IntelekIQ
Secure AI for Legal Practice
Application Note A3  |  The classification scheme

Classifying Firm Information for AI

A practical traffic-light scheme for deciding how AI may use each class of firm information and where that information may be processed.

For: IT, the DPO, and compliance leads  ·  Reading time ~8 min

The previous note explained why a firm must understand where its data goes. This note provides a practical next control. It classifies information according to sensitivity, then assigns a rule for how AI may use it and where it may be processed. Three clear classes are easier for lawyers and staff to apply consistently.

1Three classes, one rule each

Sort the firm's information into three classes by sensitivity. Attach one clear rule to each class.

Each class needs a different level of protection. Published statutes do not require the same controls as confidential client correspondence.

The traffic-light classification One scheme that decides what each kind of information may do, and where it may go Class Examples How AI may use it Where it may go GREEN Firm knowledge Statutes, case law, practice directions; firm precedents & clause libraries; SOPs; verified research memos Firm-wide knowledge index; provenance on every item Grounds answers; shareable firm-wide AMBER Confidential client work Live matter documents, correspondence, drafts, attendance notes, instructions Per-matter index; retrieval security- trimmed to cleared staff Use with private AI; processing stays under firm control RED Excluded Sensitive personal data; trust / client-account records; material under undertaking, sealing order or ethical wall; litigation-hold Not ingested into general indexes at all Scoped enclave only, by exception, with DPO approval The rule underneath: allow-list and default-deny. Name the sources you ingest; never point AI at "everything on the file server".
Figure 1. Classification has practical value only when the label changes how AI may use the information and where it may be processed.

Green: firm knowledge. This is non-confidential material that the whole firm can share. It includes statutes, case law, the firm's precedents and clause libraries, standard procedures and verified research memos. Each item should record its source, verifier and date so lawyers can check where it came from.

Amber: confidential client work. This includes live documents, correspondence, drafts, attendance notes and instructions. AI may use this material only within controls approved by the firm. With Secure Law, processing takes place on premises and access remains limited to lawyers and staff authorised for that matter.

Red: excluded. This material does not enter the general AI path. It includes special-category personal data, trust and client-account records, and material subject to an undertaking, sealing order or ethical wall. Any exception needs a separate, restricted process and the Data Protection Officer's authorisation.

2The two rules that make it work

A classification scheme depends on disciplined ingestion. Two rules provide the foundation.

The first is allow-list, default-deny. Name the sources that AI may index, such as a specific SharePoint library or precedent bank. Do not point it at every file the firm holds. Anything not explicitly allowed is not ingested.

The second is no single index for all client work. Combining documents from every matter into one searchable pool creates a risk that one client's confidential information will surface in another client's work. Confidential client files should remain in separate matter compartments with access controls for each matter. The next note explains how retrieval can enforce that separation.

Classification defines permission. A label matters only when it changes what AI may do with the information and where that information may be processed.

3The edge cases lawyers will recognise

Three categories need particular care because their restrictions may extend beyond ordinary confidentiality.

Documents obtained through litigation disclosure may carry an implied undertaking that limits their use to that proceeding. They should remain confined to the originating matter. Information behind an ethical wall must remain unavailable to excluded lawyers and staff. Trust and client-account records also stay outside the general AI processing path because Secure Law is not their system of record.

4Putting it to work

The scheme becomes useful through a short, repeatable procedure and a simple check that every lawyer or member of staff can apply before uploading information.

  1. Inventory the named sources. List the specific libraries, drives and systems that hold firm knowledge and confidential client files, by name, not "the server".
  2. Assign a class to each. Green, Amber or Red, using the table above. When in doubt, classify up, not down.
  3. Attach the rule. Green is allow-listed into the firm knowledge index. Amber stays in separate matter compartments under the firm's control. Red is excluded from general ingestion and handled through a separate restricted process, if an exception is approved.
  4. Review on a cadence. New matters, new systems and departing staff all change the picture; a quarterly pass keeps the classification honest.
Can this information go into an AI tool? A short check for lawyers and staff before they paste or upload NoNo YesYes YesNo 1 · Is it published or firm knowledge? Statutes, precedents, SOPs: GREEN 2 · Is it excluded? Sensitive personal, trust, undertaking: RED 3 · It is confidential client work (AMBER). Will processing stay under the firm's control in Secure Law, or use another tool specifically approved for it? Use it Still verify the output before you rely on it. Stop Scoped enclave with DPO approval only. Approved for this client work Inference stays under the firm's control. Not permitted here Use Secure Law or another approved route.
Figure 2. Three questions cover common decisions. Genuine edge cases should go to the DPO rather than depend on an individual's judgement.

An honest bound

Classification reduces risk but does not replace judgement. Some documents contain information from more than one class. Treat the whole document at its most sensitive level until the information is separated. Refer uncertain cases to the DPO.

With information sorted this way, the two decisions that follow have a clean foundation: which architecture keeps Amber inference under the firm's control (A4), and how the retrieval system enforces the per-matter compartments in practice (A5).

Notes & further reading

  1. Companion papers: A2 Follow the Data (the inference wall this scheme relies on); A4 Cloud, On-Premises or Sealed?; A5 Grounded Retrieval and the Two-Index Model (how the per-matter compartments are enforced).
  2. On the implied undertaking in litigation disclosure, and ethical-wall obligations, see the professional-conduct rules in your jurisdiction (Bar Council Malaysia and the Legal Profession Act 1976; the Law Society of Singapore and the Legal Profession (Professional Conduct) Rules).
  3. On special-category and trust-account data handling, see your applicable data-protection regime (the Malaysian PDPA 2010, as amended 2024; the Singapore PDPA).
IntelekIQ · Secure AI for Legal Practice · Application Note A3: Classifying Firm Information for AI. Next in the series: A4: Cloud, On-Premises or Sealed? Choosing an AI deployment model for confidential client work.