AI agents in legal and compliance work, with a record of every decision
Let legal and compliance agents gather, draft and summarize, and keep filing, signing and clearing alerts with people. With NobleCloak, policy decides which tools an agent may call, which wait for an approver, and which are denied. Every decision is recorded, so you can show an auditor or examiner what your agents were allowed to do.
100 scenarios in 5 areas
These are scenarios for how teams govern AI agents in Legal, Risk & Compliance. Some depend on capabilities still in development. Products named in a scenario are systems your team runs, named to make the example concrete. The policy applies to the AI agent's calls into them.
Contracts & Legal Ops20 examples
Sending for signature asks first
A legal ops assistant can prepare a DocuSign envelope freely, but the send tool is set to ask, so the paralegal confirms recipients and version before anything goes to a counterparty.
Non-standard NDAs to counsel
Standard mutual NDAs are sent automatically, while any NDA with changed terms goes to the assigned commercial counsel for approval.
Privilege labels confirmed
Marking a document privileged in the document management system asks first, because over-labeling weakens the privilege log.
Board consents need two signers
Sending a board written consent requires approval from both the corporate secretary and the GC, not just the person who prepared it.
No voiding envelopes
Agents are denied the DocuSign void-envelope tool, because voiding a half-signed contract mid-negotiation is a call only counsel should make.
Templates belong to legal ops
Sales reps' AI can generate an order form from Ironclad templates, but editing the template itself is entitled only to the Legal Ops role.
Signature authority by role
The execute-on-behalf-of-company tool in the CLM is entitled only to the General Counsel and CFO roles; for everyone else it is not listed and cannot be called.
Outside counsel spend on approved firms
An e-billing agent paying outside counsel invoices may pay only firms on the vendor allowlist, within each matter's monthly budget.
Legal research costs metered
Westlaw and Lexis calls made by the research agent count against a monthly budget, and usage is attributed to each attorney and matter.
Contracts don't leave through side doors
A redlining agent can reach the approved model provider only through the governed egress proxy, so contract text isn't uploaded to an unvetted summarizer.
Extraction helper reads only
A sub-agent spawned to extract renewal dates and liability caps from 400 legacy contracts gets read-only CLM access and cannot update records.
Counterparty agent negotiates in
A large customer's procurement agent may exchange redlines with our contracting agent over A2A only because its org is on the peer allowlist.
Liability cap clauses are locked
The sales copilot may adjust payment terms in an order form, but any edit touching the limitation-of-liability clause is refused.
Which version went out
When a counterparty disputes which draft was signed, legal pulls the ledger entry showing which agent sent which document version, and for which user.
Matter creation for legal only
Opening a new matter in the matter management system is entitled to Legal staff; business users' AI can request one but not create it.
Legal folders walled off
The legal team's Box folders are reachable only through a connection entitled to the Legal role, so a company-wide search assistant can't surface draft settlement terms.
Renewal notices run freely
The contract-renewal agent may email business owners about upcoming auto-renewal deadlines without a prompt; it's internal, low-risk, and saves real money.
Contract paralegal for the quarter
A contract paralegal hired for a quarter-end push gets CLM edit rights that expire on the last day of the engagement.
AI cost billed back to matters
Tokens and tool calls are attributed per matter, so legal can bill AI-assisted work back to business units or clients.
Preview the sales role
Before rolling out the CLM integration, legal ops uses "Inspect as user" on a named account executive to confirm they can draft but not execute.
Regulatory Compliance (banking, healthcare, insurance)20 examples
Provisional credit confirmed
During a Reg E dispute, posting provisional credit in the core banking system is set to require in-session confirmation by the disputes specialist.
Prior auth submissions gated
A revenue cycle agent may assemble a prior authorization packet, but submitting it to the payer portal requires confirmation by the authorization specialist.
Claim denials for adjusters only
At an insurer, the claim-denial letter tool is entitled only to licensed adjusters, and each denial requires confirmation before it is issued.
Principal review of broker communications
Retail communications drafted by an advisor's AI at a broker-dealer go to a registered principal for approval before distribution.
