# Expanding Marketing AI to Sales Support: An 8-Point Risk Audit > Marketing AI maximizes public reach, but sales AI demands strict isolation. Use this 8-point technical checklist to prevent pricing leaks and prompt injection. Source: https://www.starbox.work/en/insights/geo-guide/2026-10-25-zhaoshang-kefu-ai-risk-checklist/ (Starbox AI Insights, updated 2026-10-25) ## Why Marketing AI Corpora Cannot Power Transaction-Grade Customer Service Public marketing AI succeeds by maximizing information density, open indexing, and citation frequency across search platforms and large language models. The goal is broad visibility: making product facts, brand narratives, and technical specifications as accessible as possible to AI search crawlers. Expanding that infrastructure directly into franchise acquisition, presales qualification, and customer service introduces immediate security vulnerabilities. Sales and support workflows operate on entirely different business logic. A presales or franchise agent handles contract templates, confidential tier-pricing, partner rebate structures, and operational commitments. If you plug a dialogue interface directly into your existing marketing knowledge base without architectural boundaries, you expose your negotiation parameters to basic prompt injection attacks. Marketing AI optimizes for retrieval recall. Sales and customer service AI must optimize for privilege boundaries, deterministic authorization, and verifiable auditability. Before connecting marketing AI pipelines to transactional touchpoints, technical directors must establish physical isolation and dynamic role-based retrieval. ``` +-------------------------------------------------------------------------------+ | ARCHITECTURAL SEPARATION | +-----------------------------------+-------------------------------------------+ | Public Marketing AI | Private Transaction / Sales AI | +-----------------------------------+-------------------------------------------+ | Objective: Maximum reach & recall | Objective: Zero leak & strict boundaries | | Vector Store: Public product data | Vector Store: Tiered commercial data | | Retrieval: Open semantic search | Retrieval: Dynamic RBAC token filtering | | Guardrails: Brand tone checks | Guardrails: Deterministic hard filters | +-----------------------------------+-------------------------------------------+ ``` --- ## The 4-Layer Physical Isolation Architecture A unified company database shared between marketing and transactional bots creates catastrophic leakage risk. You must enforce physical and logical separation across four distinct layers: ### 1. Storage and Vector Store Isolation Never rely on metadata flags alone (such as `is_public: false`) inside a shared vector collection to separate marketing assets from sales policies. Maintain physically isolated vector databases or discrete tenant-level namespaces. Public marketing materials reside in an externally crawlable retrieval index; wholesale discount schedules, partner rebate matrices, and internal dispute guidelines belong in an encrypted internal cluster inaccessible to external session tokens. ### 2. Retrieval-Augmented Generation (RAG) Partitioning Dynamic authentication must precede vector similarity scoring. When a user queries the system, the middleware must validate their authenticated role before querying the index. If an unverified web visitor asks about franchise rebates, the retrieval pipeline must not query the commercial policy repository at all. By preventing sensitive chunks from ever entering the prompt context window, you eliminate the risk of context-leaking prompt injections. ### 3. Execution Context and Session Segregation Customer-facing dialogue runs within ephemeral containerized environments. Marketing discovery sessions, presales evaluations, and post-purchase customer service tickets must run on separate session runtimes. An unauthenticated web visitor chatting on your landing page must never share runtime memory, caching layers, or conversation buffers with an enterprise client negotiating a regional franchise agreement. ### 4. Downstream API and Tool Isolation If your customer service bot connects to downstream enterprise resource planning (ERP) or customer relationship management (CRM) systems via function calling, isolate write privileges behind strict service boundaries. Presales bots should hold zero write permissions; they may query predefined product availability endpoints via read-only scopes, while transactional updates require human confirmation or dedicated microservices. --- ## Adversarial Defense: Intercepting Price Leaks and Policy Extraction System prompts instructing an LLM to "never reveal wholesale prices" fail under pressure. Attackers use role-play scenarios, indirect context injection, base64 encoding, and system instruction overrides to bypass soft prompt constraints. To protect your bottom line, deploy deterministic guardrails that operate independently of model discretion: ``` [User Input] │ ▼ [Layer 1: Input Regex & Semantic Intent Classifier] ──(Violation)──► [Block & Log] │ (Clean) ▼ [Layer 2: Authenticated Retrieval (RBAC-Filtered Context)] │ ▼ [Layer 3: Model Inference Execution] │ ▼ [Layer 4: Output Guardrail (Deterministic Regex / Masking)] ──(Violation)──► [Fallback] │ (Clean) ▼ [User Output] ``` ### Deterministic Interception vs. Model Discretion Place a dual-stage filter around the inference engine. - **Inbound Filter:** Screen incoming prompts for known injection signatures, system prompt extraction patterns, and role-override commands. - **Outbound Filter:** Screen raw model outputs with deterministic regex patterns that identify phone numbers, internal policy codes, margin thresholds, or unauthorized numerical ranges before the response reaches the user. ### Red-Teaming Core Commercial Vulnerabilities Conduct routine blue/red team adversarial drills against three primary targets: - **Wholesale Price Floor:** Adversaries simulate executive emergencies or system maintenance modes to extract base distribution costs. - **Franchise Rebate Schedules:** Competitors craft nested conditional queries to deduce tier-two distributor profit margins. - **Guarantee and Liability Commitments:** Attackers push the model