The Agentic Outbound Playbook · agentic GTM
Outbound, reinvented. Run by <AI agents/> from Slack.
Not one AI SDR. A team of six AI agents that research your market, write one-to-one emails, run the sending infrastructure and learn from every reply. People press a few buttons in Slack. This page opens every layer.
- 350+B2B teams served
- 35,000+mailboxes run
- 450,000+placement tests
# outboundyou + 6 agents- Outbound Leadpicks what to test next
- Researcherfinds companies and buyers
- Copywriterwrites one email per person
- Deliverabilitykeeps every mailbox in the inbox
- Launch Opsbuilds and launches campaigns
- Analystreads replies, writes learnings
Try it · our own campaign, step by step
Ask one question.
Watch six agents do the rest.
Send the Outbound Lead one message, then press the buttons it gives you. This is how every campaign starts in our Slack.
Ilya 
MaxWant this team running on your outbound?
We set it up with you, on your data, in your Slack.
The end state · what outbound looks like now
Outbound is now
a few clicks in Slack.
No SDRs doing research by hand. Nobody copying rows between tools. A team of AI agents connects to every tool, does the work end to end and checks its own output at every step. People decide. Agents do the rest.
Four decisions per campaign
- Run Apick the test
- Approveapprove the copy
- Launchstart sending
- Approve spendnew mailboxes, new data
Everything that costs money or speaks for your brand waits for a click. Nothing else does.
What the agents do, with a check at every step
- Research
40+ sources, about 7 pages read per company
✓ a verdict and a written reason
- People
real buyers, checked live on LinkedIn
✓ 3 email checks plus catch-all
- Copy
one email per person, from a dossier
✓ swap test, claims check, inbox render
- Sending
premium Google and Azure, 4 warm-up pools
✓ placement tests every day
Limitsthe governor sets every mailbox's limit nightly
✓ six limits, the strictest wins
- Learning
the Analyst reads every reply and result
✓ insights written back to the brain
- Capacity
Work that needed an SDR team runs overnight.
Research, list building, writing and checks happen while you sleep, every night.
- Scale
New segment, new market, no new hires.
Add a vertical or a hundred mailboxes without adding headcount.
- Quality
Scale doesn't dilute it.
Every step has a check before the next one starts, at 100 emails or at 100,000.
See how it would look on your company.
Get a free strategy for your market, or talk it through with us.
Who wrote this · the people behind it
We've seen outbound from every side.
Ilya was one of lemlist's first hires, then head of growth at Expandi. Max was CMO at Expandi. Now we run Growth.Band, an agentic GTM agency: outbound for hundreds of B2B teams, and for ourselves. This year we rebuilt our own company from scratch as an agentic one. This page is the playbook of what we built.
- 350+B2B teams served
- 35,000+mailboxes we manage, for clients and for us
- 450,000+placement tests analysed
- ~10 yearsin outbound
30 minutes. We look at your outbound and say honestly where we'd start.
The recording · 45 minutes, with a live demo
Give me 45 minutes.
I'll show you how agentic GTM works.
Every layer of this page explained, with the agents running our own outbound live in Slack.
Your GTM today · new tools, old rails
Your GTM still runs
on 2021 rails.
More tools, more AI, a few new hacks. Underneath, the same system: data nobody trusts, campaigns nobody owns, lessons that never reach the next email.
Your CRM says she's your champion. She left 8 months ago.
- Q3 push v2 FINAL
- copy of Q3 push
- Founders test
- Old list, new copy
Nobody can say what worked, or what to launch next.
"Price wasn't the issue. Onboarding time was. If you'd led with that, we'd have signed in May."
The deal closed. What you learned on the call went nowhere.
> write a cold email to {{first_name}}Hi {{first_name}}, I hope this email finds you well. I noticed {{company}} is growing fast…
AI column · prompt_v7 (copy) (copy)Context limit reached. Start a new chat.One prompt, one task. Emails get faster, not better.
Each AI tool fixes one task.
The system stays the same.
Your team in agentic mode · one operator, a team of agents
Grow revenue,
not headcount.
In 2021 you scaled GTM by hiring: a researcher for data, SDRs for volume, RevOps for the CRM. In agentic mode, one GTM operator with agents does the work of 5 to 10 people, depending on your processes.
- SDR
- SDR
- SDR
- Data research
- RevOps
- Copywriter
- Deliverabilitycontractor
- Agencyhandoffs
- people
- 9
- tools
- 6
- shared context
- 0
- Outbound Lead
- Researcher
- Copywriter
- Deliverability
- Launch Ops
- Analyst
- operator
- 1
- agents
- 6
- shared brain
- 1
The real problem · tool hopping
You don't need another tool.
You need a system.
Replies drop, so the team switches tools. Instantly to Smartlead because the dashboard feels better. Apollo to Amplemarket because it's multichannel. Six months later the reply rate is lower than when you started, because the tool was never the problem.
Reply rate over 12 months: replies rise after launch and fade; each of three tool switches gives a short bump, and every peak is lower than the one before.
- Start outboundFirst campaigns get replies.
- Results fadeLists wear out, mailboxes tire, copy goes stale.
- Switch the toolInstantly to Smartlead, Apollo to Amplemarket. A short bump.
- Fade againNothing underneath changed. Back to step 2.
Every lap costs months, money and burned domains. Stop running laps.
Reply rate over 12 months: a steady climb, step by step — winner scaled, brain learns, limits hold, next test.
- Set up once. Data, copy and infrastructure, chosen and connected for you.
