Compliance audit of SaleSnail: an AI Instagram DM sales and support chatbot
We used iCOMPLY to audit our own chatbot product -- a RAG-powered conversational AI deployed in Instagram DMs. Here is exactly what we found, including the gaps.
Why we are publishing this
SaleSnail is an AimRank Labs product. When we built iCOMPLY, we committed to auditing our own AI systems first and publishing the results -- including the gaps. This page is the output of an 8-agent onboarding review of the SaleSnail codebase. The transparency gaps are real. The open questions are real. We are publishing them because honest self-assessment is what the EU AI Act expects (Article 43), and because any compliance tool that hides its own shortcomings has no business reviewing yours.
Compliance advisory -- not legal advice. This analysis reflects a draft intake review; it is not a completed Annex IV dossier and has not yet been validated by an AIGP-certified specialist.
System Overview
- System name
- SaleSnail
- Domain
- E-commerce and Customer Engagement
- Description
- AI-native Instagram DM sales and support chatbot. Trains on a business's products and brand in approximately 5 minutes, then responds to DMs 24/7 across four configurable agent types: sales, support, informational, and community. Captures leads, answers product questions, closes sales, and hands off to a human agent when needed. Distributed initially in Lithuania, then Baltic states, then EU.
- EU AI Act role type
- Provider of the AI chatbot product (businesses using it are deployers)
- Deployment
- Python monorepo on AWS. Langfuse for observability. Instagram Graph API (primary channel). Twilio (additional channels).
Model Stack (4 models, no first-party training)
SaleSnail trains no models. Deployer-uploaded knowledge base content is used at inference time only (RAG). Chapter V provider obligations sit upstream with Anthropic and Cohere; SaleSnail inherits documentation-reliance duties (Article 53(1)(b)) and must name them in its Annex IV.
Preliminary Classification
Applicable articles
Classification reasoning
Limited-risk AI system under the EU AI Act. SaleSnail is a conversational AI that interacts directly with natural persons (a business's Instagram audience), triggering Article 50 transparency duties. It is not an Annex III high-risk system (the nearest category, creditworthiness under 5(b), fails: coaching and retail offers are not essential services and no eligibility decision is made) and is not prohibited under Article 5 (persuasion is disclosed and discounts are server-capped). As a downstream integrator of third-party general-purpose models (Anthropic, Cohere), Chapter V provider obligations sit upstream; SaleSnail inherits documentation-reliance duties (Art 53(1)(b)). The dominant residual risk is GDPR, not the AI Act.
Positive controls already in place
- Hard-coded Article 50(1) AI disclosure -- injection-resistant, default-on
- Server-capped discounts -- configurable per deployer, enforced on the server side
- Sub-24h voice-note deletion -- Article 9 mitigation for biometric-adjacent data
- Human handoff path -- deployer receives escalation brief and can take over the conversation
Compliance Gaps Identified
These are real compliance findings from the SaleSnail codebase review. Each maps to a specific article with a severity rating. The first item is a MET control included for completeness -- the three that follow are genuine gaps.
Article 50(1) AI-interaction disclosure: MET. A hard-coded, injection-resistant, default-on prompt rule discloses to end-customers that they are talking to an AI. Action for AIGP reviewer: verify this fires at runtime across every channel and the white-label / generic-API paths.
Article 50(3) emotion-recognition disclosure: GAP. SaleSnail runs text-based frustration and urgency inference and writes the result to handoffs.brief. It does not disclose this inference to end-customers. The legal trigger is contestable -- text is not biometric data -- but the gap is real until the legal question is resolved.
Article 50(2) machine-readable synthetic-text marking: GAP. A conversational 'I am an AI' sentence is human-readable disclosure, not the machine-readable marking Article 50(2) requires from providers of systems that generate synthetic text interacting with natural persons.
GDPR Article 6 lawful basis for end-customer processing: GAP. No consent mechanism exists for DM end-customers. Message content is retained indefinitely. AI profiling (sentiment, urgency) is stored in handoffs.brief with no documented deletion schedule. A DPIA (GDPR Article 35) is required and not yet conducted.
Article 5(1)(b) hardening needed
The sentiment-to-persuasion firewall -- the rule that prevents inferred negative emotion from being used to apply additional persuasion pressure -- is currently an unenforced prompt convention. The discount cap is a server-side invariant with a test. The firewall should be promoted to the same status: a server-side invariant with a tested check. This is a hardening action, not an Article 5 violation: the current risk of actual prohibited-practice behaviour is low. But relying on a prompt to hold an Article 5(1)(b) boundary is not defensible posture.
Open Classification Questions
These questions were surfaced by iCOMPLY's intake analysis and flagged for human review. They require judgment calls that the AI engine deliberately does not make on its own.
- 1.Does text-based frustration and urgency inference legally trigger Article 50(3), or is it out of scope because text is not biometric data?
- 2.Across white-label deployments, is the deployer (the business) or the provider (AimRank Labs) the duty-holder for the 50(1) and 50(3) disclosures?
- 3.Does any agent type ever make an eligibility-like decision that could touch Annex III (essential services / creditworthiness)? Current read: no, because coaching and retail offers are not essential services.
- 4.What is the GDPR lawful basis and retention policy for end-customer DM content, and who is the controller versus processor (the business versus SaleSnail)?
- 5.Does multi-channel expansion beyond Instagram change the Article 50 surface or add subprocessor and transfer risk (Langfuse, Twilio)?
What happens next
- Intake analysis complete -- the system is registered in iCOMPLY with full technical details, 5-vendor subprocessor chain, and 1 deployment record.
- DPIA initiation -- GDPR Article 35 DPIA is required before the product is deployed at scale. Couples to the GDPR Article 6 lawful basis and retention policy gaps.
- Classification conversation -- the 5 open questions above will be resolved in an interactive classify session with the AI engine, producing a definitive risk tier and Article 50 duty-holder analysis.
- Obligation mapping -- once classified, the map conversation produces a prioritised obligation list with effort estimates and deadlines.
- Annex IV documentation -- generate the full 8-section technical documentation as DOCX, with the GPAI upstream reliance chain (Anthropic, Cohere) documented per Article 53(1)(b).
- Expert review + publication -- AIGP-certified specialist reviews all outputs and the full case study is published with real documents attached.
Your AI system has the same gaps
Every conversational AI has Article 50 transparency obligations. Most have GDPR gaps on end-user data. The question is whether you find them now or a regulator finds them after August 2.