Sales outreach plus AI tempts teams to automate sends. In Microsoft Outlook, that path leads quickly to deliverability damage, compliance complaints, and reps who no longer trust the draft. Guardrails first protect brand, inboxes, and auditability—then you measure whether AI saves editing time, not how many messages left the building without a human.
This playbook describes a draft-and-review pattern: AI writes into Outlook drafts; humans send. CRM holds segments and outcomes; Outlook is the send surface—not an autonomous mailbot. Connect logging to audit trails, caps to data boundaries thinking, and policy language to governance roles so Rev Ops and Legal share ownership.
Outlook workflow (v1) — draft, never auto-send
Design v1 assuming every external message has a rep’s name on it and a send click in Outlook. Automation that drops messages into Outbox or schedules bulk send without review is out of scope until eval, DLP, and legal sign-off exist for a narrower template set.
Segment in CRM first. Reps select an approved list; suppression and opt-out flags apply before any connector runs. Export includes segment ID and policy class (marketing vs transactional) so DLP rules match message type.
Draft folder only. Copilot or a connected workflow creates drafts in the rep’s Drafts folder or a shared AI-Drafts mail folder—never Outbox, never scheduled send rules to external domains in v1.
Snippet injection from approved library. Personalization pulls only from versioned fact blocks: customer name, public news cite ID, product tier from CRM. Prompts must reference snippet_id per factual claim—no invented awards, no fabricated “saw your post about X” without citation.
Rep review and manual send. Edit subject and body; confirm opt-out footer; send from Outlook with rep credentials. CRM sync records draft ID, send timestamp, override percent if body changed materially.
Numbered path for reps: Open CRM task → generate drafts in Outlook → review each for subject, footer, citations → send or discard → CRM logs outcome. Training should show the path once; guardrails live in connector config, not memory.
Guardrails table (controls and owners)
| Control | Outlook / ops setting | Purpose |
|---|---|---|
| Human send only | No auto-forward rules to external domains | Accountability |
| Daily send cap | 40 drafts generated / 25 sends per rep | Deliverability |
| Approved snippet library | SharePoint list Outreach-Snippets-v3 |
Consistent claims |
| Opt-out / suppression | CRM export filter + DLP keyword scan | Compliance |
| No fabricated references | Prompt requires citation ID per fact | Trust |
| BCC policy | Block BCC on external bulk; use CRM logging | Privacy |
Rev Ops owns caps and snippet versions; IT implements connector and DLP; Legal owns prohibited claims list. RACI should mirror governance roles even if workflow ID is outreach-outlook-v1, not support.
Rate caps and escalation
Caps prevent “AI productivity” from becoming ISP throttling. Adjust tiers by role; log when caps hit three days running—that signals bad drafts or bad lists, not lazy reps.
| Tier | Daily draft cap | Daily send cap | Escalation |
|---|---|---|---|
| SDR | 40 | 25 | Manager if cap hit 3 days running |
| AE | 25 | 15 | Rev ops review |
| Shared mailbox | 100 total | N/A — no direct external send | IT alert |
If DLP flags a draft (export-control keyword, missing opt-out), route to Legal queue folder; block send until cleared. SLA 24 hours for flagged drafts; decision logged in CRM case note with reviewer ID.
DLP and compliance triggers
Configure Microsoft Purview (or equivalent) to hold drafts when marketing class mail lacks unsubscribe block, when body matches prohibited superlatives from legal list, or when attachment types are not allow-listed. Holds are not suggestions—connector must prevent send until status cleared.
Pair DLP with eval cases: missing footer, prohibited phrase, missing citation ID. See evaluation hooks for promotion gates before expanding template coverage.
Logging compatible with audit trails
Each send produces a row aligned with audit trails: workflow_id outreach-outlook-v1, draft and sent timestamps, rep_id, snippet_ids[], override_percent, suppression_check pass/fail. Without override percent, you cannot see whether AI helped or reps rewrote entirely.
Operating outreach AI in practice
Measure reply quality and unsubscribe rate, not drafts generated. High generation with low send rate means poor drafts; high send with high unsubscribe means weak suppression or off-brand snippets.
Subject lines: ≤60 characters; ban ALL CAPS; A/B only through approved template pairs in library version bumps.
Shared mailbox: Inbound replies only in v1—outbound bulk from shared addresses confuses threading and SPF alignment.
Monday start: Disable auto-send rules on outreach connectors. Publish snippet library with version ID. Set per-rep caps. Run five test drafts through DLP and Legal before pilot.
Outreach AI in Outlook works when it accelerates editing, not sending.