- Controlled. A limit for every mailbox, every night. Placement tested daily.
- Run by method. Every campaign is a test with its rule set before launch.
- Learns on its own. Every reply and result feeds the next campaign.
- You press four buttons. Agents do the rest.
So here is the system: four layers.
A shared brain at the bottom, our method and the sending rails on top of it, and you pressing the buttons that matter.
- Pick the test
- Approve copy
- Launch
- Approve mailbox spend
- Outbound Lead
- Researcher
- Copywriter
- Deliverability
- Launch Ops
- Analyst
- hypothesis matrix
- dossier + swap test
- 10-day rule
- rulebook
- InboxBand, dedicated instance
- isolated IP
- premium Google
- Microsoft on Azure
- 4 warm-up networks
- daily placement tests
- limit governor
- Capturecalls, replies, CRM, campaign stats, chat
- StructureICP, offers, objections, learnings
- Useagents read it before every task
- Learnevery result flows back in
The playbook · eight chapters
Every part of the system,
problem first.
Each chapter starts with what breaks in most teams, then shows what we changed and how it works. Jump to the part you care about.
- 01
Agents
Problem AI tools do one task, then forget.
Fix six agents on one framework and one memory.
- 02
Second Brain
Problem every chat starts from zero.
Fix one memory people and agents read first.
- 03
Data research
Problem wrong companies, no signals.
Fix 40+ sources, sites read, buyers checked live.
- 04
Copy
Problem templates get 0.5 to 1% replies.
Fix one email per person, from a dossier.
- 05
Launch Ops
Problem tests that prove nothing.
Fix a test matrix with the rule set first.
- 06
Analyst
Problem results nobody reads.
Fix patterns and insights, written back.
- 07
Infrastructure
Problem shared warm-up, a fake 100%.
Fix dedicated instance, 4 pools, daily tests.
- 08
Governor
Problem volume jumps burn domains.
Fix a nightly limit for every mailbox.
450,000 tests: what we learnedThe architectureThe rules we run onThe 45-minute recordingYour free strategy
Three ways to start · pick what fits you
Start free.
Or start building.
- Free · 24 hours · no call
Get a free strategy
Answer 8 questions. Get your target audience, market size, signals and example emails for your company.
- Free · 20 minutes · no pitch
Free call with a founder
Bring your numbers. We look at your outbound together and say honestly what we would fix first.
- Agentic Outbound · 90 days
Build it with us
The whole system from this page, on your data, in your Slack. We build it with you, then run it together.
Chapter 1 · Agents · what they stand on
Agents do the work.
You press the buttons.
An AI SDR, or Claude wired to a few MCPs, is fast and has nothing under it: no company context, no rules, no memory of last month. Our six agents stand on four things, and every campaign makes those four things better.
Outbound Lead
strategyReads the brain, proposes three bets, briefs the team.
▶ you pick the test3 bets ranked by evidence
Researcher
data researchMaps the market, reads every site, finds and proves people.
runs on its own6,870 → 2,140 people
Copywriter
copyDossier per contact, one-to-one emails, QA before you see them.
▶ you approve copy4 segments × 2 angles
Deliverability
inbox placementBuilds and guards the infrastructure, decides limits nightly.
▶ you approve spend97.8% Google · 96.1% Outlook
Launch Ops
testingTurns a hypothesis into a test matrix, scales the winners.
▶ you launch8 campaigns · 6,420 emails
Analyst
replies & resultsFinds patterns across campaigns, writes the learnings.
feeds the brain3 insights a week
standing on
- 1 · Second Brain
Your context
Calls, replies, CRM, past campaigns, compiled into six pages.
- 2 · Framework
Our rulebook
Ten years of outbound written down as rules: ICP, signals, offers, copy, sending limits.
- 3 · Data
The market
Dozens of sources, every company read and qualified, every email proven.
- 4 · Live results
What just happened
Every reply and every campaign result flows back, so the next campaign starts from it.
Numbers from today's demo run, Campaign A · HubSpot integration partners. Rehearsal data, nothing is sent.
Want these six agents in your Slack?
We set them up on your data, with your rules and your buttons.
Chapter 2 · Second Brain · the problem
Everyone uses AI.
Almost nobody is AI-native.
Your team writes prompts all day. Each one solves one task, and then everything it learned disappears. Three reasons why.
Write the follow-up for Northwind. Same angle that worked last week.
Happy to help! Could you share some context about Northwind, your offer and what worked last week?
Previous chat: "This chat reached its maximum length. Start a new chat to continue."
ICPofferlast 40 replieswhat won
Every session starts from zero. Yesterday's context dies with the chat.
- prompt_cold_v7_FINAL.txtIlya
- icp-notes (copy).mdIlya
- objections_q3.docxMax
- claude-project · "outbound"private
- call-notes-oct.txtSDR, left
Shared with: nobody. The agents can't read it either.
Notes and prompts live on one laptop. The team and the agents never see them.
- try 1Too generic, "growing fast"
- try 2Wrong offer for this segment
- try 3Claims we can't prove
- try 4One good email
- next weekStart again from try 1
You can get one good email in four tries. You can't get emails that improve every week.
A good email is a task.
Emails that get better every week are a process.
A process needs memory.
Second Brain · what we built
One brain the whole team
and every agent reads.
Everything your market tells you is captured automatically, compiled every night into six pages, and read by people and agents before they act. What happens next flows back in.
