Keep client data separate in automated agency reports
Prevent cross-client report mixups with an explicit account map, isolated draft inputs and a synthetic two-client acceptance test.
Keep each client's report tied to an approved map of source accounts, working data, draft location and final audience. Check that identity at every stage, before AI receives the inputs and before a report leaves review. A familiar client logo or a correct total cannot prove that the underlying records belong to the right client. This guide gives small marketing agencies a practical way to test that boundary before connecting live reporting data.
Define a client boundary that survives a renamed campaign
Start with a short account map owned by your reporting lead. For each client, record a stable internal client reference, permitted source account identifiers, approved report destination and named reviewer. Keep passwords and access tokens out of this map. Have someone who knows the account check it against the actual connected systems before the first run.
Campaign names and file names can help a person recognise work, but they should not decide which client owns a record. Two clients might both have a campaign called Spring Launch. Croissant recommends using the approved source-account mapping to determine ownership, then preserving that client reference through collection and report assembly. Missing or conflicting ownership should create a review item rather than select the nearest name.
OWASP recommends denying access unless it is permitted and checking authorization on every request. [1] For this reporting workflow, ask the builder where those checks run and how they use the approved client map. A client identifier supplied in an input is something to validate against authorized access, rather than proof that access is allowed.
Separate the inputs before asking AI to write
For a client-specific draft, give the writing step only the approved facts for that client and reporting period. Croissant recommends a fresh working context for each client run, with client-specific retrieval, temporary files and cached results. Ask the implementer to demonstrate that isolation; a new chat label alone does not establish how connected storage behaves.
Reusing the layout can be useful. Reusing a previous client's filled report as the next draft can carry over a budget, recommendation, chart or source link. Keep a clean template separate from completed reports. If an example paragraph is needed, use clearly synthetic text that cannot be mistaken for either client's current performance.
Check intermediate outputs as well as the final document. A final client filter cannot undo another client's information already included in a model request or exposed in a shared review folder. The broader agency reporting workflow covers collection and factual reconciliation; this guide adds a specific check that every input and output belongs to the intended client.
Worked example: two clients, the same campaign name
This is a synthetic acceptance test, not a customer result. Fictional Client A owns test account A-101 and has USD 420 in spend. Fictional Client B owns test account B-202 and has USD 730 in spend. Both contain a campaign named Spring Launch. Put the distinctive fictional phrase BLUE KITE ONLY in B's notes to make accidental carryover visible. Run the test only in an authorized sandbox with external delivery disabled.
The approved A report should contain USD 420, reference A-101 and contain neither B's note nor B's source links. A report showing USD 1,150 fails even though the addition is correct: it combined different clients. A report showing USD 420 with B's commentary also fails. The test checks ownership of the narrative and references as well as the numbers.
Next, deliberately map a test source row from B-202 into A's collection. The expected outcome is a blocked assembly step with a visible ownership mismatch for the reporting lead. The row must not reach the writing step. Record the actual input set and output evidence; a polished refusal written after the data was already exposed does not satisfy this condition.
Finally run B and then A using the same worker, and repeat A after a simulated collection failure. A must not inherit B's cached figures, notes or draft destination. These cases are Croissant's suggested starting checks. Passing them supports a bounded pilot, but cannot certify that every possible route is secure.
Test the audience, export and review path
Ask the reviewer to open each test report using the intended audience's permissions. An administrator seeing both clients proves very little about what a client viewer can access. Check the report link, downloadable file and review folder separately. A correct dashboard view does not establish that a copied attachment has the same access rules.
If your implementation uses Power BI, Microsoft documents a specific limit: row-level security restricts workspace Viewers, while Admin, Member and Contributor roles are outside that restriction. Microsoft also provides a Test as role function. [2] Have the implementer demonstrate the actual audience configuration and supported test method. This is a Power BI rule, not a claim about every reporting tool.
For the synthetic A reviewer, attempt to open B's report through an ordinary saved test link and request B's export within the authorized test setup. Both should be denied under the proposed client boundary. Include a stale A link after test access is removed. Avoid tests against real clients or resources you do not own or have permission to assess.
Keep client distribution with an authorized person. Before release, that person checks the intended client, exact report version and approved audience together. If a draft changes after approval, send the new version through review again. This content checklist does not authorize automated email delivery or changes to anyone's permissions.
Decide what evidence you need before buying a build
Ask a supplier for the account map, a diagram of where client identity is checked, and recorded outcomes from the two-client test. Request examples of a rejected source mismatch and a denied report request alongside the successful report. OWASP recommends protecting static resources and testing authorization logic, while noting that routine tests do not replace specialist security testing. [1] Include exported files in your review rather than stopping at database permissions.
Name an agency owner for mapping changes. Adding a source account, moving a review folder or replacing a report template should trigger the relevant isolation checks again. If a mixup is found, pause the affected report path, restrict the affected output and involve the responsible account and security owners. Do not ask AI to decide whether disclosure occurred or which notification obligations apply.
If your existing reporting product can enforce and demonstrate these boundaries, use its supported configuration before commissioning extra machinery. A custom workflow is worth discussing when separate tools leave an unresolved collection or review gap. Explore the reporting demo for the general preparation pattern, then use the Croissant contact form to describe your systems and desired checks. Bring synthetic examples rather than private client files.
Implementation checklist
- Approve a stable client-to-source-account map and name its owner.
- Keep client identity attached to collected facts, drafts and destinations.
- Use a clean template and client-specific working inputs.
- Test two clients with identical campaign names and distinctive fictional notes.
- Block mismatched source ownership before the writing step.
- Check sequential runs, stale cached results and collection failures.
- Test actual viewer access, exports and review folders in an authorized sandbox.
- Approve the report version and audience together before distribution.
- Repeat relevant isolation checks when mappings, tools or destinations change.
The sources behind the sauce.
Primary research and official documentation behind the cited points. The workflow examples and checklists are Croissant’s own guidance.