Systems · 8 min read

Shadow AI is already in your marketing stack

The risk is rarely one dramatic autonomous agent. It is the growing collection of prompts, browser extensions, exports, and automations that handle prospect or customer context without a shared contract.

Key idea

Start with visibility, not a ban. Map where data enters AI workflows, what leaves them, who owns the outcome, and how an exception is handled.

The invisible workflow problem

A recent HN-linked discussion raised the risk of AI hidden inside marketing stacks. That is the right framing: the issue is not whether a team uses AI. It is whether the team can explain what a workflow sees, does, retains, and changes.

Map the exposure surface

WorkflowTypical riskControl to add
Copy generationUnverified claims or restricted context in prompts.Approved input sources and a review rubric.
Account researchStale or scraped facts treated as truth.Source links, freshness rules, and confidence states.
Lead routingOpaque decisions or incorrect ownership.Explainable rules, fallback queues, and audit logs.
Campaign activationUncontrolled volume or suppression failures.Permissions, approval gates, and stop conditions.

Build the register

  1. Inventory real workflows.

    Ask operators what they use today — including spreadsheets, personal accounts, extensions, and “temporary” automations.

  2. Classify the data.

    Record whether a workflow touches public, prospect, customer, confidential, or restricted information.

  3. Name an owner.

    Every workflow needs one accountable person who can answer for its purpose and result.

  4. Set a permission boundary.

    Separate research, drafting, recommendation, and external action. They should not inherit the same authority.

Govern the outcome, not the model

Model choice matters, but governance begins with the business outcome. If an AI workflow can change a CRM record, contact a prospect, or inform a segmentation rule, log the input references, workflow version, output, reviewer action, and final result.

This creates the feedback loop that lets a team expand useful automation while finding weak or unsafe patterns early.