Agent Compliance & Risk
ADLC · Practice Area 19

Agent Compliance & Risk

Every agent action is a data processing event. Governance decides what your organization allows; compliance is what the law requires — and it applies whether or not anyone wrote it down.

Agent Compliance and Risk title slide highlighting regulatory compliance, privacy controls, GDPR and CCPA, data governance, and industry-specific controls
Source deck: Agent Compliance & Risk, slide 1.
Overview

Three questions the law asks of your agent

Not “is the agent good?” but: what law reaches this processing, what personal data actually flows, and can you prove both to a regulator.

Know the law

Which regimes reach this agent, in which jurisdictions?

  • Regulatory compliance
  • Industry-specific controls

Protect the data

Engineer safeguards into the system, not only the policy.

  • Privacy controls
  • Data governance

Honour the rights

People can demand access, erasure, and human review.

  • GDPR / CCPA duties
  • Automated decisions
What makes agents hard: they decide their own next step. An agent can retrieve data for a purpose nobody authorised, remember what it inferred, and call a tool that ships that data across a border — all inside one task.
The Five Levers

Build compliance into the agent’s data flows

The deck organizes compliance around regulatory obligations, privacy controls, rights under GDPR and CCPA/CPRA, data governance, and sector-specific controls.

Lever 1

Regulatory Compliance

Obligations arrive on three axes at once. Most teams discover the third one late, and usually from a customer’s procurement questionnaire.

Horizontal AI lawData protection lawSectoral regulation
  • Build an obligations register with source, clause, owner, control, and evidence.
  • Map every control to evidence you could hand an auditor.
  • Re-check statutes, guidance, and deadlines on a schedule.
  • Jurisdiction is decided by where the people are, not where you are.
Regulatory Compliance slide showing horizontal AI law, data protection law, sectoral regulation, obligations register, and overlapping obligations
Lever 2

Privacy Controls

Privacy by design means the safeguard is in the code path, not in the policy document. For agents, that means controlling what enters the context window in the first place.

Minimize at retrievalFetch only the fields the task needs.
Detect and redact PIIScan inputs, outputs, and tool results before storage or transmission.
Pseudonymize earlySwap identifiers for tokens so the agent can reason without knowing who.
Scope and expire memoryPer-user and per-tenant, with a retention clock.
Keep PII out of logsCapture metadata by default; traces and prompts are personal data too.
Assess before you buildA DPIA is required when processing is likely to be high risk.
Privacy Controls slide covering data minimization, PII redaction, pseudonymization, memory retention, logging, and DPIA assessment
Lever 3

GDPR and CCPA / CPRA

Two regimes, one structural difference in the deck: Europe requires a lawful basis before processing, while California centers consumer opt-out rights — and both reach automated decisions directly.

CCPA / CPRA

Consumers may exercise opt-out rights
  • Rights to know, delete, correct, and opt out of sale or sharing.
  • Limits on the use of sensitive personal information.
  • ADMT rules for automated decisions on significant matters.
  • Pre-use notice, opt-out, and risk assessments.
  • Cybersecurity audit and certification obligations phase in.
GDPR and CCPA CPRA slide comparing legal duties for personal data and automated decisions
Lever 4

Data Governance

You cannot protect, delete, or justify data you cannot locate. Governance is knowing what you hold, where it came from, where it went, and when it goes away.

ClassificationLabel data by sensitivity so controls can be applied automatically.
Lineage & provenanceTrace every record from source to every store that touched it.
Residency & transfersA prompt sent to a model API may be an international transfer.
Retention schedulesIndefinite conversation storage is a common finding.
Sub-processor registerEvery tool and MCP server is a data flow with a contract behind it.
Training-data provenanceDeploying a third-party model does not outsource lawful acquisition.
Data Governance slide showing classification, lineage and provenance, residency, retention, sub-processors, and training-data provenance
Lever 5

Industry-Specific Controls

Sector rules predate AI and constrain what an agent may see, do, and keep — often more tightly than privacy law.

Healthcare

HIPAA · GxP

PHI, minimum-necessary access, business-associate agreements, and audit logging of record access.

Financial services

SOX · FINRA · DORA

Recordkeeping and supervision of communications, model risk management, and operational resilience.

