Frontier system · Sales · 8 stages · updated 2026-10-23

Self-learning outbound machine

Starts from your best customers and ends with booked meetings. The agent defines the ICP, finds and enriches accounts, captures a real screenshot of each prospect's site or buying signal, builds a personalised mockup and email, sends, classifies replies, and rewrites its own copy and targeting from the results.

1Is this my problem?

Outbound depends on one or two people writing copy by hand, guessing at targeting, and never learning why a campaign worked.

You'll recognise this if…

Reply rates under 1% and nobody can say which line or segment caused it.
Reps spend mornings in Apollo and LinkedIn instead of on calls.
Every new campaign starts from a blank doc, with no record of past tests.

Cost of doing nothingEach month of guessing burns sending domains, lead credits and SDR hours on lists that were never going to reply.

Companies affectedB2B SaaS and services firms, seed to Series B, 2-15 sellers, selling to a definable ICP (ACV roughly $5k+). Common in dev tools, fintech, HR tech and agencies.

2Is this solution worth it?

Reported results

Saviynt: 60% of reps booked more qualified meetings and pipeline grew by $32.2M after giving reps Clay data inside ChatGPT/Claude via MCP (source)
Template 14211 only declares a winner after 200+ sends per variant and keeps a log of the last 10 experiments (source)

Quoted from the linked source. Results vary by company and setup.

Your ROI, with the assumptions shown

Prefilled with our illustrative assumptions for a typical team (25 h/week of manual work at $60/h, 50% automated). Change them to your numbers.

Build cost (estimate)$20k–$60k

Time to first version (estimate)6-12 weeks

AI agent stages4 of 8

Human checkpoints1

Limits and when not to build it

Skip it if you have under ~20 closed-won customers (no ICP to learn from), sell mainly through partners, or have fewer than a few thousand reachable accounts.

Guardrails it needs

A person approves every new template and the first sends of each campaign.
Safety floor: if reply rate drops below a set level, sending pauses and an alert fires.
Screenshots are real captures; no invented proof or fake familiarity.
Opt-outs honoured instantly; CAN-SPAM / GDPR rules apply.
Brand and claim rules live in a separate program doc the agent cannot edit.

3How does it work?

The agency builds you one internal tool, Outbound Studio, that runs the whole process. Watch it work on a sample case:

Outbound StudioIllustrative demo with sample data

Pick a sample on the left and press Run example to watch Outbound Studio work through a case step by step.

Your tool connects to

CRMEmailSlack / TeamsCalendarLead dataWeb browserDocsAI model

Most of it is ordinary automation. AI agents handle 4 of 8 steps, and people approve at 1 checkpoint.

Under the hood: stages, vendors and the learning loop

AI agent Automation Human checkpoint

1Define the ICP
Claude reads closed-won deals and CRM fields (industry, size, time to close, expansion) and writes a versioned ICP doc with the signals that separate best customers from the rest.
ClaudeHubSpot / Salesforce MCP
AI agentDecides: Which traits separate the best customers?
2Find accounts
Daily scan of Apollo and Clay against the ICP; waterfall enrichment for verified emails; dedupe against the CRM.
ApolloClayn8n
Automation
3Research and score
Per-account brief with Tavily/Exa signal search; fit score 0-100 with written reasons; picks one angle per account.
ClaudeTavilyExa
AI agentDecides: Score and pick one angle
4Capture the real site
Headless browser takes a real screenshot of the prospect's homepage or signal page. Never generated.
PlaywrightBrowserbase
Automation
5Build the mockup and email
Generates a personalised mockup or landing page from their screenshot plus a short email on the same angle, rendered from controlled templates.
ClaudeGPT image / Figma templates
AI agentDecides: Write copy that fits the angle
6Approve
Slack review queue shows brief, mockup and draft side by side. Approve, edit or kill.
Slack
Human
7Send and track
Sequencer sends from warmed domains; opens, clicks, replies and bounces flow back to the CRM.
InstantlySmartlead
Automation
8Handle replies
Claude Haiku labels each reply (interested, not now, unsubscribe, referral) and acts: booking link, 30-day follow-up, opt-out, referral ask.
Claude Haikun8nCal.com
AI agentDecides: Classify intent and choose the action

Learning loop

Every 6 hours a champion/challenger step reads Instantly stats, waits for 200+ sends per variant, and asks Claude to pick a winner and write one new challenger with a stated hypothesis, using the last 10 experiments as memory. Monthly, meetings and replies per ICP segment go back into stage 1 so targeting tightens over time.

