At many law firms, notary offices, tax consultancies, and accounting firms, staff already use ChatGPT or Claude to summarise contracts, draft letters, or review tax documents. It does save time. The problem is that client documents are often uploaded through personal accounts, with no clear rules and no record of who uploaded what.
Banning AI outright rarely works, because staff will simply keep using it quietly. A more realistic approach is to use AI on client documents safely: understand the risks, limit access, and write an internal policy people can actually follow. This article explains how.
The risks of using AI on client documents
Professional firms hold highly sensitive information: client identities, ownership structures, disputes, financial statements, and tax data. There are several main risks when documents like these reach an AI assistant without any controls.
- Personal accounts outside the firm's control. Documents uploaded to a staff member's personal account cannot be retrieved or audited by the firm, especially after that person leaves.
- Varying data retention policies. Depending on the account type and its settings, an AI service may keep conversation history or use it to improve the service. Many users never check these settings.
- Access that is too broad. For a single question, staff sometimes upload a client's entire file when only one document was needed.
- No trail. If a client asks where their documents were processed, the firm has no answer.
- Unverified answers. AI can be wrong. Without citations back to the original source, mistakes are hard to catch before they reach the client.
Principles for using AI on client documents safely
Use accounts the firm manages
The simplest first step is to move AI use from personal accounts to accounts the firm manages. That way the firm can control who has access, check the data retention settings, and revoke access when someone leaves. Also check whether the history and data-use settings match your firm's policy.
Limit access by folder
Rather than giving AI access to the whole archive, decide which folders it may read for a given piece of work. If a staff member is handling one client, the AI only needs to see that client's folder. The principle is the same as physical access to a filing cabinet: only what is needed, for a clear purpose.
Record every access in an audit trail
Every time the AI reads a document, the system should record who made the request, which document was opened, and when. An audit trail helps the firm answer client questions, trace mistakes, and show that it is exercising due care.
Mask identity data before AI reads it
Much of the work does not require full identity numbers. NIK (national ID numbers), NPWP (tax ID numbers), and bank account numbers should be masked before a document is processed by AI. Contract analysis or a report summary can still be done without the AI ever seeing those numbers.
Ask for answers with citations
Make it a habit to ask the AI to point to the part of the document its answer is based on. With citations, staff can double-check quickly, and final responsibility stays with the professional who signs off on the work.
UU PDP and the duty to protect client data
Indonesia's Personal Data Protection Law (UU PDP, Law No. 27 of 2022) governs how personal data is collected, processed, stored, and protected. In general terms, organisations that process personal data are expected to have a lawful basis for doing so, keep the data secure, limit its use to the stated purpose, and be able to account for what they do with it.
Client documents at professional firms almost always contain personal data. Processing them with AI should therefore be treated as part of the firm's data governance, not just a question of which tool to use. Beyond UU PDP, many professions also carry confidentiality obligations towards their clients. For how this applies to your firm specifically, consult a legal adviser who knows the field.
Internal AI policy checklist
A good policy does not have to be long. What matters is that it is clear and workable. Use this list as a starting point:
- Name the AI services that are allowed, and prohibit personal accounts for client documents.
- Define which types of documents may and may not be processed by AI.
- Require masking of NIK, NPWP, and other identity data before AI reads a document.
- Limit AI access by folder or by matter, never the entire archive.
- Make sure every AI access is recorded and can be reviewed.
- Require human review of every AI output before it is sent to a client.
- Appoint a policy owner and schedule regular reviews.
- Train staff, including how to ask effective questions and where AI's limits lie.
A document workspace where you set the AI rules
Running the checklist above by hand is hard work. Folders is built for law firms, notaries, tax consultants, and accountants whose staff already use AI. Client documents live in one place, you choose which folders ChatGPT or Claude may read through MCP, and every access is recorded in an audit log. NIK and NPWP are masked before they reach the AI, and answers from your documents come with clause-level citations. A free plan is available to try it out.
