Vertical Intelligence: Give Your Agents Domain Expertise

There's a common misconception about how AI voice agents work in complex service industries: that a good general-purpose LLM, given a brief description of the business, will figure out the rest.

It won't. Not reliably. Not for the things that matter.

A dental receptionist who's been at the front desk for three years doesn't answer a question about crown procedures the same way a new hire reads from a FAQ sheet. The difference is domain knowledge — understanding not just what the practice offers, but the vocabulary, the clinical context, the typical patient anxieties, and the right escalation paths when something falls outside their purview.

The Vertical Intelligence Library (VIL) is how Workforce Wave builds that knowledge into every agent from day one.

What VIL Actually Is

VIL is a curated set of domain-knowledge documents organized by industry vertical. Each vertical — dental, medical, legal, hospitality, home services, and others — has a library of structured documents covering:

Vocabulary and terminology — the words practitioners and clients use, including industry-specific shorthand that a generalist LLM might handle inconsistently
Service category knowledge — what each service type involves, common client questions, what the agent should know versus escalate
Compliance context — the regulatory environment for the vertical and how it should shape agent behavior
Escalation templates — what triggers a transfer to a human, and how to phrase the handoff
Persona and tone guidance — the conversational register appropriate to the vertical (clinical and reassuring for healthcare, warm and anticipatory for hospitality)

These aren't prompts in themselves — they're source documents that feed the prompt construction pipeline.

How VIL Becomes a System Prompt

When Workforce Wave provisions an agent, it picks the relevant VIL documents based on the vertical classification, merges them with the business-specific information extracted from the crawl, and runs the whole thing through the prompt construction pipeline.

The result isn't a template with blanks filled in. It's a purpose-built system prompt for that specific agent in that specific vertical.

Here's a simplified view of the construction logic for a dental agent:

VIL sources selected:
  - dental/vocabulary.md          → terminology layer
  - dental/services.md            → service category knowledge
  - dental/compliance-hipaa.md    → compliance guardrails
  - dental/escalation.md          → escalation triggers
  - dental/persona-clinical.md    → tone and persona

Business-specific content merged:
  - Extracted services list       → into services block
  - Hours and location data       → into scheduling block
  - Staff and credentials         → into team block
  - Unique practice context       → into persona voice

Output: 160-line system prompt, unique to this practice

The prompt isn't 160 lines because we padded it. It's 160 lines because a dental receptionist who actually knows their job has 160 lines' worth of context to draw on.

Template Variables and Partner Customization

Partners can customize VIL documents through the admin UI. The customization system uses template variables — placeholders that get resolved at provisioning time with values from the agent's configuration.

Common template variables:

{{practice_name}}         → "Sunshine Family Dental"
{{agent_name}}            → "Alex"
{{primary_phone}}         → "(843) 555-0182"
{{booking_url}}           → "https://sunshine-dental.com/book"
{{emergency_protocol}}    → "For dental emergencies outside office hours, 
                             advise the caller to call (843) 555-0199 
                             or go to the nearest urgent care"
{{after_hours_message}}   → "Our office is open Monday through Friday, 
                             8am to 5pm. I can help you schedule an 
                             appointment for the next available slot."

A VIL document excerpt using template variables:

You are {{agent_name}}, the virtual front desk for {{practice_name}}.
Your role is to help patients schedule appointments, answer questions 
about our services, and ensure they feel welcomed and cared for from 
their first point of contact.

When a patient asks about appointment availability, use the booking 
integration to check real-time availability and offer the next 
three available slots. If no slots are available within 5 business 
days, acknowledge this, offer to add them to the waitlist, and 
provide the direct scheduling link: {{booking_url}}

For after-hours calls: {{after_hours_message}}

Partners edit these documents in the admin UI, setting defaults for their entire client base. Individual client agents can then override specific variables — a large DSO might set a different emergency_protocol for each practice, for example, while inheriting all other defaults from the partner template.

Why 150-Line Prompts Beat 20-Line Ones

There's a school of thought in the AI community that shorter prompts are better — more concise, less noise for the model to process, lower token cost.

That's true for simple tasks. It is not true for complex, multi-step service interactions in regulated industries.

Here's the difference in practice. A 20-line prompt for a dental agent might read:

You are a dental receptionist named Alex. Help patients schedule 
appointments, answer questions about dental services, and be 
friendly and professional. Escalate medical emergencies to staff.

A 150-line prompt for the same agent includes:

The full CDT procedure vocabulary so the agent doesn't confuse a crown (D2710–D2799) with a bridge (D6200–D6999) when a patient asks
Specific handling instructions for new patient intake questions versus existing patient questions
Insurance verification protocol — what to collect, what to tell the patient if their plan isn't accepted, how to handle out-of-network questions without giving dental advice
HIPAA guardrails: what PHI the agent can collect (DOB for identity verification), what it cannot collect (medical history details, diagnosis information), and how to phrase the distinction
Explicit escalation triggers: patient mentions pain that has worsened overnight → escalate immediately, offer emergency number; patient describes a cracked tooth → collect info, flag as urgent, offer same-day slot if available
After-hours specific behavior: no appointment booking if the system returns no availability, offer the waitlist, provide the direct link
Language for common anxious-patient scenarios (dental anxiety acknowledgment, questions about sedation, children's first appointments)

The 20-line version will handle maybe 60% of calls adequately. The 150-line version handles 92%+ — and the 8% it escalates, it escalates correctly.

For service businesses where every call is a potential appointment or a potential lost patient, that gap is significant.

What Partners Configure vs What's Fixed

Partners control which VIL template set is the default for each vertical, and they can edit the template documents themselves. Most partners use the WFW-curated templates as a starting point and refine based on what they learn from their clients' call data over the first few months.

What's not configurable: the compliance blocks. HIPAA guardrails, TCPA time windows, Fair Housing filters — these are injected from the compliance layer, not from VIL documents, and they cannot be removed or overridden. VIL handles domain expertise; the compliance layer handles non-negotiable rules. The two systems are separate for exactly this reason.

Partners who invest in refining their VIL templates consistently see better first-call resolution rates across their client base. It's one of the highest-leverage customization points in the entire platform — an hour spent improving a VIL template improves every agent provisioned from that template going forward.

Next in this series: The Compliance Layer Your Partners Don't Have to Build — HIPAA, TCPA, Fair Housing, and PCI masking: how ComplianceRules enforces before every tool call, and why the compliance profile is read-only by design.