to promise contractual service level agreements (SLAs), return windows, or financial guarantees that exceed standard terms. --- ## Infrastructure Strategy: Self-Hosted Models vs. Commercial APIs Deploying customer service AI across transactional touchpoints requires balancing token latency, operational cost, and compliance obligations. | Evaluation Metric | Public Commercial APIs | Self-Hosted / Private VPC LLMs | | :--- | :--- | :--- | | **Data Privacy** | Subject to vendor data policies; potential external caching | Complete tenant control; zero external data transit | | **Concurrency Cost** | Linear operational cost per token; vulnerable to query spikes | Fixed hardware/instance cost; high volume lowers per-query cost | | **Latency Consistency** | Variable based on vendor network load and regional endpoints | Highly predictable; optimized for dedicated workloads | | **Audit Log Control** | Vendor-dependent log retention; metadata often truncated | Raw token-level storage; immutable logging for compliance | | **Best-Fit Workload** | Top-of-funnel discovery, multi-turn public marketing chat | Presales pricing negotiation, contract support, claims | A hybrid deployment minimizes total cost of ownership while securing core assets. Deploy lightweight, managed commercial APIs at the top of the funnel to handle general inbound traffic and marketing inquiries. Route authenticated, sensitive franchise negotiations and post-sale account management through privately hosted models or secure Virtual Private Cloud (VPC) instances with end-to-end encryption. --- ## The 8-Point Risk Audit Checklist for Technical Directors Before signing off on migrating marketing AI to franchise and customer support channels, complete this verification audit: ``` +------------------------------------------------------------------------------------+ | TECHNICAL SECURITY AND ISOLATION VERIFICATION AUDIT | +----+-------------------------------------------------------------------------------+ | [ ] 1. Vector Database Physical Separation | | Confirm marketing data and sales policies reside in isolated instances. | | [ ] 2. RBAC-Coupled Context Injection | | Verify retrieval pipelines filter source chunks by verified session token. | | [ ] 3. Deterministic Inbound Prompt Sanitization | | Test that prompt injection vectors are dropped before reaching the LLM. | | [ ] 4. Outbound Sensitive Data Masking | | Validate hard regex filters that block unauthorized pricing terms. | | [ ] 5. Function Call Scope Restriction | | Confirm presales bots hold read-only scopes on internal microservices. | | [ ] 6. Dual-Track Inference Routing | | Verify public inquiries and sensitive transactions use split model paths. | | [ ] 7. Immutable Audit Trail Logging | | Ensure full prompt-response pairs, vector IDs, and user IDs are archived. | | [ ] 8. Human-in-the-Loop Contract Escalation | | Enforce hard fallbacks that block automated contract commitments. | +----+-------------------------------------------------------------------------------+ ``` ### 1. Vector Database Physical Separation - **Verification Method:** Inspect the database configuration. Confirm that marketing knowledge chunks and commercial policy chunks reside on distinct physical servers or isolated VPC namespaces with separate access credentials. - **Remediation:** Split unified collections into distinct repositories. Revoke shared API keys across your marketing and sales infrastructure. ### 2. RBAC-Coupled Context Injection - **Verification Method:** Attempt to retrieve commercial policy documents using an unauthenticated visitor session token via the retrieval API. - **Remediation:** Bind retrieval filtering to user session identities. If the user lacks an authorized role header, return an empty context set before vector similarity matching begins. ### 3. Deterministic Inbound Prompt Sanitization - **Verification Method:** Execute automated red-team suites containing character escapes, base64 encodings, and system prompt override payloads against input endpoints. - **Remediation:** Install an open-source or custom guardrail proxy to reject malformed inputs and known prompt injection strings prior to LLM processing. ### 4. Outbound Sensitive Data Masking - **Verification Method:** Craft prompts specifically designed to induce the model into repeating confidential margin calculations or internal policy numbers. - **Remediation:** Deploy post-processing token scanners that evaluate the generated text against deterministic blacklists, replacing matched numbers or phrases with canned safe responses. ### 5. Function Call Scope Restriction - **Verification Method:** Review API tokens assigned to customer service agent tools. Attempt to execute state-changing or write operations through the bot's interface. - **Remediation:** Downgrade agent tool permissions to least-privilege read status. Require multi-factor human authorization for order cancellations, refund approvals, or contract generations. ### 6. Dual-Track Inference Routing - **Verification Method:** Trace outbound network traffic from your customer service application during peak user simulations. - **Remediation:** Configure API gateway rules that route top-of-funnel informational queries to external commercial endpoints, while redirecting sensitive financial and contractual queries to internal VPC instances. ### 7. Immutable Audit Trail Logging - **Verification Method:** Run sample dialogues and review database write operations. Verify whether session ID, user identity, raw input, retrieved chunk IDs, and model outputs are recorded in write-once-read-many (WORM) storage. - **Remediation:** Pipe interaction telemetry directly into an isolated security information and event management (SIEM) system. Decouple logging agents from the chatbot runtime. ### 8. Human-in-the-Loop Contract Escalation - **Verification Method:** Prompt the system to finalize an unstandardized discount, agree to penalty terms, or append supplemental contract riders during a mock presales dialogue. - **Remediation:** Embed strict intent triggers that terminate the bot's conversational loop and route the ticket to a human sales representative whenever a user requests contractual commitments or bespoke pricing terms.