Governance & data control
AI Governance, Data Control & Compliance Readiness
Build AI workflows with defined permissions, review points and data controls. See how information moves, where people decide, and how each action can be checked.
For law firms, legal-tech founders, in-house teams and software companies working with sensitive information.
Technical implementation and evidence support, alongside your legal, compliance and audit specialists. Scope and applicable requirements are agreed for each engagement.
Draft a project update for Project AIllustrative · fictional data
Each control is explained below
Start here
Before a pilot becomes operational
Define permissions, review points and escalation before expanding access.
Start here
When data crosses systems
Examine applications, models, integrations and secondary copies together.
Start here
When someone asks for evidence
Connect agreed requirements to implemented controls and verification records.
01 / AI governance
Give agents a task. Define their authority.
Decide what the system may prepare, what it may execute, and where a responsible person must decide. Keep those boundaries through every tool call and handoff.
Illustrative · fictional data
Agent run
Recorded: Draft version + recipient + permission decision
Waiting for human review. Sending is held. Content and recipient are bound to this review.
W1 / In operation
Prepare the work. Hold the action.
Agents retrieve and prepare. Your team reviews consequential actions before the system executes them.
- Receive a scoped request
- Retrieve permitted sources
- Prepare a referenced draft
- Wait for human review
- Re-check permission and record the action
What this covers
Agent identity & delegated access
Define workflow identities, permitted tools and the authority attached to each request.
Approval before action
Show the proposed action, recipients and supporting information before execution.
Controlled agent handoffs
Pass necessary context and permissions; validate what the next agent returns.
Grounded, reviewable outputs
Preserve source references and flag missing or conflicting information.
Model & workflow change checks
Re-run agreed evaluations when models, prompts, tools or policies change.
Stop, escalate & recover
Bound retries, queue uncertain work and provide an operational stop path.
Where control is often lost
Did approval cover the action that ran?
Bind approval to content, destination and version. Re-check before execution.
Can a retry send the same action twice?
Use action IDs, duplicate suppression and bounded retries; reconcile partial success.
Can one agent widen another's authority?
Carry the initiating identity and limits through delegation; enforce permissions outside generated text.
Exceptions, evidence and related workflows
W1 / Exceptions and evidence
Starts with: A team requests a summary and an outbound update.
When it cannot proceed. A changed draft, recipient or permission invalidates the pending action and returns it to review.
Evidence. Source IDs, draft version, approval decision, authorization outcome and action result.
W2 / Also in scope
Keep authority intact across handoffs.
Carry the original permissions through delegated tasks and background work.
- Resolve the requesting identity
- Delegate only necessary context
- Validate returned results
- Deny unexpected tools or broader access
Starts with: A coordinator needs retrieval and drafting support.
When it cannot proceed. A delegated agent requests wider access: deny the request and escalate to an owner.
Evidence. Initiating identity, delegated scope, tool request and permission decision.
W6 / Also in scope
Release a model change with evidence.
Review what changed and what regressed before expanding a workflow's use.
- Version the proposed change
- Run agreed evaluation cases
- Investigate failures
- Approve or hold release
- Monitor with a rollback plan
Starts with: A model, prompt or tool update is proposed.
When it cannot proceed. A failing access or output case holds the release. Passing a selected test suite does not establish universal safety.
Evidence. Version comparison, test results, release decision and rollback procedure.
Boundary. An approval gate supports oversight; it does not establish that the draft is correct. Agree reviewer responsibilities, escalation owners and prohibited actions.
02 / Data governance
Make ownership and lifecycle rules operational.
Connect useful information to its owner, purpose and access boundary. Apply those decisions to source records and the copies a workflow creates.
Illustrative · fictional data
Example question · What changed in the Project A policy?
Scope · Project ARecorded: Access decision + source versions + exclusions
Assembling a cited answer. Source-linked draft ready for review, with the stale item flagged.
W3 / In operation
Answer from the right knowledge.
Retrieve information within the user's project boundary, with sources and freshness visible.
- Resolve current access
- Filter eligible sources
- Retrieve approved excerpts
- Flag stale or conflicting records
- Assemble a cited answer
W4 at a glance · fictional statuses
When a record must go, follow every copy.
Original record
RemovedSearch index & embeddings
InvalidatedCaches & agent memory
ClearedExported copy
Owner decisionBackup
Pending expiry
What this covers
Ownership & classification
Catalogue source owners, data categories, purposes and handling rules.
Permission-aware retrieval
Apply user, matter and tenant boundaries; test denied-access cases.
Provenance & freshness
Keep source versions and surface stale, conflicting or failed ingestion results.
Memory & context lifecycle
Scope conversation history, agent memory and caches; define expiry and re-use.
Retention & deletion propagation
Trace lifecycle rules through originals, derived records and secondary copies.
Controlled sharing & export
Check recipients, destinations and requested fields before information leaves.
Where control is often lost
What does the agent remember tomorrow?
Separate memory by purpose and project. Define what expires and what may be reused.
What if access changes while work waits?
Check current permissions at retrieval and execution, including queued tasks.
