Most discussions about AI agents in business treat each industry as a separate problem. Legal tech has its tools. FinTech has its automation stack. Healthcare has its compliance concerns. Customer support has its helpdesk platforms. In practice, when you look at the actual workflow steps that agent skills perform across all of these verticals, the same seven patterns keep appearing — with different data, different risk boundaries, and different approval gates, but the same underlying structure.

This matrix maps those repeating patterns across ASE’s industry skill collections and explains why understanding cross-vertical workflow anatomy makes you a better skill creator, a sharper skill consumer, and a more realistic buyer of AI automation claims.

The Seven Workflow Patterns

After reviewing the skills across ASE’s nine industry collections — Legal Ops & Compliance, Finance & Filings, Healthcare Documentation & Intake, Customer Support & Success, Real Estate Workflows, Ecommerce & Retail Operations, Product Analytics & Growth Ops, DevRel & API Documentation, and Media & Publishing — seven workflow steps appear in almost every domain:

1. Intake — Accepting raw input: a document, a ticket, a form, an audio file, a request
2. Extraction — Pulling structured data from unstructured content: OCR, table parsing, entity recognition, transcription
3. Review — A human-readable summary or comparison: diff, redline, summary, comparison
4. Routing — Deciding where something goes next: triage, prioritization, escalation logic
5. Reporting — Producing a structured output: a packet, a dashboard slice, a formatted file
6. QA — Verifying that something is correct before it moves forward: smoke test, checklist, evidence capture
7. Escalation — Flagging for human review when the agent hits a boundary: approval gates, ambiguity handlers, confidence thresholds

The matrix below shows which of these patterns each industry collection actively covers, and where the risk boundary sits for each.

The Industry Skills Matrix

✅ = pattern well-covered by existing skills | ⚠️ = partial coverage or elevated risk boundary — human checkpoint required before proceeding

What Changes Across Verticals: Risk and Data Shape

The workflow steps are similar. What differs significantly across industries is two things: the risk profile of each step and the shape of the data being processed.

Risk Profiles

In ecommerce, routing a return request to the wrong queue costs a few minutes and a customer service email. In healthcare, routing a clinical intake document to the wrong destination can have real consequences. In legal, auto-routing a contract to signature without a review step may create liability. The same “routing” step carries fundamentally different risk depending on where it sits.

This is why the ⚠️ symbols in the routing and QA columns above cluster in regulated industries. Those aren’t gaps in skill coverage — they’re intentional checkpoints. A skill in the Legal Ops collection that routes documents will surface a summary and flag for human review rather than fire off an action autonomously. That’s the right design for the domain.

Data Shape

The extraction pattern looks different depending on what you’re extracting:

• Finance & Filings: Semi-structured PDFs (EDGAR filings, invoices), tables, known field positions
• Healthcare Documentation: Handwritten notes, dictation audio, intake forms, lab report PDFs with variable structure
• Legal Ops: Long-form contracts, clause libraries, amendment trails
• Customer Support: Conversational text, ticket threads, CRM records — not primarily document extraction
• Media & Publishing: Audio/video files, WebVTT subtitle tracks, chapter markers

The underlying tools differ — Tesseract for scanned documents, WhisperX for audio, PyMuPDF for digital PDFs, table parsers for structured filings — but the workflow step (accept input → pull structure out of it) is the same. This is why ASE avoids tagging a generic OCR skill as “healthcare” without also checking that its gotchas cover healthcare-specific edge cases like handwriting variability and multi-column forms.

The Pattern You Should Bet On: Intake → Extract → Evidence Packet → Human

If you zoom out across all nine industry collections, the most reliable workflow sequence that shows up in high-quality industry skills is:

1. Accept raw input (document, ticket, media file)
2. Extract structured fields with source references
3. Assemble an evidence packet (structured summary with source links and confidence indicators)
4. Route to a human decision point before any state change occurs

This sequence appears in the Finance & Filings invoice reconciliation flow (ingest vendor invoice → extract line items → produce variance report → approval before posting). It appears in the Legal Ops contract review flow (ingest PDF → extract clauses → produce redline summary → route to lawyer). It appears in Customer Support (receive ticket → fetch CRM context → draft response with knowledge-base citations → queue for agent review).

