AI GTM & Sales Agent Pipeline
An AI-assisted sales operations system that connects inbound replies, lead context, AI-drafted responses, Slack approvals, follow-up actions, payment events, and operational handoff.
- What I built
An AI-assisted GTM workflow that connects inbound replies, lead context, AI-drafted responses, Slack approvals, follow-up actions, payment events, and operational handoff.
- Business problem
Inbound replies were arriving across channels with no shared context, sales reps spent time drafting repeatable answers, and approvals, follow-ups, and revenue events were tracked in different places.
- System pattern
Inbound reply → Normalize → Load lead + deal context → AI draft → Slack approval → Send / edit / reject → Follow-up routing → Payment event → Operational handoff.
- Reliability layer
No customer-facing message is sent without human approval. Drafts include context, edits feed back into the system, follow-ups run on deterministic cadences, and monitors catch stuck builds, errors, and reply reconciliation drift.
- Main takeaway
The AI drafts and routes; humans approve sensitive replies and the workflow owns state, follow-up, and revenue handoff.
- 01Inbound reply
- 02Normalization
- 03Context loading
- 04AI draft
- 05Slack approval
- 06Send / follow-up / handoff
- 07Logging & monitoring
Drafting is AI-assisted, but sending is human-approved. Every important event — approval, send, follow-up, payment — updates operational state and is observable.
What I designed and built.
I designed and built the sales automation architecture, mapped the inbound reply and follow-up process, implemented the n8n workflows, connected Instantly, Slack, AI models, webhooks, and operational records, and designed the human approval layer, AI drafting flow, feedback loop, routing logic, payment handoff, and reliability monitoring patterns.
Sales replies were scattered across tools, hard to triage, and difficult to follow up consistently.
Outbound and GTM workflows often create a new operational problem after they start working: replies arrive from different campaigns, leads have different context, follow-up timing matters, and sales actions can affect real customer relationships. A simple automation can send messages, but it cannot safely decide what should happen next without context, routing, and human oversight.
The business needed a system that could capture inbound replies, match them to the right deal, prepare an AI-assisted response, route sensitive cases to the right person, keep records updated, and preserve human approval before important sales actions were sent.
A human-in-the-loop AI sales system from reply intake to approved action.
The system listens for inbound replies, normalizes the data, matches each reply to the right lead or deal, checks existing context, classifies the situation, drafts a response using conversation history and playbook context, sends the draft to Slack for approval, and routes the approved action into the correct downstream workflow.
Inbound intake and lead context
- Captures replies from Instantly and webhook events
- Normalizes inbound message data
- Matches replies to existing deals
- Detects VIP or media-related replies
- Routes new or existing opportunities
AI sales conversation engine
- Loads deal context and conversation history
- Pulls playbook and prior learnings
- Builds a structured agent prompt
- Drafts a sales reply
- Parses the draft into operational fields
Human approval and action routing
- Sends proposed replies to Slack
- Supports approve, edit, reject, and design approval actions
- Stores approval decisions and timestamps
- Routes follow-up, CX, and intake actions
- Keeps operational state updated after each decision
AI-assisted GTM pipeline with human approval.
Each workflow has a focused responsibility: capture the reply, match the lead, build context, draft the response, ask for approval, route the selected action, update the system of record, and notify the right channel. AI is used to draft and classify, but human approval controls customer-facing actions.
Screenshot walkthrough.
Sanitized views of the actual production workflows. Click any screenshot to view it full-size.
Inbound reply intake
The workflow captures replies from Instantly, responds quickly to the webhook, extracts and normalizes data, stops cadence when needed, checks existing deals, detects VIP/media cases, and routes the reply into the right downstream path.
Reply normalization and deal matching
Before any sales action is taken, the system standardizes the inbound payload, identifies the sender, checks for an existing deal, and prepares the context needed for routing.
Intent and priority routing
Replies are routed based on existing deal state, VIP/media flags, and whether the message should continue to sales, alert an operator, or trigger another workflow.
AI conversation engine
The sales reply engine loads deal context, conversation history, playbook data, prior learnings, and then builds an AI prompt to generate a draft response.
Context loading for AI drafting
The AI draft is grounded in conversation history, playbook instructions, and stored learnings so the generated response reflects the current sales situation rather than a generic prompt.
Slack approval card
Generated drafts are converted into Slack approval cards so a human can approve, edit, reject, or route the next action before anything customer-facing is sent.
Approval action handler
Slack button actions are handled by a dedicated workflow that parses the event, acknowledges Slack, routes the approval type, and updates the corresponding deal or approval record.
Approve, edit, reject, and design paths
The workflow separates approval paths for sending a message, requesting edits, rejecting a draft, or approving a design-related action, keeping each business decision explicit.
Edit feedback loop
When a reviewer requests changes, the system records the revision request and feeds it back into the drafting process so the next response can incorporate human feedback.
Payment event handoff
When payment events happen, the system updates deal status, records payment activity, notifies Slack, and triggers downstream delivery or fulfillment workflows.
The same GTM system pattern, adapted across sales operations.
The sales agent pipeline was not a single automation. It was a connected set of workflows for inbound reply handling, AI drafting, Slack approvals, follow-up actions, payment events, domain callbacks, weekly analytics, reconciliation, and operational recovery.
AI redraft loop
When a reviewer requested changes, the system loaded the original deal, conversation context, playbook, prior learnings, and feedback, then generated a revised draft for another review cycle.
Follow-up cadence
A scheduled workflow found follow-up candidates, built AI-assisted follow-up drafts, posted them to Slack for approval, and tracked reminders for pending actions.