Does deletion reach every derivative?
Track embeddings, caches and exports; identify retained backups and unreachable copies.
Exceptions, evidence and related workflows
W3 / Exceptions and evidence
Starts with: A user asks a question about Project A.
When it cannot proceed. Revoked permission or an out-of-date index holds affected retrieval for review.
Evidence. Access decision, source versions and exclusion reasons without restricted content.
W4 / Also in scope
Close the lifecycle across every copy.
Track what was removed, what remains and who owns the next decision.
- Validate lifecycle request and holds
- Locate linked records
- Remove eligible active copies
- Invalidate indexes and caches
- Track backup expiry and exceptions
Starts with: An authorized lifecycle request reaches its review point.
When it cannot proceed. A retained backup stays marked pending expiry; a retention conflict needs an owner decision.
Evidence. In-scope inventory, removal results, retained copies and outstanding obligations.
Boundary. Business owners define purpose and operating rules. Relevant specialists validate retention requirements and legal holds; no universal retention period is assumed.
03 / Data sovereignty
Follow the data beyond the hosting region.
Review storage, processing and access across the full lifecycle. Make supplier dependencies and fallback behavior explicit before relying on a location setting.
W5 / In operation
Keep processing within agreed boundaries.
Make processing and fallback routes explicit across the application and its dependencies.
- Classify the request
- Choose an approved route
- Check service configuration
- Process minimized input
- Record route and secondary data paths
Illustrative · fictional data
Secondary data paths · reviewed separately
- Logs & telemetry
- Backups & recovery
- Support & administration
Recorded: Category + route + configuration version
Delivering and recording the route. Category, route and configuration recorded; secondary paths reviewed separately.
Map the information, from source to destination.
Illustrative review inventory. Actual locations, responsible people, retention periods and supporting evidence are to confirm for each engagement.
01Approved sources
- Data:
- Documents and attachments
- Location:
- Source systems
- Access owner:
- Source custodian
- Retention:
- Agreed source policy
- Evidence:
- Source/access inventory
02Ingest & classify
- Data:
- Extracted text and metadata
- Location:
- Ingestion service
- Access owner:
- Pipeline owner
- Retention:
- Temporary-copy rules
- Evidence:
- Ingestion configuration
03Retrieve
- Data:
- Chunks and embeddings
- Location:
- Retrieval store
- Access owner:
- Knowledge owner
- Retention:
- Reindex and deletion rules
- Evidence:
- Permission/retrieval tests
04Process
- Data:
- Prompts and selected context
- Location:
- Approved model service
- Access owner:
- Service owner
- Retention:
- Provider handling terms
- Evidence:
- Route/settings review
05Review
- Data:
- Draft outputs and references
- Location:
- Review application
- Access owner:
- Authorized reviewer
- Retention:
- Draft/approval retention
- Evidence:
- Versioned review record
06Deliver
- Data:
- Authorized output
- Location:
- Approved destination
- Access owner:
- Destination owner
- Retention:
- Destination and export rules
- Evidence:
- Action/export receipt
Choose a deployment by its responsibilities.
Options to assess against your requirements and operating capacity.
Managed service
Existing operational support and integration fit.
Still to validate
Validate processing paths, settings, support access, contracts and exit.
Customer-controlled cloud
More direct infrastructure configuration ownership.
Still to validate
Review external model calls, telemetry, operations ownership and recovery.
Self-managed environment
A specific isolation or operational requirement.
Still to validate
Assess model suitability, updates, capacity, security maintenance and dependencies.
What this covers
End-to-end data-flow mapping
Trace documents, prompts, outputs, embeddings and metadata across systems.
Approved processing routes
Define allowed providers, services and locations, including outage behavior.
Support & privileged access
Examine administration, remote support and emergency access alongside hosting.
Key & identity control
Document who can use keys, grant access and operate the service.
Logs, backups & recovery
Review copy locations, telemetry, retention, restore access and expiry limits.
Supplier change & exit
Plan usable exports, portability, access revocation and verification at exit.
Where control is often lost
Where does work go during an outage?
Validate fallback paths against the agreed allowlist; otherwise queue or stop.
Can debugging create another data store?
Minimize payload logging; scope telemetry destinations, access and retention.
Who can reach it through support?
Map provider administration and support access, including remaining capabilities when customer-managed keys are used.
Exceptions and evidence
W5 / Exceptions and evidence
Starts with: A request includes information with a defined handling category.
When it cannot proceed. If no approved provider is available, queue or stop the request rather than silently reroute it.
Evidence. Data category, selected route, configuration version and exception outcome.
Boundary. Hosting location, applicable legal/contractual requirements and practical control are distinct review questions. Specialists validate applicability; deployment choices alone do not guarantee sovereignty.
04 / Compliance readiness
Connect the requirement to the control and its evidence.
Translate specialist-validated requirements into technical work your team can review. Keep owners, exceptions and verification connected as the system changes.
W7 / In operation
Keep evidence connected to the control.
Prepare traceable review records and surface gaps for the people responsible for assessment.
