Context
An enterprise operations team ran a shared mailbox that had become a chokepoint. Messages arrived at a rate the team could not keep up with; manual triage took up roughly half of two analysts' time; and the reply tone and structure varied materially depending on who was on duty. Upstream teams complained about slow routing and inconsistent answers. The operations manager was looking for a way to absorb the volume without adding headcount and to standardize the outgoing communication so that upstream complaints stopped.
The obvious solutions had flaws. A generic helpdesk tool would have imposed a workflow the upstream teams would not adopt. A point-solution chatbot in front of the inbox would not have handled the attachment-heavy cases. What the ops manager actually wanted was the existing inbox and the existing team, augmented by an agent that could read every message, classify intent, route to the right owner, and draft a reply ready for one-click send.
Challenge
The requirements that drove architecture:
- Handle attachments and embedded forms. A material fraction of the mailbox traffic was not plain text. Invoices, vendor onboarding forms, expense requests — all had to be parsed, not just the email body.
- Work in the existing inbox. The solution had to operate in Gmail and Outlook on the existing shared mailbox. No new app, no separate portal.
- Consistent tone, controllable by the team. Draft replies had to match the team's standing tone, and the team had to be able to adjust the tone policy when leadership asked.
- Zero send-without-review by default. One-click send was the target; autonomous send required an explicit enable-per-intent-class decision.
Approach
Thoughtwave deployed TWSS AI Email Assistant — our production multi-agent shared-inbox automation — connected to the client's M365 tenant and to the Slack workspace where the ops team coordinated. The platform runs four stages per incoming message:
- Inbox connector. Polls or streams new messages from Gmail/Outlook via Microsoft Graph.
- Multi-agent triage. Specialist agents for intent classification, urgency detection, attachment parsing, and owner routing run in parallel; a coordinator agent reconciles.
- Routing or drafting. Messages with clear owners route to the appropriate queue; messages that fit a reply template get a draft reply in the original thread.
- Human review and send. The team sees each draft in the inbox and sends with one click; an optional Slack notification surfaces urgent items.
Engagement arc:
- Weeks 1-2. Mapped the intent taxonomy from six months of historical mailbox traffic. Identified five repeating intents that covered 80% of the message volume.
- Weeks 3-4. Stood up the platform on the client's infrastructure, connected the inbox via Graph API, and configured the multi-agent topology.
- Weeks 5-6. Ran the platform in shadow mode — drafting replies but not sending — while the team reviewed for quality. Tuned the tone-policy prompt and the routing rules.
- Weeks 7-8. Cut over to production with one-click-send drafts for all five primary intents, and enabled autonomous send for the two lowest-risk intents after an additional review.
What we built
The production system has five components:
- Inbox connector service. Microsoft Graph client with delta-query subscriptions for low-latency message arrival.
- Multi-agent orchestrator. A Python agent framework running intent, urgency, attachment, and routing agents in parallel with a coordinator on top.
- Vector store for similar-message retrieval. Previous resolutions surface automatically so draft replies can reference what was said before in comparable cases.
- Draft writer. LLM with a tone-policy system prompt; output goes back into the thread as a draft, ready for review.
- Audit and analytics. Every decision captured; daily analytics show volume by intent, response-time deltas, and confidence distribution.
Outcomes
- 70% inbox-triage automation on the pilot mailbox. Messages are classified, routed, and drafted without a human touching them until review.
- Materially faster time-to-first-response across the full message mix. Even the messages that require human judgment are faster because the research step is pre-done.
- Consistent reply tone and quality. The tone policy is a single prompt, updated in one place, and applied across every draft. No more drift based on who is on duty.
- Five supported channels at go-live. Gmail, Outlook, Microsoft 365, Slack, and Microsoft Teams — the ops team works in the surface that makes sense for the message type.
What's next
The next phase extends the platform to the client's ap@ mailbox, where attachments are heavily weighted toward invoices; routing into the TWSS AI Invoice Automation pipeline is the natural follow-on. The sales@ mailbox is the next candidate after that, where the reply-draft value is higher because the tone and CTA structure are business-critical.
Why the platform approach pays off
A single-inbox automation project that ends at 70% triage on one mailbox would already be a strong outcome. The reason this matters more is platform reuse: the second mailbox ships in weeks, the third in days. The client now has shared infrastructure — connector, agent framework, vector store, audit log — that powers every inbox automation that follows, at a fraction of the incremental cost.
For broader context, see the TWSS accelerators portfolio and our AI & Generative AI service.