Read the full exampleCFPB portal responses confirmed
Responding to a complaint in the CFPB portal requires confirmation by the compliance officer, because the response is part of the regulatory record.
SAR narratives drafted, not filed
A BSA analyst's assistant may draft a suspicious activity report narrative in the case management system, but the FinCEN filing tool is denied to all agents.
Read the full exampleSAR tools invisible outside BSA
Case management tools are entitled only to the BSA/AML role, so a branch or support assistant cannot even learn a SAR exists, reducing tipping-off risk.
OFAC hits can't be cleared by AI
An agent may run OFAC screening, but the disposition tool that clears a potential match is denied, keeping false-positive clearance with a named analyst.
No protected-class fields for underwriting
A credit union's underwriting assistant receives loan files with age, sex, and race fields removed from tool responses.
Recommend, don't decide
In the loan origination system, the AI can read applications and write a recommendation, but the approve and decline tools are denied.
Minimum necessary in Epic
A scheduling assistant at a health system is entitled to appointment tools but not chart-read tools, following HIPAA's minimum-necessary standard.
E-prescribing never reachable
E-prescribing tools are outside every agent's entitlement, so no prompt can route a medication order through AI.
Rate filings stay with actuarial
Submitting rate and form filings to state regulators is denied for every agent, even those that help prepare the exhibits.
Evidence for state AI bulletins
An insurer answering a state regulator's questions under the NAIC AI model bulletin exports ledger records showing every AI decision touching claims.
Model inventory with real usage
A bank's model risk team builds its SR 11-7 inventory of AI use from NobleCloak data: which agents ran, on which tools, at what cost.
No off-channel messaging
Agents are denied WhatsApp and SMS send tools at a broker-dealer, so customer communications stay on archived channels for SEC recordkeeping rules.
Calling hours enforced
An outbound collections agent can place calls and texts only between 8 a.m. and 9 p.m. in the consumer's time zone.
Member data stays with vetted models
A credit union's agents can reach only the model providers its vendor management team approved, through the governed egress proxy.
NYDFS access review
For its annual NYDFS Part 500 access-privilege review, a lender uses "Inspect as user" to document what each role's AI could reach.
One CUSO, many credit unions
A credit union service organization runs each client credit union as a separate NobleCloak org, so member data and policies never cross between institutions.
Audit & Evidence (SOC 2, examiners)20 examples
Prod merges gated and evidenced
A coding agent merging to main in GitHub must get confirmation, and the ledger records who confirmed, giving change-management evidence for each release.
Terraform apply for platform only
Production Terraform apply is entitled only to the Platform Engineering role and requires confirmation even for them.
Production changes go to a named approver
A production change requested through one engineer's agent is routed to the platform lead, who approves it before it runs.
Key rotation confirmed
Rotating or disabling an AWS KMS key requires confirmation, since a wrong rotation can lock the company out of encrypted backups.
Logical access evidence on demand
For SOC 2 CC6 testing, the security team exports the ledger of every allow, deny, and ask decision for the period, instead of screenshotting config pages.
Examiner asks what was blocked
A bank examiner asks what the AI tried to do and wasn't allowed to; compliance exports every deny for the quarter with agent, user, and inputs.
Evidence collector reads only
The compliance automation agent pulling AWS configuration evidence has read-only tools; write and delete tools are not on its allow-list.
Quarterly access review in minutes
Reviewers use "Inspect as user" on a sample of employees to confirm each person's effective AI permissions match their job.
Auditor's agent gets read-only access
The external audit firm's agent may call our evidence agent over A2A for read-only evidence requests, and nothing else.
Developers' AI stays on staging
Developers' assistants may query the staging Postgres connection, while the production connection is denied, which closes a long-standing audit finding.
Okta admin actions denied
Adding users to Okta admin groups is denied for every agent, so privilege escalation always goes through a person.
Logs can't be deleted
Agents are denied delete and retention-change tools on CloudWatch and Splunk, preserving the evidence trail auditors rely on.
Sampling confirmations
Internal audit pulls 25 random approvals from the ledger and verifies each was confirmed by an authorized person before the tool ran.
