ai automation pre deployment risk auditor

AI AUTOMATION PRE-DEPLOYMENT RISK REPORT

Automation: n8n Lead Qualification → HubSpot → Sales Follow-up

Execution Status: NOT PERFORMED

No live workflow, API, CRM, credentials, or external system was executed or accessed. Findings are based only on the supplied workflow description.

Executive Summary

The workflow has several material reliability and control risks that should be addressed or verified before production deployment.

The strongest concerns are CRM duplication during retries or concurrent submissions, insufficient validation of AI-generated classification decisions, and unverified handling of untrusted customer input.

The supplied evidence is not sufficient to establish that the workflow is production-ready.

Deployment Decision: FIX BEFORE DEPLOYMENT

1. CRM Retry / Idempotency Risk

Status: POTENTIAL
Severity: P1 — HIGH-IMPACT RISK

The workflow retries failed HubSpot API requests up to three times, but no documented idempotency mechanism was supplied.

A timeout can occur after HubSpot successfully commits a CRM write but before the workflow receives the response. The workflow could then retry the operation.

A separate concurrent-submission scenario could also cause two executions to independently determine that a contact does not exist before both attempt to create it.

The evidence does NOT establish that duplicate records currently occur. The exact retry scope and HubSpot operation behavior are also not fully supplied.

Required test:

1. Submit a lead.
2. Force a timeout immediately after the CRM write commits.
3. Allow the workflow retry.
4. Verify whether one or multiple CRM records are created.

Also test two identical submissions concurrently.

Recommended fix:
Use an idempotency key or deterministic unique identifier and enforce uniqueness at the CRM/data layer where possible. Do not rely solely on a search-then-create sequence.

2. AI Classification Reliability

Status: UNKNOWN / POTENTIAL
Severity: CRITICAL UNKNOWN

The AI determines whether a lead is QUALIFIED or NOT QUALIFIED, which affects the downstream sales workflow.

The supplied evidence confirms that the response is checked for valid JSON and required fields, but this does not establish that the classification itself is correct.

No evaluation dataset, accuracy measurement, confidence threshold, ambiguity handling, or representative test results were supplied.

Required verification:
Test representative qualified, unqualified, ambiguous, incomplete, adversarial, and edge-case messages.

Measure classification accuracy and identify false-qualified and false-unqualified cases.

Do not treat syntactically valid JSON as evidence that the AI decision is correct.

3. Customer Input / Prompt Injection

Status: POTENTIAL
Severity: P1 — HIGH-IMPACT RISK

The customer's project description is supplied to an AI model.

Because this field is user-controlled, a malicious customer could attempt to manipulate the model's instructions or classification behavior.

The supplied evidence does not show the actual prompt or demonstrate whether prompt-injection defenses exist.

Required verification:
Test messages containing instructions such as attempts to override the classifier, alter the expected output, inject structured data, or manipulate the qualification result.

Review the actual system/developer prompt and enforce strict separation between instructions and untrusted customer content.

4. AI Output Validation

Status: OBSERVED + POTENTIAL

The workflow validates JSON syntax and required fields.

This is useful structural validation, but it does not prove semantic correctness.

For example, a response could contain valid JSON with an incorrect budget, incorrect project type, or incorrect QUALIFIED classification.

Recommended fix:
Add semantic validation for important fields, allowed-value validation for classifications, range/type checks where applicable, and explicit handling for ambiguous or low-confidence results.

5. Human Review Control

Status: OBSERVED

A human reviews qualified leads before the sales team contacts them.

This reduces the risk of an incorrect AI classification directly reaching the sales team.

However, the evidence does not establish whether the human review is technically enforced or whether another automated action occurs before review.

Verify that no consequential downstream sales action bypasses the review gate.

6. Critical Unknowns

The following information should be verified before declaring the workflow production-ready:

• Exact AI system/developer prompt
• AI model and configuration
• AI classification evaluation results
• Exact HubSpot permissions/scopes
• Exact retry behavior and which operations are retried
• Idempotency or duplicate-protection mechanism
• Behavior during timeout-after-commit
• Concurrent duplicate submission behavior
• Confirmation email template and data mapping
• Error-handling and recovery behavior
• Monitoring and alerting configuration

Recommended Priority

P0: None demonstrated from the supplied evidence.

P1: CRM retry/idempotency risk; untrusted customer input requires verification.

Critical Unknown: AI classification reliability and several production controls cannot currently be verified.

Final Recommendation

FIX BEFORE DEPLOYMENT

Do not treat the workflow as unsafe merely because evidence is missing. However, the current evidence is insufficient to establish reliable protection against duplicate CRM writes and incorrect AI-driven decisions.

Resolve the critical unknowns and run the recommended adversarial and failure-mode tests before client handoff.

Audit confidence: LOW–MEDIUM

Execution performed: NO