Builds the inventory of AI capability an organization is actually running, across the five ways it enters. Four of those five leave no transaction behind: tools built in-house on a subscription already held, vendor features switched on inside approved products, third-party integrations attached to approved platforms, and free tools staff signed up for with a work email. Classifies entry path, evaluates ePHI contact and third-party disclosure, and produces structured findings on ownership, agreements, verification, and risk-analysis scope.

From document to output

The interview comes first

Not the IT report. Gap-driven questions to staff about workflow pressure, in-house shortcuts, shared prompts, and AI-assisted authoring surface what central administration structurally cannot see

Supplied artifacts corroborate

A vendor or BAA register, software spend export, or platform app directory confirms rows the interview produced. None of them is required to start, and none of them is treated as the source of truth

The three invisible paths are probed

Vendor features activated inside approved products, integrations attached to approved platforms, and free third-party signups each get their own pass, because an inventory built from purchasing records looks complete and is wrong

Every tool is classified and scored

Entry path, the operator whose systems actually process the data, scopes held, and ePHI contact rated confirmed, likely, or unlikely based on what the tool processes rather than what it is licensed to do

Findings state the defensibility gap

Four dimensions per tool: executed agreement, named owner, verification against a graded set, and presence in the current risk analysis. Each finding carries severity, consequence, ordered recommendations, and the CFR provision implicated

What you get back

Every finding is structured JSON: status, evidence, risk level, and remediation in one package. No interpretation required before acting on it.

Example response
{
  "tool_name": "Referral summarizer (informal)",
  "finding_type": "inventory",
  "severity": "critical",
  "finding": "Identifiable patient information is being transmitted to a model provider on a consumer-tier subscription with no Business Associate Agreement, and the workflow does not appear in the risk analysis.",
  "why_it_matters": "The obligation to identify systems handling ePHI attaches to the data rather than to how the system entered. A workflow that was never purchased is still in scope, and it currently sits outside both the contractual and the risk-analysis perimeter.",
  "recommendations": [
    "Move the workflow to an enterprise tier covered by an executed BAA, or pause it, before the next referral batch",
    "Add the workflow to the risk analysis asset list with its data flow documented",
    "Name an individual owner in the intake function",
    "Build a graded set of referral packets already summarized by hand and verify output against it before resuming at volume"
  ],
  "citation": "45 CFR 164.308(a)(1)(ii)(A); 45 CFR 164.502(e)"
}

The manual workflow the skill runs

Nothing here is hidden behind the automation. This is the process a practitioner would follow by hand, in the order the skill follows it. Read it to check the reasoning, or run it yourself without an agent.

Download the workflow (PDF)

Phase 1: Zero-ticket discovery

Run this entirely on your own authority, without requesting elevated IT access.

The interview, your primary instrument

Do not start with IT reports. Ask staff, framing questions as fixing a control gap rather than investigating misconduct. Probe functions with high document volume (intake, prior authorization, billing) for automated steps. Ask who is known for building their own shortcuts, since in-house tooling clusters around one or two people. Ask whether anyone maintains custom instructions or shared prompts colleagues use, because a shared prompt is a system. Ask whether any policies, compliance reports, or patient-facing letters were drafted with generative assistance.

Self-inspection checks

Walk through these with staff during the interview. Have them open their Microsoft My Apps portal or Google linked-apps page and record what is connected. Have them check chrome://extensions or edge://extensions for anything that reads page content or the clipboard. Have them search their work email for "verify your email", "welcome to", and "confirm your account", which catches the email and password signups that OAuth logs structurally miss.

Vendor coverage gap identification

For every approved product not acquired for its AI (EHR, practice management, transcription, phone systems), ask whether you know exactly which AI features are enabled on your tenant, what data they process, and which subprocessors are involved. If you cannot answer confidently, you have found a gap. Write to the vendor for written confirmation.

Phase 2: Artifact harvesting

Use these to corroborate the interview, not to replace it.

Vendor or BAA register

Confirms the relationships already recognized.

Software spend export

Identifies licensed and expensed tools, and misses every free tier.

Platform app directories

Identifies integrations attached to approved platforms.

Phase 3: Classification and evaluation

For every tool discovered, run all three evaluations.

Classify the entry path

Licensed purchase, acquired through standard procurement. Built in-house, assembled on a model subscription already held. Feature activated, shipped inside an approved product. Integration attached, connected to an approved platform and often with no agreement. Free third-party tool, signed up for with a work email or on a free tier.

Evaluate ePHI contact

Judge what the tool actually processes, not what it is licensed to do. Confirmed means documented processing of identifiable health information. Likely means it processes a document class, or holds scopes, that routinely reach ePHI. Unlikely means it operates only on material with no path to identifiable health information.

Assess defensibility across four dimensions

Is an agreement in place with the operator whose systems actually process the data? Is there a named owner, an individual rather than a department? Has the tool been verified against a graded set? Is it inside the organization's current risk analysis scope?

Phase 4: Producing findings

Critical

ePHI contact is confirmed or likely and no agreement is in place, or the tool is missing from the risk analysis.

High

ePHI contact is confirmed or likely, but there is no named owner or no verification.

Authoring disclosure

Applied retrospectively where documents were drafted with AI on a consumer tier.

The incident tripwire

If you find that ePHI was transmitted to an external system with no agreement in place, pause the assessment. That question is for qualified legal counsel, and this workflow does not answer it.

Two ways to run this skill

Install on ClawHub

One-click install for Claude and Cowork users. The skill runs as a native tool in your Claude environment — no configuration, no MCP setup. Apache 2.0.

Get on ClawHub →
AI Tooling Inventory SKILL.md

Install as an MCP server in Claude Code, or download the SKILL.md and use it as a system prompt with any LLM. Runs the full methodology against documents you provide. Apache 2.0. No account required.

12 more skills in the methodology.

HIPAA Gap Analysis, BAA Review, Framework Mapping, Control Assessment, Risk Assessment, AI Claims Assessment, Part 2 Notice Assessment, Compliance Q&A, Compliance Posture Intake, Document Finder, Legal Page Discovery, and Skill Router. All open source under Apache 2.0.