AI for legal

Which AI should a law firm use?

A practical guide focused on real work: ChatGPT, Claude, Gemini, NotebookLM and Copilot applied to the routines of a law firm.

G

BY

Guilherme Dantas
10 min read

The choice depends on the work that needs to be done. Each tool serves a different part of the firm's operation.

To research a recent regulation, ChatGPT and Gemini search the web. To review hundreds of pages of an old case file, NotebookLM cites the source page. For teams living in Word, Outlook and Teams, Copilot keeps the work inside the systems they already use. To standardize contract review against the firm's playbook, Claude applies the firm's own rules.

Technology choices in legal work start from the task that needs to improve and from the confidentiality level of the information involved. Benchmarks help, but they do not decide on their own.

1. The 30 second answer: where to start

To start testing next Monday, use this mapping between task and tool:

What you need to do todayWhere to start
Research a current topic on the webGemini or ChatGPT
Research many sources in depthGemini Deep Research or ChatGPT Deep Research
Study selected case files and documentsNotebookLM
Work on contracts and playbooksClaude or ChatGPT
Analyze spreadsheets and data volumesChatGPT or Copilot
Work inside Word, Outlook and TeamsMicrosoft Copilot
Standardize a recurring taskSkills, Gems or Agents
Connect AI to the firm's DMSConnectors / MCP

2. Five tools, five real situations

Compare the tools by the work they take off a lawyer's desk.

ChatGPT: data analysis and investigation

Your team received a spreadsheet with 500 labor or civil cases and needs to understand total financial exposure, risk by legal theory and concentration by client.

ChatGPT handles structured data and investigative analysis well. It reads the spreadsheet, groups the cases, builds preliminary visualizations and flags outliers that a manual check can miss.

To look outside the firm, Search and Deep Research retrieve recent case law and regulation from the web. In daily bench work, Projects keep each client's memory and context organized, so nobody has to re-explain the history of a matter with every new prompt.

Claude: document workflows and playbooks

The firm reviews dozens of contracts a month and lawyers spend their time checking the same compliance points.

Claude reads long and complex drafts. The team can turn the firm's review rule into a reusable Skill, with the same procedure for every associate.

The bench writes the procedure down: non-negotiable clauses, liability caps and fallback wording. Claude applies that checklist and returns a draft with the deviations flagged for the lawyer to decide.

The Claude for Legal repository collects examples of Skills, agents and connectors for legal workflows. Use it as a technical reference, not as a ready-made playbook: the rules have to be adapted to your jurisdiction and to the firm's practice.

Anthropic also documents how Claude marks AI generated content. That marking can support transparency about where a text came from, but it does not replace the lawyer's review and it does not by itself define how the firm should disclose its use of AI.

To test: build a Skill with four output fields: clause, problem, risk and suggestion. Load an anonymized draft and check the flags against the firm's checklist.

Gemini: regulatory research and the Google ecosystem

The central bank or the securities regulator published a new rule and the partner needs an executive summary of the impact by the end of the afternoon.

With web access, Gemini's Deep Research locates official publications, public consultation drafts and sector reporting quickly. You can instruct it to prioritize government portals and official gazettes.

For firms on Google Workspace, it works inside Drive, Gmail and Docs. A recurring answer format can also be saved as a Gem to hold the firm's standard.

NotebookLM: focused study of case files with traceable sources

You inherited a heavy dispute, with the complaint, the answer, the reply, expert reports and rulings, and you need to rebuild the timeline without reading 400 pages from scratch.

In NotebookLM, answers stay tied to the notebook's sources. You can upload documents or discover and import sources from the web. Fast Research and Deep Research in NotebookLM help find and import that material.

You can ask for the chronology of facts, the contradictions between testimonies or the list of claims in the complaint. Traceability is what lets the lawyer check the result: every passage in the answer carries a clickable citation that opens the exact excerpt of the original filing.

To test: load the complaint, the answer and two rulings. Ask for a timeline, open the citations and verify five facts directly against the sources.

Microsoft 365 Copilot: native operation inside Office

Your team spends 90% of the day moving between Word, Outlook, Teams and SharePoint. Nobody wants one more browser tab.

Here the convenience of integration beats any debate about which model is 5% smarter. Copilot runs inside the programs the team already uses: it summarizes long email threads in Outlook, finds similar clauses in contracts stored in SharePoint, supports drafting in Word and produces meeting summaries in Teams while respecting the access permissions the firm already has.

To test: open a Teams meeting, ask for a list of decisions, owners and deadlines, and check each item against the transcript before sharing it.

A quick exercise for ChatGPT and Gemini

To compare the two tools, use the same anonymized case spreadsheet. Ask for a risk classification, the criteria used, a table of exceptions and a sample of five rows for manual validation. The useful result is the one the team can review and repeat.

To test: strip names and identifiers, make the same request in both tools and record time, errors found and rework during the check.

3. Prompt, Skill, Connector, MCP and Agent in plain language

In corporate work, a tool only helps once it enters the team's routine.

These pieces do different jobs:

  • Prompt: the single question or instruction you write to resolve one specific doubt.
  • Skill: the firm's procedure, your review checklist, recorded so it repeats automatically.
  • Connector: the technical link between the AI and the system where your documents live.
  • MCP (Model Context Protocol): an open standard defining how AI applications talk to external tools and sources. Authentication, authorization and permissions depend on the implementation.
  • Agent: the full flow in which the AI uses tools, queries systems and runs steps under the team's supervision.