- Start a scheduled review
- Collect selected records
- Compare the agreed baseline
- Flag stale or missing evidence
- Prepare an owner review packet
Illustrative · fictional data
Recorded: Control reference + owner + dates + exceptions
Prepared for owner review. Control source, review dates and open exceptions included
W8 at a glance
If something goes wrong: stop, review, restart on a decision.
Pause queues
Stop affected work before it repeats.
Revoke access
Withdraw the permissions involved.
Review actions
Reconcile what already completed.
Restart by decision
Resume only on an explicit owner decision.
What this covers
Requirements-to-control mapping
Turn agreed requirements into implementation scope and acceptance checks.
Evidence collection workflows
Assemble configuration and test records with source, owner and review date.
Access reviews & exceptions
Prepare permission reviews and track exception owners, expiry and remediation.
Configuration drift checks
Compare selected settings with the baseline and route changes to an owner.
Incident reconstruction
Link requests, versions, approvals and tool results while minimizing payloads.
Operational handover
Deliver responsibilities, runbooks, limits and an ongoing review process.
Where control is often lost
Can the evidence itself expose data?
Restrict evidence access and changes; use scoped integrity controls and minimize sensitive payloads.
Who owns an exception after review?
Name a decision owner, expiry or next review, and the remediation dependency.
Can the team stop and restart safely?
Pause queued work, assess completed actions and require an explicit restart decision.
Exceptions, evidence and related workflows
W7 / Exceptions and evidence
Starts with: A scheduled control review becomes due.
When it cannot proceed. Missing evidence or changed configuration creates an exception, never an automatic pass.
Evidence. Control reference, evidence source, collected/reviewed dates, owner and open exceptions.
W8 / Also in scope
Stop repeated actions during an incident.
Provide a stop path, reconstruct the event and restart under an explicit decision.
- Detect a signal
- Pause affected queues
- Revoke applicable access
- Review completed actions
- Remediate and authorize restart
Starts with: An unexpected action or suspicious access signal is detected.
When it cannot proceed. An already-sent message cannot be undone by pausing. Completed actions need reconciliation or remediation.
Evidence. Request/action IDs, pause and revocation records, incident decisions and restart authorization.
Boundary. I implement technical controls and support evidence preparation. Legal interpretation, compliance decisions and independent assessment stay with the responsible specialists; I do not issue certifications or guarantee compliance.
05 / What you receive
From an unclear boundary to a reviewable system.
Agree the outputs your engagement needs. Each connects an operational decision to working controls and evidence.
Step 01
Map
Data-flow mapCurrent-state systems and data flows: sources, recipients, copies, owners and unknowns.
Step 02
Prioritize
Control-gap registerA control-gap register with impact, dependencies, decision owners and next actions.
Step 03
Implement
Agreed changesAgreed permission, workflow and data-handling changes with acceptance criteria.
Step 04
Verify
Verification recordsTraceable test results, limitations, findings and remediation records.
Step 05
Hand over
Runbook & responsibilitiesConfiguration notes, operating runbooks and responsibility for review, change and exit.
Deliverables and verification depth depend on agreed scope; this is not a fixed package or an independent assessment.
Clear responsibilities throughout.
Your team
Purpose, operating ownership, approvals and risk acceptance.
Legal & compliance
Applicability, interpretation and policy decisions.
- My scope
Engineering
Agreed implementation, verification records and remediation.
Independent assessors
Assessment conclusions, reports and certifications where applicable.
Related work & services
See the adjacent engineering scope.
- Legal AI engineeringBuild and test legal AI workflows inside Word and the firm's tools, with human review and any model.
- Human-governed intakePublished intake engineering with human review. It demonstrates that workflow's approved scope, rather than every control on this page.
- AI agent developmentBuild multi-step agents around approved knowledge, tools, permissions and escalation.
- Application security testingCommission an authorized assessment of agreed web, API and AI surfaces.
Questions before you start
Scope, evidence and next steps.
Can you help with an existing system?
Yes. Start by mapping its workflow, dependencies and gaps, then agree what needs implementation or verification before changing it.
Does regional hosting settle the sovereignty question?
Review processing, secondary copies, support access and supplier arrangements alongside the hosting location. Relevant specialists validate the applicable requirements.
Can you guarantee compliance or issue certification?
No. My scope is technical implementation and evidence support. Legal interpretation, compliance decisions and independent assessment stay with the responsible specialists.
What would you need to scope the work?
A high-level workflow description, the systems involved, data categories and known constraints. Start without sending confidential records or credentials. Deliverables depend on the agreed scope.
How does this relate to security testing?
Governance defines responsibilities and intended controls. A separately agreed security assessment tests relevant implementation boundaries. Active testing requires authorization.
Start with one workflow
Make the control questions concrete.
Describe the systems involved, the decisions you need to make and where you need technical help. We can agree the assessment or implementation scope from there.
Discuss your governance requirementsTechnical implementation and evidence support, alongside your legal, compliance and audit specialists. Scope and applicable requirements are agreed for each engagement.
Last reviewed