Client Status Report
Grow Big Ventures AI Email SDR System
Current implementation status for the email-only Instantly AI SDR/BDR system discussed in the client meeting. The operating workflow is intentionally simple: Instantly replies come in, Claude classifies and drafts where appropriate, Slack is used for approval or human intervention, approved replies are sent back through Instantly, and GoHighLevel supports CRM tracking plus calendar availability.
Executive Summary
SOW Reference And Current Interpretation
| SOW Area | What The SOW Says | Current Interpretation |
|---|---|---|
| Email channel | Email via Instantly only. HeyReach / LinkedIn is excluded. | Keep launch focused on Instantly replies and approved Instantly sends. |
| Response time | Target response time is approximately 30 seconds to 5 minutes, depending on tuning. | AI processing can be fast, but co-pilot mode also depends on human approval speed. |
| AI behavior | Claude Sonnet drafts direct replies that push toward booking and avoid unnecessary back-and-forth. | Positive and information-request replies should be concise, specific, and booking-oriented. |
| Booking | Use GoHighLevel round-robin calendar to propose times and support booking. | Interested replies should include the booking link plus two available calendar times. |
| Lead intelligence | Score and qualify leads, detect intent, and tag no-follow-up cases. | Store intent, action, risk, confidence, score, and suppression/follow-up state. |
| Co-pilot mode | AI drafts go to Slack for approve, edit, or reject before sending. | Slack should be the client-facing approval surface. Dashboard remains available for edit/audit. |
| Follow-up engine | Track booking-link sent but not completed and trigger re-engagement. | Needs GBV-approved follow-up timing and copy before production. |
| Training mode | Sandbox conversations and feedback-driven configuration without code changes. | The new CSV should become structured training examples, rules, and tests. |
| Commercial terms | $1,500 one-time email-only project fee; operational costs are separate. | Maintenance is covered by the partnership exchange in the SOW unless both sides agree otherwise. |
Current System Coverage
| Capability | Status | Coverage |
|---|---|---|
| Instantly inbound reply handling | Built, needs live credentials | Webhook handling, reply normalization, campaign/prospect mapping, and readiness checks exist. |
| AI classification | Built | Classifies intent, recommended action, risk, confidence, and lead score for replies. |
| AI response drafting | Built, needs training update | Drafting exists. The new training sheet should be applied before production traffic. |
| Slack approval | Built, needs production callback smoke | Approve/reject buttons route through the same approval system as the dashboard. |
| Human intervention | Foundation built | Handoff, internal tasks, and Slack notification patterns exist. GBV-specific triggers need final approval. |
| GoHighLevel CRM | Built, needs account test | Contact lookup/upsert, notes, tags, opportunities, tasks, and sync history are implemented. |
| GoHighLevel calendar | Adapter built, needs account wiring | Calendar availability and booking are supported. The live GBV calendar must be connected and tested. |
| Follow-up manager | Foundation built | Scheduling and cancellation exist. Production timing and copy need client approval. |
| Dashboard and audit | Built | Useful for conversation review, decisions, training, analytics, settings, and debugging. |
Training Data Processed
What was reviewed
- 237 total rows in the client training CSV.
- 233 usable lead reply plus AI response pairs.
- 87 rows include correction notes or special guidance.
- Rows are split across original examples, V2 examples, and a Final section.
Main findings
- Replies should be short, human, and booking-oriented.
- The agent should greet by name when available and never send only a standalone name.
- Information requests should use concise "for context" style positioning.
- GBV positioning centers on AI-powered email and LinkedIn outbound systems, owned infrastructure, at-cost setup, and aligned upside.
- The sentiment column is useful but not enough by itself; several rows need deeper intent classification.
Confusion Points To Resolve
| Topic | Why It Is Confusing | Decision Needed From Client |
|---|---|---|
| Not interested replies | The training sheet often sends a polite close plus booking link, while the live discussion suggested negative replies may not matter. | Should the AI ignore, send one polite close, or decide case by case? |
| Out-of-office replies | They are low value, but they can also indicate a future date to follow up. | Should the system ignore them or schedule a delayed follow-up? |
| Prospect sends their own calendar | One training example suggests sending the GBV link, but the meeting direction was to let a human book on the prospect's calendar. | Confirm this always routes to Slack human intervention. |
| Positive reply format | The training sheet mostly sends only the booking link; the meeting direction included two available times as well. | Confirm booking link plus two GoHighLevel available times. |
| Approved claims | The sheet uses claims such as 4x exited founders, at-cost work, $50k+ cash, and aligned upside. | Confirm which claims are approved for automated replies. |
| Tracked booking links | Examples use unique IDs in booking links. | Confirm whether every prospect needs a unique tracked link. |
| Wrong-fit prospects | Some examples disqualify public sector or fundraising requests; others still route toward a call. | Confirm when to disqualify versus send to human review. |
Access Needed
The client does not need to find API keys manually. The fastest path is temporary admin access to the right accounts so the delivery team can retrieve the technical IDs, configure webhooks, and run controlled tests.
| System | Access Needed | What We Will Configure |
|---|---|---|
| Instantly | Admin or manager access to the correct workspace. | API key, webhook, webhook secret, campaign IDs, mailbox IDs, test lead, and approved-send permissions. |
| GoHighLevel | Temporary admin/staff access to the correct location. | Private Integration Token, location ID, calendar ID, round-robin users, pipeline/stage IDs, and test appointment flow. |
| Slack | Permission to install/configure the Slack app in the approval workspace. | Approval channel, bot token, signing secret, interactivity callback URL, channel ID, and approver list. |
| Anthropic / Claude | GBV-owned API key or approval for managed pass-through billing. | Model configuration, monthly budget guardrail, and usage tracking. |
| Business rules | Final approval of reply behavior and claims. | Stop rules, follow-up policy, approved claims, escalation rules, and examples that override the sheet. |
Pricing And Operational Cost
SOW commercial terms
- Email-only project fee: $1,500 one-time.
- Phase 1 payment trigger: $750 after MVP delivery and confirmation.
- Phase 2 payment trigger: $750 after full deployment.
- Ongoing maintenance is exchanged for 100,000 emails/month of sending infrastructure.
- Operational costs remain separate from the project fee and maintenance exchange.
Operational planning range
- At 4M emails/month and 0.7%-1% reply rate, expected replies are about 28,000 to 40,000/month.
- Infrastructure and backups: roughly $80 to $230/month, depending on final hosting shape.
- Claude Sonnet classification-only or suppression actions can be below $0.01 when prompts are optimized.
- Classification plus drafted response should be planned at about $0.01 to $0.03 per AI action until real token logs are measured.
- Estimated AI reserve at 28k to 40k replies: about $300 to $1,000/month before prompt caching and suppression optimizations.
The original SOW states Claude usage is expected to be under $0.01 per interaction at typical volumes. Because the current plan includes richer training context, classification, lead scoring, and drafted responses at higher volume, the safer launch estimate is a usage reserve until production token logs confirm actual cost. Claude Sonnet pricing source: Anthropic Claude API pricing.
Recommended Next Steps
Resolve behavior questions
Confirm not-interested handling, out-of-office handling, own-calendar handling, approved claims, and follow-up timing.
Grant temporary access
Provide account access for Instantly, GoHighLevel, Slack, and Claude billing or API usage approval.
Apply training data
Turn the 237-row CSV into structured rules, examples, and sandbox tests for GBV-specific replies.
Run live smoke tests
Validate one Instantly inbound reply, one Slack approval, one approved send, one GHL contact sync, and one calendar availability test.
Start co-pilot production
Run human approval first, measure results, then selectively automate low-risk positive reply types.