance becomes 30-second clearance with better recall on indirect and lateral-historical conflicts.

The privilege architecture matters here especially. The conflicts data is highly sensitive. The AI tool running this analysis must be on the enterprise tier with full tenant isolation and training exclusion. This is not a use case for any consumer or trial tool.

The privilege-safe prompts that actually work

After watching mid-size firms build research assistants for the better part of two years, the difference between an AI research workflow that protects privilege and one that creates exposure comes down to four prompt moves and one architectural rule.

Specify the legal frame and verification requirement. Jurisdiction, procedural posture, standard of review, and the citation verification expectation. 'Research X under Texas law on summary judgment, with KeyCite indicators on every citation' produces a different output than 'research X.'

Specify the firm voice and format. Upload three to five prior firm research memos as voice and structure samples. Without samples, the output reads like generic legal AI. With samples, it reads like your firm.

Specify what stays inside the tool and what moves to the DMS. Drafts and analytical work product can live inside the AI tool's workspace temporarily. Final output moves to NetDocuments or iManage under the matter folder with proper access controls. The AI tool is not the system of record.

Specify the verification list at the end of every prompt. Tell the AI which citations the human must verify, which factual claims need cross-checking, which conclusions need attorney sign-off. This embeds the verification protocol in the prompt rather than relying on a separate checklist that gets skipped under deadline pressure.

The one architectural rule: every AI session for client matter work runs under the enterprise contract with a Data Processing Addendum, training exclusion, and tenant isolation. No exceptions. No 'just this once' on a personal account.

The privilege and malpractice non-negotiables

This section is short because the rule is simple, but it is the most important section in this guide.

Do not put any of the following into the consumer tier of any AI tool (free ChatGPT, free Claude, Gemini personal, Microsoft Copilot personal, any free chat product):

Privileged attorney-client communications
Attorney work product (mental impressions, legal theories, case strategy)
Client identities tied to matter substance
Witness names, deposition content, or substantive testimony
Settlement positions or negotiation strategy
Sealed pleadings or protective-order materials
Trade secrets disclosed under NDA
Internal firm conflicts data

Use enterprise legal AI tools (CoCounsel, Lexis+ AI, Westlaw Precision AI, Harvey) under signed enterprise agreements that include training data exclusion, tenant isolation, audit logging, and a Data Processing Addendum that names the firm as the data controller. These contract terms are the privilege firewall. Without them, opposing counsel can argue privilege was waived through disclosure to a third party whose terms permit training on legal materials.

The state bar AI opinions are converging on the same rule. The 2024 New York State Bar Association opinion, the 2024 California State Bar guidance, the 2024 Florida ethics advisory, the 2024 Illinois opinion, and the 2025 Texas opinion all hold that AI use is permitted under the rules of professional conduct, but the lawyer remains responsible for competence, supervision, confidentiality, and verification. Mata v. Avianca remains the canonical lesson on what happens when verification fails. Two attorneys, six fictional cases, real sanctions, and a story that carries through every state bar CLE.

The practical workflow that respects these rules: build research prompts and templates inside the AI tool, run all client-matter work through the enterprise-licensed product associated to the correct matter folder, verify every citation with KeyCite or Shepard's before any document leaves the firm, document the verification step in the matter file, and route the final output through the DMS rather than emailing AI tool exports as attachments.

Malpractice insurance carriers as of 2026 require AI disclosure in annual applications. ALPS, ProAssurance, and the major specialty malpractice carriers all have AI riders. Some require attestation that human attorneys verify all AI output. Some require a written firm AI policy. Some impose a small premium adjustment for firms using AI in production work. Call your broker before your next renewal. Get the disclosure right. Underwriters reward firms that show their work.

If your firm has signed a vendor enterprise agreement with a Data Processing Addendum, the rules can be different on permitted use. Ask your general counsel or the firm's risk partner what is covered. Do not assume.

