This document captures the vision for replacing the traditional dashboard paradigm with an AI analyst that communicates via email—proactively briefing customers, answering questions on demand, and surfacing insights no chart could reveal. This is out of scope for the current Workerline Enhancements sprint and is documented here for future planning.
Instead of a dashboard that passively shows data, the primary interface becomes a conversation over email. A fleet manager asks: "How are crews doing on my Asia-Pacific vessels this month?" and receives a direct, contextualized answer with evidence—no clicking through filters, no interpreting charts.
"You're hiring an analyst, not buying software. They monitor every vessel 24/7, speak 200+ languages, and never take a day off."
This reframes the product from a SaaS subscription into an intelligence service—a framing that resonates with maritime companies accustomed to paying for services (P&I clubs, classification societies, crew management) rather than software access.
Pushed to you weekly (daily for large fleets). A narrative summary covering what changed, what needs attention, what's trending. Serves 80% of the "checking in" use case.
Reply to any email to ask follow-up questions. "Compare sentiment between my tanker and bulk carrier fleets." The analyst responds with data-backed answers.
Every claim includes a one-time link to a focused evidence page—no dashboard, no login. Shows the underlying data, anonymized excerpts, and trend lines.
The primary product touchpoint. A narrative email that reads like an intelligence report from a trusted advisor—not a data dump.
A longer, more reflective fleet wellbeing analysis—the difference between a daily news briefing and a monthly magazine feature. Covers bigger-picture trends, benchmarking, and strategic recommendations.
Charts rendered server-side as inline PNG images (not attachments). Compact trend lines, comparison bars, RAG status tables, and mini heatmaps—woven between written analysis. Styled HTML elements (colored status indicators, progress bars, metric tables) for lightweight data without images. Consistent visual templates the customer learns to read quickly.
Customers reply to any briefing email to dig deeper. The analyst maintains thread context and handles ambiguity naturally.
| Customer Reply | Analyst Response |
|---|---|
| "What about the Pacific fleet?" | Contextual breakdown of Pacific vessel sentiment, cases, and engagement |
| "More on that second point." | Expanded analysis of the referenced insight with supporting evidence |
| "Send this to my DPA at john@osm.com" | Forwards a clean, shareable summary to the specified recipient |
| "Send me a Q4 fleet report" | Generates a narrated PDF with executive summary, findings, trends, and recommendations |
| "Send audit-ready summary for MV Nordic Spirit" | Compliance packet: engagement rates, case history, escalation response times, wellbeing trends |
Every substantive claim in an email includes a small link that opens a single-purpose evidence page. Not a dashboard—a focused, read-only page that exists to answer one question: "Where did this insight come from?"
The analyst can be invited to Teams meetings (crew welfare reviews, safety committees) and participate via chat—surfacing data, answering questions in real time, and providing meeting summaries afterward.
The first week isn't a product walkthrough—it's the analyst introducing itself:
"I've been assigned to your fleet. Here's what I've observed so far. Here are the kinds of things I can help you with. What matters most to you?"
This first exchange calibrates ongoing behavior—one customer cares about retention risk, another about compliance readiness, another about mental health specifically. The analyst remembers and adapts.
| Challenge | Risk | Mitigation |
|---|---|---|
| Discoverability | Email becomes a confirmation machine—only tells people what they expect | Editorially opinionated briefings with a "something you might not have noticed" section. Suggested questions at the end of each email. Monthly deep-read for bigger patterns. |
| Depth on demand | Email format strains for deep, open-ended exploration | Generated PDF reports on request. Temporary evidence pages with richer data. Depth feels like escalation within the same conversation, not a different tool. |
| Proof & credibility | "An AI email told me" won't cut it in a boardroom | Inline evidence attribution with specific numbers. One-click evidence cards behind every claim. Forward-friendly design so emails work up the chain. |
| Trust | Maritime industry is conservative and compliance-heavy | Every claim traceable to actual conversations/data points. Evidence layer exists as a verification backstop. Measured, understated tone builds credibility. |
Three components, intentionally simple:
| Component | Purpose | Notes |
|---|---|---|
| Email composition engine | Generates data-rich HTML emails with inline chart images and styled metrics | Scheduled (weekly/daily) + on-demand. Reusable chart templates (trend line, bar chart, RAG table, mini heatmap). |
| Conversational reply handler | Interprets customer replies, maintains thread context, generates follow-up emails | Handles ambiguity, asks clarifying questions. Generates PDFs for report requests. |
| Evidence card renderer | Serves single-purpose web pages behind expiring tokens | No login, no sessions, no frontend framework. Server-rendered. Content pre-generated alongside emails. |
The current Workerline Enhancements sprint (§4.1–4.7 of the main spec) builds the data layer that makes the AI Analyst viable. The table below references both the existing production schema (sessions with summaries, sentiment scores, cases with messages/actions, departments, survey_instructions) and the new tables being added in the current sprint. The main spec's ERD is simplified to show only new/changed fields — the existing production tables are more comprehensive.