The teams getting useful work out of agent skills in regulated industries are not trying to eliminate the human step. They’re trying to make everything before it faster, more reliable, and better evidenced.

How ASE Tags Industry Skills Without Creating a Junk Drawer

One editorial challenge with industry skill collections is noise. It’s easy for a generic PDF-to-text tool to acquire a “healthcare” tag just because someone in healthcare used it. That kind of drift makes industry collections useless — they become indistinguishable from the main catalog.

ASE applies a four-part test before tagging a skill with an industry collection:

1. Domain objects: Does the skill description reference domain-specific entities? (clinical notes, EDGAR filings, MLS listings, merchant invoices)
2. Domain workflow: Does it describe a workflow step that is meaningful in that industry context? (not just “extract text” but “extract invoice line items for three-way match”)
3. Risk or compliance context: Does it acknowledge the elevated consequences of failure in that domain? (HIPAA-sensitive fields, legally binding documents, financial reconciliation)
4. Domain-native integrations: Does it connect to tools that practitioners in that industry actually use? (Epic, Stripe, Salesforce, WooCommerce, court filing systems)

A generic Tesseract OCR wrapper fails this test for healthcare. A skill that wraps Tesseract specifically for multi-column intake forms, includes a gotchas section about handwriting variability, and routes output to a structured intake schema — passes.

Using the Matrix as a Gap Finder

If you are building a skill for a regulated industry, use the matrix as a quick gap check. Before publishing:

• Is your skill handling intake robustly? Does the SKILL.md explain what input formats it accepts and which it cannot handle?
• Does extraction produce source references — not just extracted values, but evidence of where they came from?
• Does the review output include a human-readable summary rather than just a data blob?
• Does routing or escalation include an explicit checkpoint — a step where the skill surfaces its output and waits rather than acting?
• Is QA covered — either by verifying the skill’s own output before surfacing it, or by producing an evidence artifact that makes downstream QA fast?

The gaps most commonly found in rejected or low-performing industry skills are: missing escalation paths (the skill takes action when it should have asked), missing source references in extraction outputs, and no differentiation between high-confidence and low-confidence results.

Cross-Industry Skill Reuse: Where It Works and Where It Breaks

Some skills are legitimately cross-industry because the workflow step is domain-neutral. A well-built PDF text extraction skill using PyMuPDF does the same job whether the input is a lease agreement or an EDGAR 10-K. A WhisperX transcription skill works for both a medical dictation and a podcast episode.

Where cross-industry reuse breaks down is at the edges: when the skill needs to know what to do with ambiguous results, how to handle sensitive fields, or which downstream system to connect to. A generic transcription skill cannot know that a phrase in a medical dictation should trigger a review flag. That’s where you need a domain-specific wrapper skill that uses the generic tool but adds the industry context in its gotchas, its escalation logic, and its output schema.

The Media & Publishing collection and the Healthcare Documentation collection both use transcription skills — but the healthcare version adds field-level sensitivity flags, and the media version adds chapter-marker generation and subtitle alignment. Same core tool; very different wrapping.

What This Means for Skill Creators

If you are building a new skill for any of ASE’s industry collections, the most useful framing is: which of the seven workflow steps does my skill cover, and where does it deliberately stop and hand off to a human?

A skill that clearly owns its step in the sequence — and is explicit about its boundaries — is far more useful than a skill that implies end-to-end automation. The former integrates cleanly into a real workflow. The latter creates false confidence and gets abandoned when it breaks at the boundary it pretended not to have.

For cross-industry skills that can be reused, make the domain-neutrality explicit in the description. Let the consuming team know what they need to add in their own wrapper. That kind of honest scoping turns a narrow tool into a trusted building block.

Browse ASE’s full set of industry collections at agentskillexchange.com/industry-skills/, and explore cross-industry skill patterns alongside the catalog’s general skill guides and tutorials.