Example run

Illustrative sample data showing what moves between stages. Not taken from the sources.

1
Define the ICP · in42 closed-won deals from HubSpot
outICP v7: B2B SaaS, 20-150 staff, hiring SDRs, uses HubSpot
2
Find accounts · inICP v7
out118 new accounts, 96 with verified emails, 22 dropped as CRM duplicates
3
Research and score · inacme-analytics.com
outScore 82: hiring 3 SDRs, pricing page changed last week. Angle: speed-to-lead
4
Capture the real site · inacme-analytics.com/pricing
outScreenshot captured 1440x900
5
Build the mockup and email · inScreenshot + angle
outMockup of their pricing page with an instant-callback widget; 74-word email
6
Approve · inReview queue item
outApproved with one edit to the subject line
7
Send and track · inApproved email
outSent 09:12; opened 09:40; replied 11:05
8
Handle replies · in"Interesting, can you send times for next week?"
outLabel: interested (0.94). Booking link sent; AE alerted in Slack

Vendors commonly used

ClaudeHubSpot / Salesforce MCPApolloClayn8nTavilyExaPlaywrightBrowserbaseGPT image / Figma templatesSlackInstantlySmartleadClaude HaikuCal.com

Where AI decides and where people do

AI agents decide

Define the ICP: Which traits separate the best customers?
Research and score: Score and pick one angle
Build the mockup and email: Write copy that fits the angle
Handle replies: Classify intent and choose the action

Human checkpoints

Approve: Slack review queue shows brief, mockup and draft side by side. Approve, edit or kill.

Maturity ladder

What each level adds. Most teams start at level 1 or 2 and climb. Highlighted: where this build sits.

Level 1Basic automation
A static list from Apollo goes into one sequence with a mail-merge first name.
Level 2AI-assisted
An LLM writes a first line per lead from their website; a rep reviews each batch.
Level 3Autonomous agent
Agents pick accounts from the ICP, capture the real site, build a mockup and email, classify replies and route them. Humans approve templates.
Level 4Self-learning system
Champion/challenger tests run every 6 hours and results flow back into the ICP and copy. The system writes its own next experiment.

Building blocks

Start with one of these smaller workflows and add stages.

Signal-based outbound with governed AI researchSaviynt: 60% of reps booked more qualified meetings, with a $32.2M increase in pipeline (as reported by Clay)

Ops builds enrichment and contact-finding workflows once (waterfall across many data vendors); reps call them from the AI chat they already use to get account context, buying signals and verified contacts..

SalesAdvanced~5–15 h/week$3k–$12k Competitor displacement research across your TAMA-LIGN replaced a $60K/year manual research contract; 2,000 accounts researched, 1M+ data points, and $6.8M pipeline generated or identified in 6.5 months (as reported by Clay)

For every target account, an AI web-research step checks which of your service categories they use, why, and which competitor provides it, then writes the answers into CRM fields reps can filter on..

SalesAdvanced~10–30 h/week$5k–$15k Closed-lost "wake the dead" campaignSendoso: 16 meetings, 23 leads and a closed deal within two weeks of launch, at a 75% open rate and 11% response rate

Pull last year's closed-lost deals and stalled meetings, have an LLM summarise why each prospect originally engaged from the pre-call notes, then send a short personal email that references the AE, the last meeting and that reason..

SalesStarter~3–6 h/week$1k–$4k Lead enrichment and routing in Slack

New HubSpot contacts or Slack-posted domains are enriched with firmographics, scored against your ICP and replied to in-thread with a summary, so the right rep picks them up..

SalesStarter~2–5 h/week$800–$3k Executive LinkedIn engagement to ABM pipeline

When someone from an ideal-customer account engages with an executive's LinkedIn post, a webhook sends the lead into a workflow that enriches, filters on qualification criteria, and stores qualified leads for sales follow-up..

SalesAdvanced~2–4 h/week$1.5k–$4k

Seen in the wild

Links checked 2026-10-23. The pipeline combines patterns from these public builds; we are not affiliated with these companies.

4Who can build it?

Matched on the services, platforms and industries agencies list on their own sites, and on their published projects.