Concepts

The AI layer

How AI agents work in Servanex — voice, text, copilot, workflow — and why they're co-workers, not features.

Servanex ships with four AI agent classes — Voice, Text, Copilot, and Workflow — each scoped to a specific surface and permission boundary.

The four agent classes

• Voice answers your phone, books jobs, qualifies leads, handles after-hours emergencies, and follows up on estimates by phone.
• Text handles inbound SMS, email, and webchat — booking, rescheduling, status updates, payment links.
• Copilot sits inside the staff web app, ready to draft estimates, summarize customers, suggest dispatch moves, and write outbound messages on a CSR’s behalf.
• Workflow runs as a background actor inside tenant-defined automations — escalating overdue invoices, scoring sales-play candidates, drafting follow-ups.

How they work

Every agent is configured in the AI Agents module: which model it uses, which prompts and system messages it loads, which tools it’s allowed to call, which approval policies wrap its sensitive actions, and which voice and persona it presents.

When the agent runs, it composes tool calls — customer.find_by_phone, dispatch.find_slot, job.create, comm.send_sms, and so on. Each call goes through the platform’s action layer, the same code path a human CSR’s click would take. Same validation, same permissions, same audit log, same events.

Why this matters

The AI doesn’t have a separate code path or a special database connection. It’s a first-class actor in the platform. That means:

• You can’t accidentally give the AI more access than a human equivalent role would have.
• Every action the AI takes is in the audit log with model, prompt version, tokens, cost, and reasoning.
• The AI’s actions trigger the same downstream events as a human’s — workflows fire, notifications send, financial entries post.

This is the architectural difference between “AI as a feature” and “AI as a co-worker.”