How information moves through the flow: from upload to decision

Think about reviewing an NDA. The first gain is finding the right file, the valid version and the playbook before anyone starts reading.

Exemplo: revisão de NDA

Produtividade não é a IA decidir. É ela tirar a busca e a primeira leitura do caminho.

Antes

O advogado abre pastas, procura o último contrato, relembra o padrão da bancada e só então começa a comparar cláusula por cláusula.

Com o fluxo conectado

  1. 01

    Prompt

    O advogado faz o pedido

    “Revise o NDA do cliente XPTO.” É a instrução que inicia a tarefa.

  2. 02

    Skill

    A regra da casa entra

    O checklist de revisão do escritório é aplicado sempre do mesmo jeito.

  3. 03

    Connector

    O histórico é localizado

    A IA encontra contratos e versões anteriores no sistema onde eles já vivem.

  4. 04

    MCP

    A comunicação é padronizada

    Define como a IA conversa com o sistema. Autenticação, autorização e permissões dependem da implementação.

  5. 05

    Agent

    O fluxo prepara a análise

    Ele organiza busca e comparação, entrega os alertas e deixa a decisão com o advogado.

When the lawyer asks "Review the NDA for client XPTO", the pieces work together:

  1. the Skill brings the rules and the checklist the bench already approved.
  2. the Connector fetches the contracts and the history from the system where they already live.
  3. the MCP standardizes the communication between the AI and that system. Authentication, authorization and permissions have to be configured and audited by the implementation.
  4. the Agent organizes the sequence: retrieve, compare, flag deviations and prepare the draft.
  5. the lawyer receives the flags and makes the decision that stays under their responsibility.

The team then spends less time assembling context and more time assessing risk, negotiating and advising the client. AI speeds up the first pass. Legal judgment stays with the lawyer.

4. Before pasting client documents: governance and bar rules

Before using generative AI in the firm's routine, answer five questions. Recommendation 001/2024 of the Federal Council of the Brazilian Bar Association sets national guidelines on confidentiality, privacy, human supervision and professional responsibility. In June 2026 the Council launched the National Plan for Integrating Artificial Intelligence into the Legal Profession, which consolidates and expands those recommendations.

  1. Which plan and contract are we on? Check training policy, retention, data location, administrative controls, the DPA and the terms that apply to the connectors in use, explicitly.
  2. What information am I sending? Remove names, qualifications and document numbers that the analysis does not need.
  3. Does the AI actually need that data? To analyze a termination clause, for instance, the draft does not need to carry the names of the parties.
  4. Who has access internally? The firm's confidentiality controls have to hold inside the AI tool as well.
  5. Who signs the answer? AI produces drafts. Technical, ethical and civil responsibility toward the client and the bar remains the lawyer's.

A firm does not need to ban AI. It needs clear guidelines on which categories of document may enter which tool.

5. Which AI would I choose for my firm?

Firm profileWhere I would start
Microsoft 365 / SharePointMicrosoft 365 Copilot
Google WorkspaceGemini + NotebookLM
iManage / NetDocumentsAI plus a compatible connector
Contracts and playbooksClaude
Excel / structured dataChatGPT

Do not buy five tools at once. Pick one workflow, test two alternatives and measure time, quality and rework.

6. How to start on Monday

To get going, pick one task and follow the execution:

  1. Choose a real bottleneck: take a repetitive task that costs the team hours every week.
  2. Time the current process: know exactly what doing it manually costs today.
  3. Run a controlled pilot: use the most suitable tool on a small batch of documents.
  4. Evaluate the draft: measure the accuracy rate and the time spent on human review.
  5. Turn it into a standard: if it works, record the instruction as a Skill for the team.
  6. Connect it to your systems: only after the manual process is validated does it pay to invest in direct connections through MCP to your DMS or ERP.

Sources and official documentation

To go deeper into the technical and contractual aspects of each platform, consult the official channels:

Tools and plans change. Check the documentation before putting a workflow into production. This text is informational and does not replace your organization's own legal and security compliance analysis.

CatechLabs works on integrating models with iManage, SharePoint, ERP, case data and playbooks, preserving permissions, traceability and governance. See how we structure AI for legal operations.

By Guilherme Dantas, COO at CatechLabs.

Frequently asked questions

What is the best AI for lawyers in 2026?
There is no single best tool for every task. ChatGPT and Gemini handle research and data well; Claude handles contracts; NotebookLM handles case files; and Copilot suits firms already running on Microsoft 365.
ChatGPT or Claude: which is better for lawyers?
ChatGPT is strong on data analysis and research. Claude is the better fit for contracts and review playbooks.
Which AI should be used to analyze case files?
NotebookLM is a good option when the documents are already selected and you need to check the sources behind every answer.
Which AI should be used for contracts?
Claude or ChatGPT. Claude stands out when the firm works from a written review playbook.
Can client documents be sent to an AI tool?
Check the plan, the data retention terms, the access permissions and the firm policy first. Anonymize whatever the analysis does not need.
Does the Brazilian Bar Association allow the use of AI?
Recommendation 001/2024 of the Federal Council of the Brazilian Bar Association sets national guidelines for the use of generative AI in legal practice. In June 2026 the National Plan for Integrating Artificial Intelligence into the Legal Profession consolidated and expanded those recommendations.
How should a firm start using AI?
Pick one repetitive task, run a controlled pilot and measure accuracy, time and rework before expanding to anything else.

Facing a similar challenge?

Talk to the founders
BACK TO BLOG