Pipeline Hygiene Agent: deployment guide

Pack Version: 1.0.0 · Time: 0.5–1 day · Owner role: RevOps (config), Sales leadership (weekly digest consumer)

Follow in order. Each step ends with a check; don't proceed on a failed check.

Step 0: Prerequisite

Open pipeline of meaningful size (50+ deals); recurring forecast-accuracy or pipeline-review pain; HubSpot as CRM. If that's not true yet, fix it before starting this build.

Step 1: Write the config from real data (30–90 min)

Pull real examples from your CRM (closed-won deals, real accounts, or real historical decisions depending on the agent) and derive the workflow's rule thresholds and weights from them, not from a generic default. Save as a versioned config file and commit it.

Check: every rule/threshold in the config traces to something real, not a guess.

Step 2: Create CRM properties (20–30 min)

Follow `hubspot/setup-guide.md` to create the agent-owned properties: `pipeline_hygiene_score`, `pipeline_hygiene_flags`, `pipeline_hygiene_recommendations`, `pipeline_hygiene_scan_date`.

Check: all internal property names resolve in HubSpot; no naming collisions with existing sales-owned fields.

Step 3: Build supporting integrations (30–60 min)

Follow `clay/clay-setup-guide.md` for any enrichment/sourcing tables this agent needs, and confirm required tool connections listed in `agent-spec.md` under Required Tools.

Check: a manual test row flows through the integration end to end.

Step 4: Import and configure the n8n workflow (1 hour)

1. Import `n8n/workflow.json`.
2. Configure every node marked CONFIGURE: credentials, schedule, versioned config.
3. Leave the schedule/trigger disabled.

Check: manual run on `demo-data/sample-input.json` reproduces `demo-data/sample-output.json`.

Step 5: Pilot before go-live

Run against a small, real sample with CRM writes disconnected. Review results with the primary user listed in `agent-spec.md`. Tune the config from their corrections, then re-run.

Check: the pilot reviewer agrees with the agent's output on the large majority of sampled records.

Step 6: Go live

1. Reconnect write nodes, enable the schedule/trigger.
2. Complete `qa/qa-checklist.md` with a non-builder reviewer; file the signed checklist.

Check: first live run writes only agent-owned fields — diff a sample record before/after to confirm.

Rollback

Disable the schedule/trigger. Agent-owned fields are inert and can be bulk-cleared without affecting any sales-owned field.

What not to do

Don't wire this agent's output fields into automated downstream actions (sequence enrollment, auto-routing, auto-closing) without an explicit governance decision — see `agent-spec.md`'s Governance section for this agent's specific hard limits.