Skip to main content

What does success look like — the agent is the interface, forms are the fallback?

OUTCOME MAP — MULTIMODAL AGENT INTERFACE
════════════════════════════════════════════════════════════

  DESIRED OUTCOME
  50%+ of drmg-sales tasks completed via conversation
  within 60 days of launch
       │
       ├── Contributing Factors
       │     • 8 WorkCharts already built and tested
       │     • Skill router handles text→WorkChart matching
       │     • Vercel AI SDK (useChat) already in package.json
       │     • Modalities knowledge documented (7x7 matrix)
       │     • A2A protocol enables multi-agent orchestration
       │
       ├── Obstacles
       │     • No conversational UI exists (all form-based)
       │     • WorkCharts accept string-only inputs (no multimodal)
       │     • No session memory across conversation turns
       │     • Agent Platform (identity + memory) not yet at L3
       │     • Identity & Access (auth) required for any user-facing work
       │     • AI feature pricing not validated
       │
       ├── Investigations
       │     ? Which modality (text, file, voice) proves value fastest?
       │     ? Does the existing skill router generalise to conversational input?
       │     ? What session memory architecture supports multi-turn context?
       │     ? How do competitors (HubSpot AI, Einstein) handle modality routing?
       │
       ├── Success Measures
       │     ✓ Agent processes RFP from PDF drop to proposal draft < 5 min
       │     ✓ Pipeline summary via "What's my pipeline?" in < 3 seconds
       │     ✓ 50%+ completed tasks originate from conversation (60 days)
       │     ✗ KILL: < 10% conversational task rate after 30 days
       │
       ├── Roles
       │     A: Product (defines conversational UX)
       │     R: Engineering (builds chat UI, modality router, normaliser)
       │     C: Sales team (validates agent usefulness)
       │     I: Agent Platform team (provides identity + memory)
       │
       └── Next Actions
             1. Wire useChat to a streaming endpoint (T0, day 1)
             2. Build text-only chat widget with feature flag (T0, day 3)
             3. Validate skill router handles conversational intent (T0, day 5)

════════════════════════════════════════════════════════════

Headline​

Forms serve databases, not users.

P-Q-D​

ProblemQuestionDecision
8 WorkCharts with zero conversational entryWhat if the agent were the interface?Build chat-first, keep forms as fallback

Gate​

  • Desired outcome is crisp (50%+ tasks via conversation within 60 days)
  • Contributing factors have evidence (8 WorkCharts built, useChat in package.json)
  • Obstacles have mitigations (Agent Platform dependency = feature flag for incremental rollout)
  • Investigation questions have owners (product owns modality prioritisation, engineering owns skill router validation)
  • Success measures are objective and binary (< 5 min RFP, < 3s pipeline, 50%+ rate)
  • RACI is complete
  • Next actions are assigned

Context​