Reps reviewed weekly
- Baseline
- ~30% (top + loudest reps)
- Target
- 100% of reps, every week
Count of distinct rep IDs that received an agent-generated coaching digest, divided by total active reps.
← Use casesUSE CASE
§ USE CASE · SALES OPERATIONS
Sales supervisor agent
An always-on agent that reviews each sales rep's pipeline against historical patterns, flags deals at risk, and drafts coaching messages for the manager — sourced from the company's CRM and a governed semantic layer of revenue metrics. The agent never writes to the CRM without approval; it produces evidence-backed recommendations a sales leader can act on.
Executive summary
Sales leaders cannot inspect every rep's pipeline every week. The work that does happen is reactive, late, and skewed toward the loudest reps.
The Sales supervisor agent runs every weekday morning, reviews each rep's open opportunities and recent activity, and produces a short, sourced coaching note for the manager: which deals slipped, which look stuck, and what one specific intervention would change the trajectory.
The agent never edits CRM records, never sends customer-facing messages, and never makes a forecast number on its own. It produces evidence — every claim resolves to a CRM row — and the manager retains full agency on what to do next.
What this case fixes
In typical mid-market sales orgs, pipeline review is weekly, conversational, and uneven. Managers focus on the deals reps surface; risk in the rest of the pipeline accumulates silently. By the time a stalled deal is noticed, the recoverable window has often closed.
Generic AI tools either summarize CRM activity (low signal) or suggest forecast numbers (no auditability). Neither helps a manager run a coaching cadence with evidence behind it. What the org actually needs is a governed agent that knows the team's pipeline schema, applies the same risk heuristics every week, and produces decisions a manager can defend.
How it runs
This reference pattern uses three phases. Read access is scoped to the required CRM and governed metric data; selected writes, such as coaching messages or CRM notes, can require manager approval before execution.
- 01DATA
Ingest CRM and semantic layer
Pulls the rep's open opportunities, last 30 days of activity, and the headless semantic layer's win-rate / cycle-length curves by segment. Read-only.
- 02AGENT
Score and explain
The supervisor agent applies a fixed risk policy: stage age vs. segment median, missing decision-maker, no activity in N days, decreasing meeting cadence. Each flag links back to the source rows.
- 03AGENT
Draft coaching
For each flagged deal, the agent drafts one specific, evidence-backed coaching action (e.g., 'Schedule executive sponsor intro by Friday — the last one closed in 21 days'). No generic advice.
- 04HUMAN
Manager review and approval
The manager sees the digest in the Xophia console (or via Slack). They approve, edit or dismiss each recommendation. Only approved items proceed.
- 05AUTOMATION
Apply and log
Approved coaching is sent to the rep through the existing channel (Slack/Email) and a note is written into the CRM as the manager's user. Every step — read, decision, approval, write — is logged in the immutable audit trail.
Systems and connectors
| System | Role |
|---|---|
| HubSpot / Salesforce / Pipedrive | CRM (system of record) |
| Headless semantic layer | Governed metrics (win rate, cycle length) |
| Slack / Microsoft Teams | Manager channel |
| Email gateway | Outbound channel |
| Xophia audit log | Compliance and reconciliation |
KPI targets
Time-to-first-touch on stalled deals
- Baseline
- 11 days median (post-stall)
- Target
- ≤ 2 business days median
From the moment a deal trips a stall heuristic in the agent, to the moment a coaching action is approved by the manager.
Coaching evidence rate
- Baseline
- Anecdotal (no source)
- Target
- 100% of recommendations link to source rows
Audit-log check: every coaching item must reference at least one CRM record id and one semantic-layer query.
Manager review time
- Baseline
- 60-90 min weekly per rep
- Target
- ≤ 15 min weekly to triage agent digest
Self-reported via the Xophia console; cross-checked against approval timestamps.
Illustrative execution trace
A synthetic example of how an execution could be recorded. Names, identifiers, volumes, costs and outcomes are fictional; the sequence is illustrative, not a customer result.
Rep MR-014 has 12 open opportunities. The agent identified three at risk and drafted coaching for two; one was dismissed by the manager.
| Tool | Input | Output |
|---|---|---|
crm.list_open_opps | owner_id=MR-014, status=open | 12 records, fields: stage, amount, last_activity_at, decision_maker_present |
semantic.win_curve | segment=mid_market, stage=proposal, cycle_to_date=33d | median_to_close=27d, p75=44d → flag if cycle_to_date > p75 with no DM contact |
semantic.activity_baseline | segment=mid_market, stage=proposal | expected meetings/wk=1.6, observed for opp_id=O-9112=0.4 → flag |
agent.draft_coaching | opp_id=O-9112, signals=['stale_cycle','low_activity'] | Suggest scheduling executive sponsor intro by Fri; reference 3 prior won deals at 21-day cycle in same segment. |
console.publish_digest | manager_id=MGR-04, items=3 (2 drafted, 1 informational) | digest_id=D-2026-05-08-MGR-04 awaiting review |
Outcome. Manager approved 2 coaching items at 09:14 UTC. Approved items written to Slack and to CRM activity feed. Audit row D-2026-05-08-MGR-04 closed. Total agent cost: $0.043; latency p95: 6.2s.
Want to scope this for your stack?
Bring the systems, policies and approval boundaries involved. We will map a practical first scope and reply with next steps.