Skip to content

AI Communication Architecture (Email Layer)

Updated
2026-05-29

Summary

The enterprise has multiple functional email addresses across multiple domains ([email protected], [email protected], [email protected], [email protected], and others). Brandon's day historically involved manual triage of this volume. Andrea and Karin do NOT do office operations (they are cleaning staff, per 2026-05-27 correction). The substrate must handle the email layer autonomously where it can, with Brandon's surface limited to curated digest and structured todos for items genuinely requiring his judgment.

The 4-stage rollout

Stage 1: Read-only intake and Brandon-curated digest (1-2 weeks build)

  • SE-Application wires OAuth read access to brandon@ and ap@ inboxes
  • AI Communication Monitoring (substrate-native role) reads incoming mail every 15 minutes
  • Each email classified into rough buckets (spam, customer, vendor, banking, internal, personal)
  • Daily digest at 6pm Phoenix time delivered to Brandon: "Today's email — N received. X customer (Y need response), Z vendor (W invoices for AP), V banking, U personal, T archived as spam"
  • Brandon reads the digest, sees curated view; responses still come from his real inbox

Outcome: visibility without changing email handling. Noise filtered, daily picture clear.

Stage 2: AP pipeline and todo generation (2-3 weeks after Stage 1)

For ap@: - Invoice content classified and parsed (PDF extraction, vendor, amount, due date, terms) - Structured AP queue rendered in dashboard - QB Assistant identifies routine (recurring vendors, expected amounts) vs. needs-attention (new vendor, unusual amount, terms questions) - Todo list: "AP processing — 5 invoices ($X total), 3 routine + 2 need review"

For brandon@: - Items requiring Brandon's response get structured todos with AI-drafted suggestion - Items handleable by substrate agents route directly to them - Strategic items flagged for next chat-MC1 session

Stage 3: Autonomous AI responses (4-6 weeks after Stage 2)

Specific categories where confidence is high get autonomous response: - Standard customer inquiries (hours, location, basic pricing, dealer info) - Vendor acknowledgments (invoice received, payment scheduled) - Internal acknowledgments - Spam confirmations

Bedrock harness enforces: never claim availability without verifying inventory; never commit pricing outside policy thresholds; never apologize on company's behalf (legal implications); never agree to terms or obligations; never share financial info.

Start narrow — 5% autonomous response rate in month one, growing to 30-40% by month three, eventually 60-70% by month six if pattern proves itself.

Stage 4: Full substrate email handling (3-6 months out)

Substrate agents own their own email surfaces: - AMM handles sales@ inquiries autonomously - SMM handles marketing-side outreach - QB Assistant handles AP processing end-to-end - AI Communication Monitoring becomes routing layer, not operating layer - Brandon's surface: daily digest + specific escalations only

Critical design points

Inbox segregation strategy

Two architectural options:

Option A — Multiple physical inboxes (each functional address is its own Google Workspace mailbox at ~$6/month/user) Option B — Single physical inbox with forwarding/aliases (all functional addresses forward to one mailbox; substrate classifies by "delivered to" header; $6/month total)

Recommendation: Start with B (single intake), migrate to A if specific inboxes need genuine human separation later.

Brandon's surface

Brandon's inbox becomes: - Daily digest at 6pm - Structured todo list with AI-drafted context - Immediate escalations ONLY for genuinely urgent (cash threshold events, regulatory, legal)

Not raw email volume. Curated surface.

No human triage relay

Important — Andrea and Karin are cleaning staff

Earlier framework assumptions had Andrea handling informal triage. The 2026-05-27 correction clarified Andrea and Karin clean the building and are not in office operations. Email triage must be FULLY substrate-driven, not Andrea-assisted.

Customer-facing autonomous response risk

This is the highest-stakes question. Wrong information to a customer = real damage. Specific protections: - Bedrock truthfulness: AI cannot claim availability, pricing, or specs without verifying live inventory - Bedrock external-party: customer communication is by definition external; autonomous response must be tightly scoped - Reviewer at milestone: AI response patterns reviewed weekly initially, monthly later, to catch error patterns before they compound

Conservative start. Expand autonomous category as patterns prove out.

Substrate-native sequencing

AI Communication Monitoring should be migrated to substrate-native operation (manager-as-substrate-agent pattern, established by AM-agent 2026-05-26). Reading email and classifying it benefits from continuous availability rather than session-spinup.

In the ratified migration order, AI Comm Monitoring was at position #7 (after OM, IT, WDM, SE-App, SMM, AMM). The email workload may justify moving it up if email handling becomes priority.

Open questions

Information needed

  • Email infrastructure choice: Google Workspace vs. Microsoft 365 — which is current/preferred?
  • Functional address inventory: Full list of email addresses across all domains
  • Stage 1 build approval: Cost is small ($6-12/month + SE-Application engineering time). Brandon to approve scope and timing.
  • Customer-trust threshold: At what point does autonomous response feel safe enough to expand? Brandon's risk tolerance defines this.

Resolution requires: Brandon Tomkins.


Living document. Last updated 2026-05-27 by chat-MC1.