How the Second Brain works
- 1 · Capture · automatically
- Sales callsevery 30 min
- Replieshourly
- CRM & dealsnightly
- Campaign resultsnightly
- Team chatevery 30 min
- Founder-led B2B teams, 20 to 200 people, hiring their first SDR12 calls
- Why now: a new sales leader or a funding roundCRM
- Agentic Outbound for teams that want the whole channel run by agentsoffer
- Proof used in copy: CRM-2.1, 5.4% reply rateresults
- Plain words, one idea per email, one askrulebook
- Never "hope this finds you well"47 replies
- "We already use Apollo" → lead with signals, not data9 calls
- "Not now" → back on the list when the timing changesCRM
- Oct 14 · founder angle beat the hiring angle in segment Bcampaign
- Oct 9 · onboarding time matters more than pricecall, Marc O.
- Q4 reactivation of lost deals · scale above 30% interestthreshold
- HubSpot partners, new angle · running, verdict in 10 dayslive
↳ every line links to the call, reply or result it came from · illustrative
- 3 · Use · people and agents
- Copy & sequences
- Reply handling
- Campaign bets
Call prep for your team
4 · Learn · ↺ every outcome goes back into the brain, nightly · people decide, the system remembers
Was: It forgets
It remembers
Every call, reply and result is stored word for word and compiled every night. Nothing depends on one chat window.
Was: It's local
It's shared
Plain files in your own repository. Your team opens the pages, your agents connect with a scoped key. One source for everyone.
Was: One task
It's a process
Learnings move the other pages. A new objection updates Objections; a campaign that won updates Offers. Next month starts smarter.
Second Brain · why you can trust it
Four rules that keep
the brain honest.
The fear with AI is an agent inventing a price or a promise, or someone seeing what they shouldn't. We solved both with structure, not with a better prompt.
Sources word for word: full call transcripts, emails, CRM snapshots. Never edited, never deleted.
- Writes
- collector robots only
- Reads
- AI and people
The six pages. Every number, date and quote links to the raw file it came from.
- Writes
- the nightly AI
- Reads
- team and agents
Drafts and thinking of the founder or team lead.
- Writes
- a human only
- Reads
- no AI access
You can always go back to the source
Objections page"Price came up in 4 of 9 calls in August; resolved 3 times."
Call transcript · 14 Aug · 23:41"…honestly the price is the thing I'd need to sell internally…"
Transcripts are stored in full, not as someone's summary. Every claim on a page carries a link, and a check runs on every change: a figure or quote that can't be found in a source fails the page.
Access by zone, not by trust
| Who | Strategy & clients | Finance | Ops & code | Other clients |
|---|---|---|---|---|
| Founder | yes | yes | yes | no |
| Sales rep | yes | no | no | no |
| Copywriter agent | read | no | no | no |
| Analyst agent | read + propose | no | no | no |
- Founder
- Strategy & clients
- yes
- Finance
- yes
- Ops & code
- yes
- Other clients
- no
- Sales rep
- Strategy & clients
- yes
- Finance
- no
- Ops & code
- no
- Other clients
- no
- Copywriter agent
- Strategy & clients
- read
- Finance
- no
- Ops & code
- no
- Other clients
- no
- Analyst agent
- Strategy & clients
- read + propose
- Finance
- no
- Ops & code
- no
- Other clients
- no
Each zone is a separate repository. People and agents get a scoped key: issue it to connect, revoke it to disconnect. Every request is logged: who asked what, and when.
Money and promises need a person
- Prices
- Commitments
- Decisions
- Legal
An agent can propose a change to these pages, with the quote and the link. It cannot write one. A person approves, or it never lands.
✓ Human approvalLive numbers are never stored as text
Live (revenue, reply rates) is always read from the source system. Status (deal stage, mailbox count) is written by a robot into a fenced block. Judgement (ICP, offers, playbooks) changes when evidence arrives, not on a schedule.
So the brain can't hold a nice, stale "we have 32 clients" that nobody remembers to update.
Second Brain · the rhythm
It runs every day.
You give it fifteen minutes a week.
Collecting is done by simple, reliable scripts. Thinking is done by AI once a night. If the AI fails one night, nothing is lost: the sources are already saved and get compiled the next night.
- 02:30 · 03:00snapshots, then synthesis
- 09:00morning brief
- collectors every 30 min · your replies picked up every 10 min
- collector run
- AI step
Collect
- Call transcripts, in full
- Slack channels
- Mailbox replies, hourly
- No AI touches them
Compile
- Reads only what is new
- For each source: new page, add to a page, contradiction, or nothing
- Prices and promises become proposals
- Rebuilds the index, runs the checks
Brief
- What changed overnight
- What needs a decision
- How much is waiting in the inbox
- Reply in one word to merge or reject
Clean
- Sunday: a cleanup proposal for contradictions, stale and unlinked pages
- 1st of the month: a devil's advocate pass looks for evidence against recent claims
- Monthly 10-question interview on strategy
Brain · morning brief · 09:00
Changed overnight
- 3 new calls compiled → Objections: "we already get leads from the directory" added (2 calls)
- Learnings: founder-led segment beat agencies with SDRs (campaign A, day 10)
Needs your decision
- Offers page: a rep quoted a different volume on Tuesday's call. Keep, change or flag?
Every page stays current without a meeting
The robot collects, the AI compiles, a person decides where a mistake costs money. Because cost grows with new material, not with the size of the brain, it stays cheap to run as it grows.
Example brief · illustrative
Second Brain · how others build one
Not a notes app.
Not a chatbot over a pile.
