Key facts
| Matter attribution | Workspace and scoped keys per matter make every event attributable |
| Retrieval traceability | Document references from retrieval can be logged alongside model metadata |
| Content policy | Metadata by default; prompt content only under an approved purpose |
| Review records | Lawyer edits and approvals stay in the document and workflow systems |
| Export | SIEM platforms including Splunk, Sentinel, QRadar and Datadog |
| Legal hold | Retention is configurable, but holds must be preserved deliberately in your records systems |
| Deployment | Cloud, VPC, on-prem and air-gapped options for privileged material |
| Status | Audit export is live; assistants and responses endpoints are coming soon |
TL;DR
- Log per matter: access, retrieval, model version and lawyer review.
- Prefer metadata over content; privilege makes over-logging a liability.
- Join AI events with document management records for a full chain.
- Plan legal holds before a dispute, not during one.
- Where privilege demands it, keep logs inside the perimeter.
How it works, step by step
- Define the defensibility questions: who saw which document, what did the model produce, and who verified it.
- Provision a workspace and scoped keys per matter so events carry attribution without extra plumbing.
- Log retrieval references and model metadata for every AI-assisted review, with a matter-level correlation ID.
- Keep prompt content out of logs unless a specific, documented purpose requires capture.
- Join AI events with document management and review records so one query shows the full chain of custody.
- Configure retention per matter and document how legal holds override purge schedules.
- Rehearse a challenge: produce the review history for one document from logs plus the review system.
Try it yourself
Open the AI API key security checklist →
What defensibility requires
A challenge to AI-assisted work usually asks four things: who had access, what was reviewed, what the tool produced, and what a qualified lawyer did with it. Design the log to answer all four. Per-matter workspaces and keys give attribution; retrieval references show which documents informed an output; model and prompt versions show what produced it; the review system shows the human decision.
Plugsky exports access, key, administrative and inference events to internal SIEM. Your document management and review systems hold the substantive record. The correlation identifier is what turns those two sources into a coherent chain rather than parallel histories.
Privilege-safe logging
For legal work, more logging is not automatically safer. Storing document text or full prompts in a log system widens the set of places privileged material resides and complicates production in litigation. Prefer metadata: identifiers, references, versions, timestamps and actors.
- Minimise content: capture prompts only with a documented purpose and restricted access.
- Separate systems: keep log access distinct from matter data access.
- Perimeter option: where privilege demands it, run and store logs inside an on-prem or air-gapped estate.
Hold, retention and the challenge drill
Retention defaults should be short enough to reduce exposure and configurable per matter. But holds override defaults: when litigation is reasonably anticipated, the firm must preserve the relevant records — including AI logs — through documented processes in systems where holds are actually enforceable.
Prove the pipeline before it is tested. Pick a document, reconstruct its review history from logs and the review system, and time how long that takes. Repeat with a different reviewer and a different matter type. The exercise reveals missing joins and retention gaps while they are still cheap to fix.
Honest comparison
| Defensibility need | Plugsky | Typical AI platform | In-house tooling |
|---|---|---|---|
| Matter attribution | Workspace and keys per matter | Shared tenancy | You design it |
| Retrieval traceability | Log retrieval references with model metadata | Not exposed | Your implementation |
| Privilege safety | Metadata-first, configurable content capture | Often opaque | Your policy |
| Chain of custody | Correlated with your DMS and review records | Not addressed | Your integration |
| Hold readiness | Configurable retention, holds managed in your systems | Vendor-dependent | Fully yours |
| Perimeter option | On-prem and air-gapped deployment | Rarely | Possible, costly |
Frequently asked questions
Do we have to log the documents the model read?
Logging retrieval references — identifiers of the documents retrieved — is usually enough and far safer than logging content. It shows what informed an output without duplicating privileged text.
How do we attribute AI activity to a matter?
Use a workspace and scoped keys per matter. Every access and inference event then carries matter attribution without additional tagging logic.
What about legal holds?
Holds are managed in your records systems, but AI logs are within their scope. Configure retention so hold obligations can override purge schedules, and document the process.
Should logs live inside the firm's network?
If privilege or engagement terms require it, run an on-prem or air-gapped deployment so logs and documents never leave the perimeter.
How do we prove the chain of custody?
Join inference events with document management and review records through a shared identifier, and rehearse a reconstruction exercise for a real document.
Who can read the logs?
A restricted set of roles separated from matter access — typically risk, security and a designated partner — with access itself logged.
How long do we keep AI logs?
Per matter type and firm policy, with short defaults and documented rationale. Review the settings whenever review tooling or workflows change.