Follow-up action handler
The action handler processed approvals, edits, skips, and follow-up actions, then updated approval state, sent messages, advanced cadence counters, or posted revision requests.
Domain purchase callback
When a domain purchase event was received, the system parsed the callback, routed by status, marked the deal live, logged the purchase event, and notified Slack of live, unavailable, premium, or error states.
Weekly analytics digest
A scheduled digest loaded active deals, fetched weekly stats, formatted the report, checked traffic conditions, sent an email digest, and logged the event.
Operational workflows around the sales system.
The production system included support workflows that monitored stuck builds, reconciled reply data, caught errors, and notified Slack when human intervention was needed. These workflows helped keep the sales system reliable after launch.
Error catcher
A centralized error workflow receives workflow failures, formats the error report, and posts the alert to Slack so failures are visible instead of silent.
Stuck build monitor
A scheduled monitor finds stuck builds, decides whether each one should retry, triggers auto-retry when safe, and sends Slack alerts when manual intervention is needed.
Daily reply reconciliation
A scheduled reconciliation workflow checks recent Instantly replies, compares them against existing deal records, computes missing or dropped events, and posts a report to Slack.
The building blocks.
Inbound and lead state
- Instantly reply webhooks
- Webhook response handling
- Inbound message normalization
- Deal lookup by sender
- Existing deal detection
- VIP / media routing
AI drafting workflow
- Conversation history loading
- Playbook retrieval
- Learnings retrieval
- Structured prompt building
- AI draft generation
- Draft parsing
Human approval system
- Slack approval cards
- Approve / edit / reject buttons
- Design approval paths
- Approval status updates
- Slack timestamp persistence
- Human-in-the-loop control
Sales action routing
- Stop cadence on reply
- Forward to sales when needed
- Follow-up action routing
- CX and intake action forwarding
- Payment event handoff
- Deal status updates
Follow-up operations
- Candidate selection
- Follow-up approval creation
- Reminder routing
- Cadence advancement
- Skip / edit / approve handling
- Sent message records
Revenue and delivery handoff
- Stripe payment event handling
- Deal won status updates
- Domain purchase callback handling
- Purchase status routing
- Delivery workflow triggers
- Slack delivery notifications
Monitoring and recovery
- Error catcher workflow
- Stuck build monitor
- Auto-retry routing
- Manual intervention alerts
- Reply reconciliation reports
- Weekly analytics digest
Reliability
- Explicit workflow stages
- Deterministic routing before AI actions
- State updates after key events
- Human approval before customer-facing sends
- Slack notifications for operational visibility
- Separate workflows for intake, drafting, approval, and payment handoff
Proof of production thinking
The workflow tracks customer, lead, reply, order, or approval state instead of relying on one-off executions.
Sensitive customer-facing actions are reviewed before being sent.
Unclear, invalid, or risky cases are routed away from automatic execution.
Failures generate logs or alerts instead of failing silently.
Inputs are cleaned and structured before any AI or downstream action.
The final output is written back into the tools people already use.
AI drafts the message, but humans approve the action.
The system uses AI to prepare responses and classify context, but the final control layer remains deterministic and human-reviewed. Customer-facing actions, payment-related handoffs, and sensitive routing decisions are kept outside the model and handled through explicit workflow paths.
- Inbound replies are normalized before routing.
- Existing deals are checked before new actions are created.
- Cadences can be stopped when a lead replies.
- VIP or media-related replies can trigger special routing.
- AI drafts are sent to Slack before customer-facing delivery.
- Reviewers can approve, edit, reject, or reroute the action.
- Feedback loops support redrafting instead of forcing a bad send.
- Payment events update deal state before downstream actions.
- Slack notifications keep humans aware of important workflow events.
- Stuck builds are detected and retried only when retry conditions are met.
- Manual intervention alerts are sent when auto-retry is not safe.
- Error reports are formatted and sent to Slack.
- Reply reconciliation catches mismatches between Instantly replies and deal records.
- Weekly analytics digests keep performance visible.
- Follow-up actions require explicit approve, edit, or skip decisions.
Faster sales follow-up, safer AI assistance, and better operational control.
Exact metrics are not public. The workflow outcomes below describe the operational changes the business consistently observed after rollout.
Inbound replies could be captured, matched, triaged, and prepared for review without manual copy-paste across tools.
Sales responses followed playbook context, conversation history, and approval rules instead of relying on ad hoc manual drafting.
AI helped prepare the response, but humans controlled approval, edits, rejection, and sensitive routing.
Slack cards, state updates, payment events, and logs made it easier to see what happened and what needed attention.
The same architecture can support outbound replies, follow-up cadences, deal updates, payment events, and human review workflows.
The system connected reply intake, AI drafting, approval, follow-up, and revenue handoff into one coordinated workflow layer.
What this build taught me.
- 01
AI sales systems need human approval before customer-facing actions.
- 02
Lead replies become much easier to handle when every message is tied to deal state and conversation context.
- 03
The best AI drafts come from workflow context, not from isolated prompts.
- 04
Slack can work as an effective human review interface when the backend tracks state correctly.
- 05
Revenue workflows need feedback loops, retries, and operational visibility — not just message generation.
Where this system goes next.
The same pattern, hardened for broader GTM operations.
- Unified dashboard for open replies, approvals, follow-ups, and payment handoffs.
- Queue-based retries for external APIs and Slack actions.
- Prompt regression tests for sales reply quality.
- Centralized approval analytics by reviewer, campaign, and reply type.
- Role-based permissions for sales, CX, and operations teams.
- Multi-campaign playbook management.
- Deeper CRM integration for pipeline and lifecycle reporting.