Supernal
Statement of Work for Heritage AI
← All proposals

Help Desk AI Employee

Sam

An in-app chat widget that handles inbound support across the Heritage AI platform: answers producer and firm-admin questions in-app, retrieves documents, opens tickets, and hands off cleanly to humans whenever judgment, advice, or platform engineering is required.

01 / Overview

The brief

The Help Desk AI Employee runs on doc-grounded retrieval with citations on every answer, archives every conversation for recordkeeping, and operates inside Heritage AI's compliance posture. The Heritage AI team approves voice and escalation rules. The Help Desk handles the rest.

02 / Objectives

What success looks like

  • Resolve over 50% of partner-firm (L2 to L5) support inbound without human involvement, cutting producer and admin time-to-answer from hours to seconds.
  • Let HNW end clients (L1) ask the AI for documents, book status, and scheduling.
  • Surface auditable citations on every response. No hallucinated facts reach users.
  • Take routine support questions off Heritage's employees and partner-firm directors.
  • Build one data layer and governance setup that serves both audiences.

03 / Scope

Scope of work

Stage 1 · Weeks 0 to 4

Knowledge base, personas, and escalation design

Lock the foundations Sam will run on. The knowledge base is treated as a product, not an artifact. The personas and escalation rules are approved before any conversation goes live.

  • Curated knowledge base: module-level help docs, glossary, process docs.
  • Versioned RAG retrieval with visible citations on every answer.
  • Conflict detection, freshness SLAs, deprecation workflow.
  • Quote-grounded answering for high-stakes content (rates, deadlines, policy).
  • Single editorial owner per topic.
  • Partner-firm help desk (L2 to L5) answers platform questions directly and takes actions where permitted.
  • HNW end-client chat (L1) answers transactional questions only (documents, status, scheduling) and routes anything else to the client's advisor.
  • Per-persona writing rules: how the AI talks, what it avoids, and how its tone shifts by situation (warmer for complaints, neutral for compliance questions).
  • Recommended visible AI disclosure: "Hello, I'm Heritage's AI assistant…"
  • Banned-language list, formatting spec, signature.
  • 5 to 20 example Q&A pairs per persona, by Heritage, as the AI's reference for what good looks like.
  • Four escalation triggers: user asks for a human, AI is unsure, user is frustrated, topic requires a human.
  • Handoff to the human includes user identity, conversation summary, what the AI already tried, sentiment indicator, and a suggested next step.
  • Real-time human override and takeover at any moment.
  • CS Manager workspace: a ticket inbox showing every open and recent ticket with status, assignee, priority, age, channel, and last AI response. Filters and saved views by team, persona, or topic. SLA timers per ticket. Bulk actions (close, reassign, tag). Ticket merge and split.

Stage 2 · Weeks 4 to 6

Data layer, integrations, and compliance

Single database serves both the partner-firm and end-client surfaces. Compliance posture (PII redaction, audit trail) is in place before any user sees a Sam response.

  • Single database used by both the partner-firm and end-client AI: conversation transcripts, citations, action logs, sentiment scores, escalation history.
  • Heritage AI platform integration: read-only access to user and firm context, book status, document inventory.
  • Email notifications routed by topic to the right human owner.
  • Full audit trail: every conversation, draft, and action archived.
  • PII redaction in conversation logs before storage.
  • Conversation archive available for Heritage's compliance review.
  • Monthly performance report: usage volume, containment rate, escalation reasons, CSAT.

Stage 3 · Weeks 6 to 12

Shadow mode, co-pilot mode, and autonomous launch

Three rollout modes that progressively shift load from humans to Sam. Heritage holds the brake until the data shows Sam is ready for each step.

  • Shadow mode (Weeks 6 to 8): AI drafts, humans send all responses. Reference Q&A bank built from real traffic.
  • Co-pilot mode (Weeks 8 to 10): AI suggests, humans approve. Prompt and knowledge base iteration on real data.
  • Autonomous with audit (Weeks 10 to 12): AI sends on transactional topics, humans QA samples. Weekly QA and knowledge base updates.

04 / Out of scope

Out of scope

ItemStatus
Email and voice channelsNot in scope (potential future add-on)
Multi-language supportNot in scope (English only at launch)
Personalized investment, tax, estate, or legal adviceAlways handed off to a human director

05 / Timeline

Twelve weeks to autonomous

PhaseDurationOutcome
Knowledge base curation and governanceWeeks 0 to 2Knowledge base consolidated, conflict-detected, versioned. Freshness owners assigned.
Persona, voice, and escalation designWeeks 2 to 4Voice and tone rules locked. Escalation matrix approved.
Data layer and integrations buildWeeks 4 to 6Database schema, platform read-only access, PM/CRM ticketing, audit pipeline live.
Agent build and shadow modeWeeks 6 to 8AI runs in shadow. Humans review and send all responses. Reference Q&A bank built.
Co-pilot mode and tuningWeeks 8 to 10Humans approve. Prompt and knowledge base iteration on real data.
Production launch and hypercareWeeks 10 to 12Autonomous on transactional. Weekly QA and knowledge base updates.

06 / Required inputs

What Heritage provides

For the Help Desk AI Employee to start handling inbound on day one, Heritage AI provides the following in week 0 to 2. Approval delays extend cadence day-for-day. Several items may already exist within Supernal's stack from prior Heritage work. We will inventory what is reusable in week 0 and only ask Heritage AI to provide the gaps.

6.1   Engagement and approvals

  • List of approvers with helpdesk authority.
  • Weekly 30-minute review call during build.

6.2   Knowledge base and content

  • Module-level help docs covering every feature live on the Heritage AI platform.
  • Glossary of estate-planning and Heritage-specific terms (approach book, asset sheets).
  • Process docs (book production, QA workflow, PDF delivery, end-client vault access).
  • 5 to 20 example support Q&A pairs per persona, hand-written by Heritage.
  • Banned-language list, tone preferences, and visible disclosure copy.

6.3   Escalation and routing

  • Routing matrix: who handles what.
  • Topic auto-escalation list.
  • Heritage AI employees' availability windows for live takeover.
  • After-hours escalation rules.

6.4   Operational and security access

  • Heritage AI platform read-only API access (user and firm context, book status, document inventory).

Investment

Pricing

$7,000 / month