| Capability | Status | Relevance to AI Analyst |
|---|---|---|
| Session data with AI summaries | ✅ Exists (production) | Each conversation already produces a GPT-4 summary, sentiment breakdown (positive_pct, negative_pct, neutral_pct), and risk analysis fields. These exist in the production sessions table (not shown in the main spec's simplified ERD). This is the raw material for briefings. |
| Risk scoring (0.0–1.0) | ✅ Exists | Severity level, category, warning signs, follow-up timeline — all stored per session. Directly feeds "vessels of concern" and anomaly detection. |
| Cases with messages & actions | ✅ Exists (production) | Full case lifecycle (open → assigned → resolved) with cases, case_messages, and case_actions tables in production schema. Provides the "case updates" section of briefings. |
| Org/site/department hierarchy | ✅ Exists (production) | Multi-vessel fleet structure: orgs → sites → departments already modeled in production schema. Enables fleet-level and vessel-level queries. |
| Resend email integration | ✅ Exists | Already used for escalation emails and notifications. Outbound email infrastructure is ready. |
| Bull queues (Redis-backed) | ✅ Exists | Async job processing for risk analysis, email, push. New briefing generation jobs slot in naturally. |
| OpenAI GPT-4 integration | ✅ Exists | Chat completions, structured outputs, streaming — all working. The analyst's synthesis and reply generation use the same infrastructure. |
| Escalation system | 🔧 Current sprint | Auto-escalation with configurable thresholds. Feeds "proactive alerts" in the analyst's briefings. |
| Org settings & feature flags | 🔧 Current sprint | Per-org configuration. Briefing preferences (frequency, focus areas) would live here. |
| Inbound email processing | ❌ New | Resend supports inbound webhooks but it's not wired up. Needed for conversational replies. |
| Server-side chart generation | ❌ New | Needed for inline email visuals. QuickChart API or chart.js + canvas → PNG. |
| Teams bot framework | ❌ New | Microsoft Bot Framework registration, webhooks, adaptive cards. Significant new surface area. |
Broken into four phases that can be delivered incrementally. Each phase is independently useful—the client doesn't need to commit to all four at once.
| Phase | Scope | Builds On | Effort | Estimate |
|---|---|---|---|---|
| Phase 1: Email Briefings | Scheduled Bull job for weekly/daily briefing generation. GPT-4 synthesis across sessions, cases, and risk data. HTML email templates with inline chart PNGs (QuickChart or canvas-rendered). Fleet overview, vessels of concern, anomaly narratives. | Existing sessions, risk analysis, cases, Resend, Bull queues | 1 week | $4,400 |
| Phase 2: Conversational Reply | Resend inbound webhook handler (SPF/DKIM validation). Thread context management with deduplication. Natural language → data query → GPT-4 synthesis. On-demand PDF report generation. | Phase 1 email engine + existing GPT-4 integration | 1–1.5 weeks | $5,500 |
| Phase 3: Evidence Cards | Single Express route with signed token auth. Server-rendered evidence page — trend line, anonymized excerpts, fleet comparison. Pre-generated alongside emails. Expiring links, access logging. | Phase 1 content generation | 2–3 days | $1,800 |
| Phase 4: Meeting Integration | Microsoft Teams bot registration + Azure AD auth. Real-time chat participation via Bot Framework. Meeting context tracking. Auto-generated post-meeting summaries and action items. Tenant governance and app policy setup. | Phase 2 query engine + existing data layer | 2–3 weeks | $9,900 |
5–7 weeks · agentic coding workflow · incremental delivery · each phase independently deployable
Briefings + replies + evidence cards · ~2.5 weeks
Teams meeting integration · 2–3 weeks