AI risk report for the board
The risk committee gets a quarterly report on AI tool usage, denials, and spend by department, drawn from NobleCloak's usage and ledger data.
Reconstructing a suspicious export
After a suspicious data export, incident response reconstructs exactly which agent called which tools, and for which user, minute by minute.
Red team on staging only
A security red-team agent is allowed to run attack tools only against the staging connection, and production connections are not on its allow-list.
Evidence tickets frozen
Jira tickets labeled as audit evidence can't be edited by agents once the audit period closes.
PHI access evidence for HITRUST
A healthcare SaaS vendor shows HITRUST assessors which agents read PHI-bearing tools, how often, and for which users.
Third-party AI calls for the examiner
A community bank shows its FDIC examiner that every AI call to an outside service went through the governed proxy, with the destination logged.
Exceptions recorded
Every per-user override is recorded with the admin who granted it, giving auditors a complete list of exceptions to the standard policy.
Third-Party & Vendor Risk20 examples
Questionnaires sent with confirmation
A TPRM agent can draft SIG questionnaires, but emailing them to vendor contacts asks first so the right contact gets the right tier.
Auto-renewals over threshold ask
An agent renewing a vendor subscription pays without a prompt under $5,000 and needs approval above it.
Purchase orders confirmed
Creating a Coupa purchase order from an agent asks first, so commitments aren't made from a misread request.
Bank detail changes need two people
If the vendor bank-detail tool is ever opened to AP, every change requires a second AP approver beyond the requester.
AI buys only from approved vendors
An agent purchasing SaaS seats or API credits may pay only vendors on the allowlist maintained by procurement, within a monthly budget.
Unreviewed MCP servers stay dark
A new MCP server a developer installs is not entitled to the org until vendor review completes, so its tools are invisible to every AI app.
No calls to unvetted APIs
An agent trying to call a new translation vendor's API is blocked at the egress proxy until that domain is approved.
Partner agents by name only
Only partner agents on the peer allowlist may call in over A2A; an unknown agent presenting valid-looking credentials is refused and logged.
No fourth-party hops
Our outsourced collections vendor's agent may call our agent, but our policy refuses calls relayed for that vendor's own subcontractor.
Offboarding cuts the connection
When a vendor contract ends, removing its connection makes every tool tied to it unreachable at once, across all AI apps.
LLM spend per provider
Monthly budgets are set per model provider, and cost attribution shows each team's spend by vendor.
High-risk vendors by role
Tools backed by vendors rated high-risk in the TPRM program are entitled only to the roles named in the vendor's approval.
No customer files to vendor portals
Uploading files that contain customer PII to a vendor's support portal is refused.
Risk ratings stay with analysts
A bank's TPRM agent may read vendor monitoring scores from BitSight, but changing a vendor's internal risk rating is denied.
Model provider concentration
Cost attribution shows most AI spend runs through one provider, which the bank's vendor risk team flags as concentration risk.
Vendor AI activity for examiners
During an NCUA exam, the credit union shows every call its vendor-hosted agents made into member systems, from the audit ledger.
Vendor remote access for one day
A core processor's support engineer gets agent access to a diagnostics tool for 24 hours, and the grant ends without anyone revoking it.
Vendor master changes denied
Agents are denied tools that create vendors or edit remittance bank details in NetSuite, a common path for payment fraud.
SOC reports from trust portals
The vendor review agent may download SOC 2 reports from vendor trust portals through the egress proxy, limited to approved domains.
Personal AI apps get the same limits
An employee who connects a personal Claude account to company tools still hits the same gateway policy, so shadow AI doesn't bypass vendor controls.
Privacy, Records & eDiscovery20 examples
eDiscovery exports for litigation support
Running and exporting eDiscovery searches is entitled to Litigation Support, and each export requires confirmation.
Breach notification letters confirmed
Sending breach notification letters through the mail vendor requires confirmation, because breach notices carry legal deadlines.
Breach drafts go to the GC
Draft breach notices produced by the incident agent are routed to the General Counsel for approval.
Review vendor spend per matter
Hosting and processing charges from the eDiscovery vendor are capped per matter by a monthly budget, with approval above it.
