AI Outreach Agent

The stack

What it’s built on, and how the pieces talk.

No dashboard is required to run it day to day; the operator’s interface is a Telegram notification. The pipeline itself is four cron jobs against one Postgres schema.

Layer 1

Discovery

Two-stage Apify scrapes by a curated list of local hashtags, with a follower window that shifts by time of day, then email recovery for profiles without one.

Apify TikTok + Instagramemail findergeo filterdedup

discover · 9 AM and 3 PM

Layer 2

Reasoning

A fast model writes the personal line and classifies replies; a stronger model drafts the reply that goes out under the client’s name.

Claude Haiku 4.5Claude Opusrules tabledate anchor

send every 15 min · read every 5 min

Layer 3

Delivery + ops

Sends from the client’s own mailbox with human spacing, watches the inbox over IMAP, and reports to the owner’s phone.

nodemailer SMTPimapflow IMAPVercel cron ×4Telegram Bot API

Wired in

Apify

Anthropic Claude

SMTP + IMAP

Vercel cron

Supabase Postgres

Telegram Bot API

Seeded variants

Guardrails

It has been rebuilt three times with increasing ambition: a Claude Code skill run by hand, then an always-on cron agent, then a reusable outreach module inside the CRM so any client can run it.

Where it stands todayThe numbers in this study come from the campaign’s real send and reply history. The fully unattended cron version is deployed and its schema is live, but the results cited belong to the agent’s first generation, run from a Claude Code session. We say that so the claim holds up if you ask.

Inside the product

The five surfaces of an agent that emails people.

The pipeline board lives in the CRM. The rest is what the creator and the owner actually see: the email, the reply and its verdict, the approval card, and the 5 PM digest. All invented names.

Pipeline

Every prospect on a seven-stage board.

Queued, sent, replied, interested, booked, posted, passed. The agent moves cards as replies come in and pauses automation the moment a human takes over a thread.

crm_outreach_prospects · stage · needs_you · automation_paused

Outreach · Pipeline

The email

One personal line, a real offer, sent at a human pace.

A short note from the owner with a single clause written about that creator, an offer of a comped tasting for two in exchange for one post, and a mention that top performers get paid. Sent from the restaurant’s own domain with 35 to 75 seconds between sends.

templates.ts · SUBJECT_VARIANTS · personal_line · seeded PRNG

Gmail · the cold email

The reply

Classified, date-anchored, drafted, and sent if it is safe.

The classifier tags the stage, checks geography and scans for unusual asks. The drafter resolves “this Saturday” against the day the creator wrote, confirms the visit, and auto-sends because the stage is safe and the daily cap has room. The booking is written back to the pipeline.

classify() · hasUnusualAsk() · date anchor · AUTO_SEND_SAFE_STAGES

Reply · classified and drafted

Approval card

When the agent is unsure, it asks, with buttons.

A private-event question trips the unusual-ask detector, so instead of guessing, the agent sends the owner a card: the message, why it was routed, the proposed draft, and Send, Skip, Edit. Batches get an Approve-all.

bot_drafts · Telegram inline keyboard · webhook callbacks

Telegram · needs you

Daily digest

The whole day in one message.

Leads scraped, cold emails sent, replies in and out, visits today, the pipeline funnel, and the next seven days of bookings, delivered to the owner’s phone at 5 PM.

cron-daily-digest · bot_daily_stats

Telegram · 5 PM digest

Why clients switch

The math a client sees before they say “I want this.”

Replaces $10,456.70 a year of software and outreach labor with $528 a year of running cost, gives back 26.8 hours for every 100 creators contacted, and it already produced 137 contacted, 29 replies, a 21.2% reply rate, 5 bookings at 1.28 touches per lead, and paid collabs closed.

Software you stop paying for

$10,457

per year, at mid-tier plans, across 3 tools. $31,370 over three years.

Hours given back

27 hrs

per year at 100 prospects, across 6 everyday actions. Worth about $1,339 at $50/hr.

Net position

$9,929

year one, after Agent running cost (build included in the marketing retainer) ($528). $29,786 over three years, before counting the hours.

The stack this replaces

Current published list prices, mid-tier plans, billed annually.

Modash (influencer discovery, vetting, contact data, tracking)Performance (mid tier of Essentials / Performance / Enterprise)$5,988 / yr

Instantly (cold email sending, sequencing, inbox rotation)Growth$451 / yr

Part-time virtual assistant (outreach labor: sourcing, vetting, writing, sending, reply handling, follow-ups, reporting)US VA at $25/hr, 160.7 hours/yr (6 campaigns x 100 prospects x 26.78 h)$4,018 / yr

Annual stack$10,457

AI Outreach Agent$528 once

Apify Starter $19/mo x 12 = $228 (https://apify.com/pricing), Anthropic API budgeted at $60/yr, Vercel Pro $20/mo x 12 = $240 (https://vercel.com/pricing). No separate build fee: the agent is delivered inside the monthly retainer, so only this pass-through runs.

Time saved at 100 prospects

Minutes to do it by hand in the old tools versus inside the system, scaled to a year.

