What's in a Mosaic assessment.
A walk-through of an Agentic Privacy Impact Assessment for a fictional agent. Every name, finding and number on this page is made up for illustration.
SYNTHETIC EXAMPLE. Not a client, not a real finding.
SYNTHETIC EXAMPLE. Not a client, not a real finding.
One agent, one report.
- Report
- Agentic Privacy Impact Assessment
- Agent
- HR helpdesk agent for a fictional 400-person company. It answers staff questions, reads the shared HR inbox and files leave requests.
- Team
- a lead practitioner and a named senior reviewer
- Typical elapsed time
- build support 4–8 weeks, alongside the build; retrofit of a live agent 2–4 weeks
SYNTHETIC EXAMPLE. Not a client, not a real finding.
Section by section
- 01
Executive summary and launch recommendation
Can it launch, and on what conditions?
- 02
System description and purpose
What is it for, and who uses it?
- 03
Data-flow map and data inventory
What personal information moves where?
- 04
Permission matrix
What can it reach, against what it needs?
- 05
Privacy analysis
Which principles and laws apply, and does it meet them?
- 06
Agent risk analysis
Can it be manipulated, what does it do alone, can you stop it?
- 07
Risk register
Each risk with a rating, owner, treatment and status.
- 08
Control requirements
What your builder must change, and how we will verify it.
- 09
Five-stop evidence table
Which stops exist, and the proof each works.
- 10
Behaviour test results
What our scripted tests tried, and what happened.
- 11
Residual risk
What is left after the fixes, and who accepts it.
- 12
Sign-off page
Where your privacy officer signs.
SYNTHETIC EXAMPLE. Not a client, not a real finding.
Four excerpts
Permission matrix (excerpt)
| Connector | Needed | Granted | Finding |
|---|---|---|---|
| ConnectorHR records (read) | NeededThe asking employee's own record | GrantedAll employee records | FindingOver-scoped · High |
| ConnectorHR inbox (read) | NeededThe shared HR inbox | GrantedThe service account's whole mailbox | FindingOver-scoped · Moderate |
| ConnectorLeave system (write) | NeededCreate a leave request | GrantedCreate, approve and delete | FindingOver-scoped · High |
Risk register (excerpt)
| ID | Risk | Rating | Owner | Treatment | Status |
|---|---|---|---|---|---|
| IDR-03 | RiskInstructions hidden in an emailed attachment could make the agent approve leave | RatingHigh | OwnerBuild lead | TreatmentRemove the approve permission; require a person to approve | StatusClosed on retest |
| IDR-07 | RiskThe agent repeats another employee's salary when asked indirectly | RatingHigh | OwnerBuild lead | TreatmentLimit HR records to the asker's own file | StatusClosed on retest |
Five-stop evidence (excerpt)
| Stop | Evidence | Result |
|---|---|---|
| StopIdentity | EvidenceAgent runs under its own service identity; permission export reviewed | ResultProven |
| StopPolicy | EvidenceData-loss rule blocks Social Insurance Numbers in outbound messages | ResultProven |
| StopRuntime | EvidenceNo documented way to pause the agent | ResultNot shown → condition 1 |
| StopMonitoring | EvidenceActions logged; nobody assigned to review them | ResultPartial → condition 2 |
| StopA person | EvidenceLeave approvals now need a manager | ResultProven |
Launch recommendation (excerpt)
Launch with two conditions: a documented pause procedure, and a named owner for weekly log review. Both verified closed on retest.
SYNTHETIC EXAMPLE. Not a client, not a real finding.
Every risk on the same scale.
High, Moderate or Low, rated on likelihood and impact, the same way on every engagement. Our privacy impact assessments use the same skeleton and the same scale.
SYNTHETIC EXAMPLE. Not a client, not a real finding.