"Second brain" usually means one of two things. Both are useful. Neither changes what your team sends tomorrow.
 Company brain for agentswhat we build | Personal second brainnotes app + capture, organize, distill, express | Upload and chata pile of docs with a chatbot on top | |
|---|---|---|---|
| Who fills it | Collectors, automatically, from calls, mail, CRM, campaigns | You, by hand, when you remember | Whoever uploads files |
| Shape | Six pages on how your company sells | Folders and notes for one person | An unstructured pile |
| Can you trust a number | Every claim links to its source, checked on every change | Only if you wrote it | Fact and guess look the same |
| Shared | By zone, with scoped keys for people and agents | Personal | The whole pile or nothing |
| Output | Copy, replies, call prep and campaign bets | Your own thinking | An answer, once, to one person |
| Gets better | Every result is written back as a learning | Only if you keep it up | No |
Your calls and CRM already hold the answers.
Bring one recorded sales call. We'll show you what the brain would write from it.
Chapter 3 · Data research · the problem
Your list is why outbound
"doesn't work".
Pull a list from Apollo, ZoomInfo or Clay and you hit the same three problems. The first one everybody knows. The other two quietly kill the campaign.
| Name | Title | |
|---|---|---|
| Dana W. | VP Sales | dana@acme.combounced |
| Marc O. | Head of Sales | marc@brightl… |
| Priya R. | COO | p.r@lumen.iobounced |
| Tom H. | CRO | tom@kite.co |
Marc O. started a new position at Nimbus · 8 months ago
Some people left. Some emails bounce. Annoying, and the easy part.
Their own website: not a HubSpot partner at all.
Many companies don't do what the filter says. A category label is not what a company sells.
| Company | Hiring? | New leader? | Why now? |
|---|---|---|---|
| Northwind | ? | ? | ? |
| Brightline | ? | ? | ? |
| Lumen Labs | ? | ? | ? |
| Kite Health | ? | ? | ? |
| Acme Freight | ? | ? | ? |
Nothing says this company, this person, this month. Without a reason to write, every message is generic.
What a bad list does, step by step
- WRONGCOMPANIESfilters, not websites
- GENERICMESSAGEno signal to write from
- NOREPLIES0.5 to 1% reply rate
- SPAM FLAGS,BURNED DOMAINScomplaints hit the domain
- "OUTBOUNDDOESN'T WORK"
Data research · what we changed
You describe the customer.
We hand back the people.
First we find far more than you need. Then we merge it, read it, judge it and check it, until only the right people are left. In the demo it takes seconds; a real run takes a few hours.
- 8,960 listings found across sources
- 6,870 companies after merge and de-dup
- 3,940 in target countries, sites read
- 2,310 qualified as a real fit
- 2,140 people, proven and signalled
Step 1 · find more than you need
Dozens of sources, searched as one
Every database is strong somewhere and blind somewhere else. So we don't use one. The Researcher maps what each source can filter on, builds up to 20 filter variations per source, pulls every list, merges duplicates and drops broken and dead companies.
40+sources, not only the ones you already pay for
- Apollo
- ZoomInfo
- Clay
- Crunchbase
- BuiltWith
- G2
- Exa
- Wellfound
- Indeed
- Glassdoor
- Perplexity
- HubSpot
- GitHub
- Clutch
- Capterra
- Product Hunt
- databases
- partner directories
- review sites
- job boards
- tech detection
- marketplaces
- web search
- our own base
mergeone de-duplicated list
Step 2 · read the website
We read what they say about themselves
For every company we open 6 to 10 pages, about 7 on average. When a site blocks simple requests, we climb to a real browser and full automation. A "checking your browser" screen counts as a failure, never as a description.
~7pages read per company
northbeck-ops.com- Home✓
- Services✓
- Case studies✓
- Partners✓
- Pricing✓
- Team✓
- Careers✓
- Blogbot wall
- Request≈80% of sites end here
- Browsersites drawn in JavaScript
- Automationsolves challenge pages
- Anti-botthe last few percent
Step 3 · qualify with an LLM
A verdict with a reason you can argue with
An LLM reads those pages against your brief and decides what the company really does. Every company gets a verdict, a confidence score and a written reason. Change the brief mid-order and it applies from the next company.
3,940read → 2,310real fits
Northbeck OpsFIT · 0.91- HubSpot-only partner agencyservices page
- 11 Salesforce to HubSpot migrationscase studies
- Two SDRs on the teamteam page
- 34 people, inside 5 to 200careers
Reason: sells HubSpot implementation and migrations, already has a sales function, matches segment "HubSpot-only, with SDRs".
Brightpath · not a fit: Shopify agencyKestrel · not a fit: existing customer
Step 4 · find the buyers
Wide search, then judgement
Job-title filters miss most buyers. So we search wide across several people sources, cheapest first, and stop as soon as we have enough. Then an LLM reads every title and decides who really buys.
~3decision-makers per company
- Daniel K.Founder & CEObuyer
- Sara L.Head of Partnershipsbuyer
- Emma T.HubSpot Consultant · delivery, not buyingnot a buyer
- Lucy M.Marketing Internnot a buyer
- Our own database2 people found$0.00
- Apollo3 more · 5 people, enough. Stop.$0.014
- Sources 3 and 4never called·
Step 5 · live LinkedIn check
Still there, still that job, still that company
Databases keep people for months after they leave. Before anyone gets an email, we open their LinkedIn live and check three things: they still work there, the title still matches, and the company on the profile is the company we qualified.