When NOT to use AI for legal research

AI legal research tools are powerful but not universal. They are the wrong answer for:

Anything client-facing without attorney verification. A research memo going to a client, an opinion letter, a regulatory advisory. AI drafts; lawyers verify and sign.
Novel legal questions in unsettled areas. AI is excellent on settled doctrine and weaker on emerging case law where treatment is contested or first-impression issues are pending. For these questions, the AI gives you a starting point, not an answer.
Anything involving sealed pleadings or confidential settlement terms outside the licensed enterprise tool. The audit trail and tenant isolation matter most for these materials.
High-stakes credibility analysis. AI can summarize witness statements but cannot evaluate credibility, intent, or what a witness deliberately did not say. The judgment call stays with the attorney.
Final legal positions under tight scrutiny. Briefs going to the Supreme Court, opinion letters with significant deal exposure, regulatory submissions that will be reviewed by enforcement staff. AI assists; partners decide.

A simple rule: AI is an unfair advantage on the 80 percent of legal research where speed and structure matter. Trust the official channels and human judgment for the 20 percent where the document or decision has career-defining or client-defining weight.

The quick-start template

Here is the prompt scaffold that works across most law firm research workflows. Copy it, fill in the brackets, paste into your enterprise legal AI tool.

Research [legal question] under [jurisdiction] law, focused on [procedural posture or transactional context].

Specific outputs needed: [list 3 to 5 specific things you need: leading authority, recent treatment, statutory text, regulatory cross-references, factual analogs].

Verification: flag every citation with KeyCite or Shepard's treatment. Confirm jurisdiction on every cited case. Flag any case where the procedural posture does not match the question.

Voice and format: use the firm's [research memo / client advisory / brief] format as reflected in the uploaded samples.

Confidentiality: this matter is privileged and runs under the firm's enterprise agreement. Output goes to the matter folder in [NetDocuments / iManage] under matter number [X].

That is the whole pattern. For 80 percent of mid-size firm research work, this is enough. For complex matters, extend the scaffold with matter-specific risk categories and prior firm work product as voice samples.

Bigger wins beyond research

Once a firm has the privilege-safe research workflow running, the next layer of value shows up in places that are not single research memos.

Practice-group knowledge management. Build structured AI prompts and corpora for each practice group. Litigation has its corpus and its prompt library. Corporate has its corpus and its prompt library. Real estate has its corpus and its prompt library. Each playbook codifies the firm's preferred legal frames, citation patterns, and writing conventions. Associate onboarding accelerates because the institutional knowledge is searchable. Senior associates supervise consistently because the standard work product is consistent.

CLE and internal training automation. Feed the AI the firm's prior CLE materials and recent regulatory updates. Generate practice-group CLE at scale. The materials need partner editing but the structural drafting becomes a 30-minute task. Firms that run this well also pull CLE credit for the partner-edit time the firm was already absorbing.

Pitch and proposal generation. AI drafts pitch responses by combining prior pitch decks, relevant practice-group experience, and the new client's specific needs. Senior associate edits, partner approves. Pitch turnaround compresses from three days to four hours, which matters on inbound RFPs with short deadlines.

Matter-budget and pricing analysis. Feed the AI the firm's time entries from past matters, the matter outcomes, and the client billing arrangements. Ask it to identify which matter types have predictable budgets, which have high variance, and where alternative fee arrangements would have produced better outcomes. The firm's pricing partner gets data-driven inputs for the next pitch. The next AFA on a similar matter becomes more confident.

The law firm AI consulting connection

This is one tool category in one practice area. The bigger AI question for mid-size firms is structural. The firms that figure out where AI fits, where it does not, and how to deploy it with the right privilege architecture and supervision protocols end up with better realization rates, faster matter turns, and a competitive position against BigLaw firms that previously won every cross-jurisdiction pitch on resources alone. The firms that wait usually end up either banning AI awkwardly, allowing it under the table without supervision, or both.

If your firm is wrestling with the bigger AI question, the AI Consulting for Law Firms page covers the full scope: where AI actually fits in mid-size firm operations, what the common failure modes look like, how the privilege architecture works under the current state bar opinions, and what an engagement looks like when it works.

Closing

The goal is not for partners to become AI architects. It is for the firm to deploy AI legal research without trading away the privilege protections, work product doctrine, or supervision obligations that define what a law firm is. The architecture is what makes it work. Get the contract right, get the DMS integration right, get the verification protocol right, and the rest of the value follows.

Pick one practice group. Sign the enterprise agreement. Run a 90-day pilot with a written policy and malpractice disclosure on file. Then extend it.

If you want to talk about how AI fits into your firm at the program level, the AI Consulting for Law Firms page lays out the full picture and how an engagement works.