Action	By hand	In the system	Saved / yr
Find and vet a local creatorOpen the hashtag or location feed, scroll to a post, open the profile, eyeball follower count, read the bio for a city or a neighborhood, scan the last few posts to confirm they actually live and shoot locally, check that engagement is not bought, tap the email button or copy the address out of the bio, paste name plus handle plus email into a spreadsheet. Roughly two or three profiles get opened for every one that survives vetting, and 5 minutes is the blended cost per kept prospect.	5 min	0.3 min	8 hrs
Write a personalized first emailRe-open the creator's profile, find something real to reference (a recent dish, a venue, a series they run), write an opening line that proves you looked, paste the offer body, adjust the subject so it does not read like a blast, proofread. Six minutes is the low end of the realistic 6 to 8 minute range.	6 min	0.2 min	10 hrs
Send with spacing and log itPaste into the mail client, set the subject, send, deliberately wait so the sends do not land in a burst, then go back to the sheet and log the date, mark the row contacted, and check the person was not already emailed from an older list.	2 min	0 min	3 hrs
Read and classify a reply, then draft a response with correct datesOpen the reply, work out whether it is interested, a question, a rate request, or a no, pull up the calendar to find real open dates, check the offer terms, write a reply that answers the actual question and proposes specific days, proofread, send. Eight minutes is the conservative end of the realistic 8 to 10 minute range.	8 min	1 min	2 hrs
Run 3-touch follow-ups on timeScan the sheet for who was emailed and when, work out who is due for touch two or touch three, re-read the original thread so the follow-up is not a copy-paste, write it, send it, update the row. In practice this is the step that quietly stops happening when the week gets busy.	3 min	0 min	1 hrs
Daily reportingOpen the sheet, count what went out, count what came back, check the inbox for replies that were never logged, update the booked column, and write the day's status somewhere the owner will see it.	10 min	1 min	2 hrs
Total hours given back per year			27 hrs

Runs while you sleep

⚡

Geo-filtered creator discovery

Scrapes hashtag and location feeds through Apify, filters to the follower band and the service-area radius, and recovers public business emails, so the list arrives already vetted instead of arriving as raw handles.

On demand at campaign start, then on a refresh cadence as new local creators post.

⚡

Personalized first-touch composer

Haiku writes an opening line from the creator's actual recent content and the agent selects a seeded template variant, so 100 emails read as 100 emails rather than one blast.

Immediately before each send.

⚡

Human-paced sender with dedup

Sends over SMTP with randomized 35 to 75 second spacing, writes an IMAP copy into Sent so replies thread naturally in the real mailbox, and blocks any address already contacted.

Continuously through the sending window.

⚡

Reply classifier and auto-responder

Reads each reply, classifies intent, and has Opus draft a response anchored to the real current date and live availability, auto-sending when the classification is safe and escalating when it is not.

On every inbound reply.

⚡

Follow-up cron

Runs the 3-touch sequence on schedule and suppresses anyone who has already replied, which is the step that reliably gets dropped when a human owns it.

Daily, on the schedule set per lead.

⚡

Daily Telegram digest

Pushes sends, replies, classifications, and bookings to the owner's phone, replacing the manual spreadsheet update.

Once a day.

All prices verified against the vendors' own published pricing pages on 2026-09-28 (September 2026).
Every replaced tool is priced on ANNUAL billing, which is the cheaper of the two published rates. Annual totals are the published yearly figure, not monthly x 12 marked up.
Influencer platform: Modash has three tiers (Essentials, Performance, Enterprise). Performance is the mid tier and is what is counted here at $499/mo on annual billing, $5,988/yr. Floor note for a conservative reader: Essentials is $199/mo, $2,388/yr, and swapping to it drops the annual stack total to $6,856.70.
Cold email: Instantly Growth is the plan that actually fits this volume (1,000 contacts, 5,000 sends/mo). Annual billing is $37.60/mo, $451.20/yr. The larger Hypergrowth tier at $358/mo was deliberately NOT used because the volume does not justify it.
Apify Starter is NOT counted as a saving. The agent uses the same Apify subscription, so it appears only in the agent's running cost. Counting it on both sides would inflate the number.
Labor is priced from measured hours, not a guessed retainer. 6 campaigns a year at 100 prospects each = 160.7 hours of manual work, at a US VA rate of $25/hr = $4,017.50/yr. A 5 hr/week part-time VA would bill $6,500/yr, so the derived figure is the conservative one. The $25/hr rate sits at the market average (ZipRecruiter reports $24.40/hr, Glassdoor $25/hr for US virtual assistants in September 2026).
Time savings are stated per 100 prospects. Reply-handling and follow-up volumes are derived from the campaign's real measured rates (21.2% reply rate, 1.28 touches per lead), not from a theoretical 3-touch maximum.
Owner time is valued at $50/hr for the headline labor figure. At the $25/hr VA rate the same 26.8 hours per 100 prospects is worth $669.58.
Anthropic API estimate: Haiku 4.5 at $1/$5 per million tokens for the personal opening line and the reply classifier, Claude Opus 5 at $5/$25 per million tokens for reply drafts. Metered math on 600 prospects and 126 replies a year lands near $16; budgeted at $60/yr to absorb retries and thinking tokens.
Vercel is budgeted at Pro ($20/mo, $240/yr) rather than Hobby, because cron and commercial use belong on Pro. Choosing the higher number keeps the net-savings figure honest.
Campaign results quoted verbatim from the live run: 137 contacted, 29 replies, 21.2% reply rate, 5 bookings, 1.28 touches per lead, paid collabs closed.
The build itself is not billed as a line item. It is included in the monthly marketing retainer, so only the agent's marginal running cost is charged through.

An outreach agent that books the table itself.

Creator outreach was a part-time job nobody had time for.