Live profile check · today
- Daniel K.Founder & CEO at Northbeck Ops · since 2019✓ still there
- Sara L.Head of Partnerships at Northbeck Ops · since 2023✓ still there
- Owen P.VP Sales at Kestrel → now at Arden Labsleft in May · removed
Step 6 · relevance signals
A reason to write this month
Signals turn a list into a list worth writing to. For each company and person we collect what changed recently: hiring, posts, leadership moves, events, tech. They are dated and sourced, and they become the dossier the Copywriter writes from.
- Oct 14Job post: HubSpot adminhiring
- Sep 22Post: "10 new retainers this year"post
- Jul 03New CMO joinedleadership
- Oct 09Speaker, HubSpot user groupevent
- Aug 18
Added a CPQ tool to the sitetech
Step 7 · prove the email
Found and proven are different words
A bounce damages your domain for months. So every address goes through a chain of checks, including catch-all validation for domains that accept everything. One confirmation from a provider is enough to keep it; one rejection is enough to remove it.
- Format and domainsyntax, MX records existpass
- Mail server handshakethe mailbox answerspass
- Verifier 1independent provideraccepts mail
- Verifiers 2 and 3only called when 1 is unsurenot needed
- Catch-all validationdomain accepts everything? test the real mailboxreal mailbox
- 2,140 people for Campaign A
- 4 segments, each with its own story
- 3× email verification, catch-all checked
- 12.3M companies in our own base, queried first
Curious what this finds in your market?
Your free strategy names real target companies and the signals behind them.
Chapter 4 · Copy · the problem
AI made emails faster.
Not better.
Copy decides two things: whether people answer at all, and whether your next campaign is better than this one. Most teams lose on both.
- Alex · LeadlyQuick question, Daniel. I noticed Northbeck is growing…
- Sam · PipewiseDaniel, scaling outbound? We help B2B companies…
- Jo · GrowthlyHi Daniel, I hope this email finds you well…
- Kim · SalesAIFollowing up on my last note about…
0.5–1%reply rate, and only from people who already had the problem today
If your email reads like the other four, it gets treated like them.
"Hi Daniel, I noticed Northbeck uses HubSpot and is growing fast. Companies like yours often struggle with…"
True for 10,000 other companies. Feels personal to the model. Not to Daniel.
A merge field with an adjective is still a template.
The AI tool works inside one task. Nothing makes the next emails better than these.
Copy · what we changed
Every email is written from
facts about one person.
Eight steps from the brain to the inbox, and one back into the brain. You see the result in Slack; this is what happens before it gets there.
Context from the Second Brain
Start from what already worked
The Copywriter reads the ICP, the offer, the messaging DNA, the objections from sales calls and the last 30 days of learnings: which lines got replies, which got "not now".
- ICP · Offers · Messaging
- 34 sales calls
- 312 classified replies
41 past campaigns- Learnings, 30 days
Extra research when context is thin
A new vertical gets researched first
For an audience we've never sold to, the brain isn't enough. The Copywriter researches the vertical: the problems these companies talk about, the dates that matter to them, what peers say in reviews and posts.
- Partners depend on the directory for new workpartner forums · reviews
- Q4 is when agencies plan next year's pipelineindustry posts
- Migrations are the bigger-ticket servicecase studies across 40 partners
The brief
Three insights the copy must carry
Context and research become a short brief: who, which problem, which offer, which proof, and the three insights to lead with.
- Directory leads plateau: partners can't double them on demandfrom the brain · objections and calls
- Migration buyers show up months ahead, through hiring and leadership changesfrom research · learnings
- Offer: signal playbook + 20 companies, sent within an hour of a replyfrom the brain · offers
Deep research per contact
A dossier on every person
For each contact we look for five to ten facts and stop at five. Each fact has a source. Some go into the first email, the rest are kept for the follow-ups.
DanielFounder & CEO · Northbeck Ops- 11 Salesforce to HubSpot migrationscase studies
- Two SDRs on the teamteam page · LinkedIn
- HubSpot-only partner agencyservices page
- HubSpot admin job postcareers · kept for follow-up
- Spoke at a HubSpot user groupevents · kept for follow-up
5 signals found → search stops
Write
Signals combined into one email
The email combines the dossier facts that fit the brief. Different people get different facts in a different order, so no two emails are the same.
To Daniel · Northbeck Ops migration signals for your SDRs
Hi Daniel,
Northbeck lists 11 Salesforce to HubSpot migrations1 in its case studies, and you have two SDRs on the team2.
Migration buyers show up months ahead: a new CMO, a HubSpot admin job post, a RevOps hire. Most SDR lists use none of it.
I put together the signals our HubSpot partner clients use to find migration deals, plus 20 companies near you showing them right now. Want me to send it over?
QA
Checked before you ever see it
Every claim is checked against proof on file. The swap test sends Daniel's email to Rachel on paper: if it still works, it's a template and it gets rewritten.
- Every claim has proof on file
- One idea, one ask
- No spam words, no broken links
- Merge fields filled on real rows
- Swap test: rewrite if it still works
- Matches segment and offer
Swap test · Daniel's email sent to Rachel0 / 2facts are true for her. One-to-one. It ships.
If it still works after the swapTemplateIt's a template in disguise, and it gets rewritten.
Render check
How it looks in their inbox
Before launch, every email is rendered on real rows: the shortest and longest names, empty fields, the preview line in Gmail and Outlook, the phone screen. Broken merges and ugly previews are caught here, not by the prospect.