Payments

PCI DSS

Cardholder data must never enter a prompt, a log, or an agent’s memory.

Insurance & credit

Sector + fair-lending

Adverse-action explanations, anti-discrimination testing, and actuarial justification.

Industry-Specific Controls slide for healthcare, financial services, payments, insurance, and credit
The Hard Part

Personal data does not stay in one place

A user asks to be forgotten. Deleting their source row may leave personal data in vector indexes, agent memory, logs, caches, backups, and potentially model weights.

Source databaseEasy
Vector indexRe-embed
Agent memoryOften missed
Logs & tracesOften missed
Caches & backupsHard
Fine-tuned weightsVery hard

Why agents make this worse

  • Memory holds inferences and derived attributes never in the source system.
  • Embeddings encode the record; deleting the row leaves the vector.
  • Every tool call may have copied data into a system you do not control.
  • Weights that saw personal data cannot simply be edited.

What actually works

  • Maintain a provenance map from source record to every downstream store.
  • Design so weights never see raw personal data — retrieve, do not train.
  • Keep identifiers in one place; hold tokens everywhere else.
  • Set retention clocks so most data expires before anyone has to ask.
Where Personal Data Actually Hides slide, showing source database, vector index, agent memory, logs and traces, caches and backups, and fine-tuned weights
The Hard Part

When the agent decides about a person

An agent that approves, prices, screens, or denies is making an automated decision. The deck highlights duties around notice, contestability, and meaningful human review.

Is a decision being made?
Solely automated?
Legal or similarly significant effect?
Duties attach: notice, opt-out, human review
HiringCreditPricingInsuranceFraudEligibilityAccess to services

Safeguards people are owed

  • Advance notice that an automated system will decide.
  • Meaningful information about the logic involved.
  • A way to contest the outcome and obtain human review.
  • In California, the deck notes an opt-out right for automated decisions.

The trap: rubber-stamp review

  • A human who always clicks approve does not make a process non-automated.
  • The reviewer needs authority, context, and the ability to say no.
  • “Solely automated” is about substance, not the mere presence of a person.
  • Document who reviewed, on what evidence, and what changed.
When the Agent Decides About a Person slide showing the automated decision assessment and safeguards for human review
Putting It Together

Compliance by Design

Retrofitting compliance onto a deployed agent is the expensive path. Design the data flows first; the legal obligations follow them.

IFThe agent will touch personal data
›
Lawful basis, mapped per retrieval
IFSomeone asks to be forgotten
›
Provenance map → every downstream store
IFThe agent decides something significant
›
Notice, human review, right to contest
IFContext leaves your jurisdiction
›
Treat it as a transfer and document it properly
IFConversations are being stored
›
Retention clock, or it becomes a finding
IFYou added a tool or MCP server
›
New sub-processor, new data flow
IFYou are in a regulated sector
›
Sector rules bind before privacy law does

The page preserves the source deck’s legal disclaimer: verify current obligations with counsel because this area changes quickly.

Source Deck

All 10 slide images

Each slide uses responsive image variants from the Slide600, Slide700, Slide900, and Slide1000 folders. Click a slide to open a larger view.

Slide 1 of 10

Click to enlarge
Agent Compliance and Risk slide 1 of 10

Slide 2 of 10

Click to enlarge
Agent Compliance and Risk slide 2 of 10

Slide 3 of 10

Click to enlarge
Agent Compliance and Risk slide 3 of 10

Slide 4 of 10

Click to enlarge
Agent Compliance and Risk slide 4 of 10

Slide 5 of 10

Click to enlarge
Agent Compliance and Risk slide 5 of 10

Slide 6 of 10

Click to enlarge
Agent Compliance and Risk slide 6 of 10

Slide 7 of 10

Click to enlarge
Agent Compliance and Risk slide 7 of 10

Slide 8 of 10

Click to enlarge
Agent Compliance and Risk slide 8 of 10

Slide 9 of 10

Click to enlarge
Agent Compliance and Risk slide 9 of 10

Slide 10 of 10

Click to enlarge
Agent Compliance and Risk slide 10 of 10
Agent Compliance and Risk slide enlarged