Holds placed by legal, never released by AI
Legal's assistant may place a litigation hold in Microsoft Purview, while the release-hold tool is denied to every agent.
No mailbox deletions during holds
The Google Workspace delete-user tool is denied for IT agents, since a deleted custodian mailbox can mean sanctions.
Productions through the governed proxy
Agents can deliver productions only to the review platform's approved domain through the egress proxy; other upload destinations are unreachable.
Retention deletes inside the schedule only
The records-disposition tool is callable only by the scheduled retention workflow after records management signs off.
PIAs open to the privacy team
Privacy analysts' AI may create and update privacy impact assessments in OneTrust freely; it's their daily work.
Redact PII in productions
Account numbers and SSNs are redacted from documents before the production agent stages them.
PII mapping from metadata only
A data-mapping agent may read Snowflake schema metadata on the catalog connection but is denied row-level queries on the production connection.
EU data stays in the EU org
A company's EU subsidiary runs as its own NobleCloak org, so agents in the US org cannot reach EU personal data.
Slack retention settings locked
Changing Slack workspace retention settings is denied for all agents, so a cleanup bot can't shorten retention in the middle of a dispute.
HIPAA records can't be deleted
For a healthcare provider, delete tools on the medical records system are not entitled to any role's AI, protecting the six-year retention requirement.
Consent records readable
Support and marketing assistants may read consent records from the consent management platform but are denied the tool that edits them.
Chain of custody from the ledger
Litigation support shows opposing counsel when each collection ran, which agent ran it, and who authorized it, from ledger entries.
Custodian under investigation
While an employee is under internal investigation, a per-user override removes their AI's export, forwarding, and delete tools.
Review cost by matter
AI first-pass review costs are attributed per matter, so legal can compare them with contract-attorney review rates.
Outside counsel's agent pulls documents
Outside counsel's agent may call our records agent over A2A to retrieve documents for a named matter, because the firm is a peer on the allowlist.
| Action | Who | Decision |
|---|---|---|
| draft_sar_narrative | BSA analyst's assistant | allowed |
| file_sar | any agent | denied |
| clear_sanctions_alert | any agent | denied |
| send_for_signature | legal ops agent | needs approval |
| export_evidence | audit team's assistant | allowed |
| release_legal_hold | any agent | denied |
Approval requests go to the legal ops or compliance lead in Slack or the web inbox (Teams and email coming soon).
Illustrative: names no package, import or API call
AI agents in Legal, Risk & Compliance: common questions
Can an AI agent file a regulatory report?
Not if the filing tool is denied. The usual setup lets the agent draft the narrative and keeps the filing with a person.
What can we show an auditor or examiner?
A record of each decision: which agent called which tool from which person's session, what policy decided (allowed, held for approval, or denied) and how any approval request was answered. It covers the calls policy blocked as well as the ones it allowed.
Who can see the sensitive compliance tools?
Only the roles you scope them to. BSA tools, for example, can stay invisible outside the BSA team, and Inspect as user shows what each person's agent would get.
Can one action need sign-off from two people?
Two or more approvals on one action are coming soon, and the scenarios that need them carry the Coming soon label. Until then an approval request goes to one approver you name. That is a single approval, not a two-person control.
Set this policy up for your own agents.
Start with your own agents, your tools and the approvers you already have.
Other functions
Finance & Accounting
Payments, the close, treasury, budgets and payroll.
Software Development & Engineering
Code review, releases, production systems and on-call.
IT & Security Operations
Identity, devices, threat response, SaaS admin and cloud.
People & HR
Hiring, onboarding, compensation, employee relations and training.
Sales & Revenue
Outreach, the CRM, quotes, contracts and renewals.
Marketing & Communications
Content, campaigns, paid media, social and the website.
Customer Support & Success
Ticket replies, refunds, privacy requests and escalations.
Operations, Procurement & Facilities
Purchasing, vendors, inventory, facilities and projects.
Data, Analytics & AI Operations
Warehouse access, pipelines, agent-to-agent work, models and AI cost.