Inbox · preview
- Lena · PipedriveYour invoice for October9:12
- Ilya Azovtsevmigration signals for your SDRs · Northbeck lists 11 Salesforce to HubSpot…9:40
- HubSpotPartner newsletter · October8:05
migration signals for your SDRsIlya Azovtsev · to meHi Daniel, Northbeck lists 11 Salesforce to HubSpot migrations in its case studies…Final version and the loop back
You approve. The results teach the brain.
You see the final sequence in Slack and press approve. After launch, what got replies and what didn't is written back into Learnings, so the next brief starts from it.
- you approve in Slack
- launch
- replies classified
- what worked, what didn't
- Learnings page
- next brief
Email from today's demo run. Contacts and research items are illustrative.
See an email written this way for your buyer.
Your free strategy comes with example emails for real people in your market.
Chapter 5 · Launch Ops · campaigns as a test matrix
Most teams don't test.
They launch and hope.
Five campaigns, five lists, five versions of copy, launched in different weeks. When one works, nobody can say why. A/B tests answer one question once; they don't create continuous learning. Launch Ops designs the test before anything is sent.
Before · what usually happens
| Campaign | List | Copy | Week | Reply |
|---|---|---|---|---|
| Q3 push v2 FINAL | Apollo, new | v4 | 32 | 2.1% |
| copy of Q3 push | old list | v2 | 35 | 0.8% |
| Founders test | Clay | v5 | 36 | 3.4% |
| Agencies Oct | mixed | v5b | 40 | 1.2% |
Everything changed at once. Was it the list, the copy or the week? Nobody knows, so nothing is learned.
After · one hypothesis, a designed test
| Segment | Migration deals | Support retainers |
|---|---|---|
| HubSpot-only, with SDRs230 people | keep watching | stop · learning |
| HubSpot-only, founder-led520 people | winner · scale | keep watching |
| HubSpot + other, with SDRs340 people | stop · learning | keep watching |
| HubSpot + other, founder-led1,050 people | keep watching | winner · scale |
One variable per axis: who (segment) and what (angle). Segments from the demo run, outcomes illustrative.
- A · New
HubSpot integration partners
WhoFounders and BD heads at partner agencies, 5 to 200 people
ProblemSteady inbound from the directory, no way to double it next quarter
- B · Proven
Companies hiring SDRs
WhoFounders, Heads of Sales and RevOps with an open SDR role
Evidence3.8% reply · 1.0% interested
- C · Warm
Reactivation
Who47 lost deals and 30 positive replies that went quiet
Why nowThey said "not now" in Q2 and Q3; Q4 planning is happening now
- Day 08 cells launched small
Every segment × angle gets a small batch. Each campaign built as a draft, 8/8 renders checked.
- Day 1 to 10Results come in
Replies classified daily. Losers can stop early when they clearly fail.
- Day 10The verdict
At least 1 positive per 250 emails wins. The rule is set before launch, not after.
- Day 11Winners get the leads
The next 2,000 contacts go to the winning cells. Verdicts go to the brain and the Analyst.
Chapter 6 · Analyst · patterns, insights, the next bet
Dashboards show numbers.
The Analyst tells you why.
It reads every reply and every campaign, finds patterns across them, and comes back with three insights and what to change. Then it writes them into the brain and talks them through with Outbound Lead and with you.
- Founder-led · migrations2.2
- Generalists · retainers1.8
- With SDRs · migrations0.9
- With SDRs · retainers0.4
- 1Founder-led agencies reply about twice as often as agencies with SDRs. The founder still owns new business.
- 2"Migration" wins with HubSpot-only partners, "retainers" with generalists. Same segment, different story.
- 3New objection in 6 replies: "we already get leads from the directory". Added to Objections with a suggested answer.
Into the brain
Insights are written to Learnings with dates and links to the replies they came from.
To Outbound Lead
The next bet starts from them: founder-led, migration angle, the new objection answered up front.
To the Copywriter
The next brief carries the winning lines and drops the ones that lost.
To you
One message on Monday. You decide what to apply; nothing changes without a click.
Example review · illustrative numbers
Want this test-and-learn loop on your outbound?
Every campaign teaches the next one. We run it with you.
Chapter 7 · Infrastructure · the problem
Your dashboard says 99%.
Your prospects see spam.
Most teams take the same path: buy cheap domains and mailboxes inside their sending tool, warm them up there, and trust the number the tool shows. Here is what happens next.
A shared pool means your warm-up partners are other cold senders on cheap domains. It often makes reputation worse, not better.
Warm-up stays flat at 100% from month 1 to month 4 while real inbox placement falls steadily.
After a few months every mailbox in it has seen your emails and marked them "not spam". Warm-up says 100%. Real inboxes disagree.
- tool dashboard
- 99%
- reply rate, 6 weeks
- ↓ 60%
- Google50%
- Microsoft8%
You only find out when it's already broken.
Replies fall, the dashboard is green. A test finally shows spam at Microsoft and half at Google. Now you rebuild from zero.
Nobody replies from the spam folder.
And nobody tells you when you got there.
Infrastructure · part 1, the setup
Built right on day one.
Sending after three weeks.
Our Deliverability agent does two jobs. The first is building the infrastructure, step by step, so every email starts from a clean reputation.
- 1 · Domains
Lookalike domains, in your name
Names close to your brand, never your main domain. Each one redirects to your website through a captcha, so scanners and blocklists see a real site and nothing to flag.
- trynorthbeck.comnorthbeck.com via captcha
- northbeck-hq.comnorthbeck.com via captcha
- getnorthbeck.ionorthbeck.com via captcha
✓ blocklist check before purchase
- 2 · Mailboxes
Premium Google, or Microsoft on enterprise IPs
Premium Google Workspace mailboxes with DNS on Cloudflare, or Microsoft Azure SMTP on enterprise-level IP addresses that get through Proofpoint and Mimecast.
- Google Workspace
- DNS on Cloudflare
- Microsoft Azure
- 3 · Identity
Authenticated and human
SPF, DKIM and DMARC set and checked every night for drift. Signatures, names and sender details set on every mailbox, so it looks and behaves like a real person.
- SPF ✓
- DKIM ✓
- DMARC ✓
- Signature ✓
- Sender name ✓
- 4 · Sending platform
InboxBand, dedicated instance, isolated IP
Mailboxes connect to InboxBand on a dedicated instance with an isolated IP address, so no other sender's reputation touches yours. Daily limits are preset per mailbox.
- InboxBand · dedicated
- isolated IP
- limits preset
- 5 · Provider matching
Google writes to Google, Microsoft to Microsoft
Every recipient's mail provider is detected first. Senders are matched to it, which is where placement is highest.
- Google sendersGmail and Google Workspace
- Azure sendersOutlook, Proofpoint, Mimecast
- 6 · Warm-up and tests
Four warm-up networks, daily seed tests
Three premium networks plus one of our own, rotated so no single pool gets biased. Inbox placement tests on seed pools are set up before the first cold email.
- premium 1
- premium 2
- premium 3
- our own
Infrastructure · part 2, governance
Checked every night.
Fixed before you notice.
The second job is watching. Every night the Deliverability agent checks every mailbox, decides what to change, acts, and tells you what it did and why.
- PlacementSeed tests on Google and Microsoftreal inbox or spam, per domain
- Warm-up healthGoogle and Microsoft, separatelythe early signal, about two weeks ahead
- BouncesLast 400 leads per domaina bad list burns a domain faster than volume
- DNS and blocklistsSPF, DKIM, DMARC, listingsdrift caught overnight
- VolumeWhat each mailbox really sentagainst what it was allowed
- CapacityHeadroom for the next 10 daysand a mailbox order when it runs short
- Deliverability · 07:10
One domain dropped to 74% at Microsoft overnight, and warm-up at Microsoft has been sliding for 9 days. I cut the whole domain to half for three weeks and moved its volume to the rotation pool. Your sending is unchanged. Nothing for you to do.
- Deliverability · 07:11
Capacity: you're at 89%. Rolling the winner out to all partners needs about 60 more mailboxes. Warm-up takes 3 weeks, so order by Oct 24 to be live Nov 14.
Prepare the orderLater
- 35,000+ mailboxes run
- 450,000+ placement tests
- 98.9% average placement
- 21 days of warm-up, never shortened
Not sure your emails reach the inbox?
We test your domains on Google and Microsoft and tell you what we see.
Chapter 8 · Limit governor · how much each mailbox may send
Nobody controlled the shape
of sending. So we built a governor.
Before, limits were set once, by hand, at setup. After that, a mailbox's day was shaped by the client's lead flow, and nothing watched the jumps.
Six limits compete every night.
The strictest one wins.
Three sensors feed six limits. Each limit can only lower the number, never raise it, so a mailbox grows only when everything is fine at once. Move the sensors and watch which rule takes over.
- Spam-test ceilingsensors 1 + 212
- Bounce cutsensor 3off
- Ramp+1 per earned day10
- Return capafter a long pauseoff
- Spam quarantine2+ hits: domain halvedoff
- Platform capGoogle and Microsoft15
Ramp holds it: one step up per earned day.
Read together: 0 spam hits with warm-up at 98–99.5% gives a ceiling of 12. One hit with perfect warm-up is noise; the same hit with sagging warm-up is a real warning.
Bounces under 3%: no cut. Ramp earned a step today.
Climb, one step a day
Sensors clean: the mailbox earns one more email per day it actually used, up to the platform cap.
Volume down, warm-up up
Cold volume steps down, warm-up share rises, the fault gets fixed. The mailbox never goes silent.
The whole domain rests
Three weeks at reduced volume for every mailbox on the domain. A pre-warmed replacement covers the gap.
Cut in one day. Earned back one step at a time.
Growth counts from what the sending tool confirmed yesterday, not from the level the mailbox once had. Domains that kept their volume after a spam event were still broken a month later 29% of the time; domains that were cut recovered.
- spam event · whole domain cut
- idle day · no step earned
The research · production data, April to August 2026
What 450,000 deliverability
tests taught us.
Before writing a single rule into the governor, we analysed 450,000 inbox placement tests across our whole fleet, mailbox by mailbox. As far as we know, it is the largest cold email deliverability study anyone has run. Seven findings.
The biggest spam trigger: sudden jumps in volume.
Every extra email costs a little inbox rate, but a jump costs far more. The damage arrives one to two weeks later, so this week's dashboard shows last week's decisions. The level you send at matters about four times more than how fast you got there, and the comeback after a pause is where most teams break their domains.
- Ramp at most 2× per week
- Never jump overnight
- Come back from a pause in two steps: 50% → 100%
Spam rate after a pause, by how you come back
- at old volume or below≈0%
- at ~2× to "catch up"6–7%
- full speed right after spam27%
Most teams come back at about 2×: campaigns restart with a bigger list than they stopped with.
- 1 · Spam is domain-level
35%vs 1.45%
When one mailbox starts hitting spam, its neighbours on the same domain follow within two weeks with 35% probability. On other domains: 1.45%.
Treat every spam hit as a domain incident. Scale with new domains, never subdomains.
- 2 · Warm-up decides a mailbox's life
19.7%vs 0.8%
Spam rate over the next 4 weeks after 14 days of warm-up, against 21 days. Mailboxes that start too early hit a ~91.5% inbox ceiling for good; after day 30: 99.5%.
21 days is the minimum, 30+ wins. For Google; Azure doesn't care about age.
- 3 · Failure shows two weeks early
98% → 92%
A warm-up drop that still looks green in most tools almost guarantees inbox placement falls to about 85% two weeks later. Google warm-up says nothing about Microsoft.
One sensor per ecosystem. Cut volume at the drop, before the crash.
- 4 · There is no safe volume
Inbox rate by emails per mailbox per day: 97.7% at 5, 94.6% at 10, 91% at 15.
Inbox rate falls in a straight line with every email you add per mailbox. There is no magic threshold; you pay for every extra email.
Volume is a business decision: pick the inbox rate you're buying, then set the cap.
- 5 · Your dashboard is a week late
1–2 weeks
Damage from a volume increase lands one to two weeks after it. The level you sit at matters about 4× more than the speed you got there.
If this week looks fine, you're looking at last week.
- 6 · Recovery: cut to half, wait 3 weeks
- cut to 50–75%healed
- cut deeperno faster
- no cut29% broken
A domain-wide cut heals in about three weeks. Cutting deeper doesn't heal faster. Domains that didn't cut: 29% still broken a month later.
Moderate cut, whole domain, then patience.
Every rule in the governor (chapter 8) comes from one of these findings.
Where is your outbound on all of this?
Eight questions, a score right away, and a free strategy in 24 hours.
How it's built · the architecture
The whole system
on one screen.
No magic and no single clever prompt. A chain of simple, boring parts, each doing one job, each one replaceable. Here is how data moves through it.
calls · mailboxes · CRM · campaigns · chat · 40+ data sourcesscheduled, every 30 min to nightly
once a night, only new files
checks: every claim has a source · formats · links · secrets
gateway: scoped key per person and agent · read freely · write only by proposal · every request logged
- Outbound Lead
- Researcher
- Copywriter
- Deliverability
- Launch Ops
- Analyst
↺ replies, results and calls go back to the collectors, and the loop starts again
Eight decisions that made it work.
Each one came from something that broke. If you build your own, these are the parts worth copying.
- 01
Collect without AI. Think once a night.
If the AI fails, nothing is lost: the sources are already saved and get compiled tomorrow.
- 02
Plain text in git, not a database.
Readable by a person, an agent or the next vendor. Every change can be seen and rolled back.
- 03
Split by who may write.
Robots write sources, AI writes pages, humans write notes. Fact and guess never mix.
- 04
Access means separate repositories.
A key for one zone can't reach another. Connect with a key, disconnect by revoking it.
- 05
Humans only where mistakes cost money.
Prices, promises, launches and spend need a click. Everything else runs.
- 06
Cheapest source first, stop at enough.
Pay for the result, not for five vendors agreeing with each other.
- 07
Test before you scale.
Every campaign is a cell in a matrix with a rule set before launch. Winners get the volume.
- 08
The domain is the unit of reputation.
Six limits, the strictest wins, and every spam hit is treated as a domain incident.
If you build it yourself · the honest checklist
You can build this.
Here is what it takes.
No single piece is exotic. The cost is that there are a dozen of them, they only pay off together, and each one fails quietly.
| Piece | What it needs | The hard part |
|---|---|---|
| Second Brain | Collectors for calls, mail, CRM, campaigns; a nightly synthesis; source checks; access zones | Keeping it honest: no number without a source |
| Data research | Contracts with many sources, a filter map per source, website reading with blocks handled, LLM qualification | Coverage without paying five times for one contact |
| Email proof | Several verifiers, catch-all validation, live LinkedIn checks | A bounce costs months of domain reputation |
| Copy system | Brief from the brain, dossiers, claims check, swap test, render check | Personal at scale without templates in disguise |
| Testing loop | A test matrix, a winning rule set before launch, an analyst writing learnings back | Changing one variable at a time |
| Infrastructure | Lookalike domains, premium Google and Azure mailboxes, 4 warm-up networks, daily seed tests | Seeing a failure two weeks before it happens |
| Governor | Nightly limits per mailbox from placement, warm-up and bounces | Rules you can only learn from hundreds of thousands of tests |
Playbook · best practices you can use without us
The rules we run on.
Take them.
A part of what sits under the hood, written as rules. Most of them you may already know. The difference is doing all of them, every day.
- .comDomains
Only .com, close to your brand, never your main domain. Check blocklists before you buy.
- ↪ captchaRedirects
Every sending domain redirects to your site through a captcha, so scanners see a real site and nothing to flag.
- 3Mailboxes per domain
No more than three. Reputation lives on the domain, so fewer mailboxes means a smaller blast radius.
- GooglePremium Workspace
Premium Google Workspace mailboxes, DNS on Cloudflare. Never cheap resold inboxes.
- AzureMicrosoft, enterprise IPs
Microsoft Azure SMTP on enterprise-level IP addresses, for recipients behind Proofpoint, Mimecast and Microsoft.
- 15 / 5Max per mailbox per day
15 cold emails a day on Google, 5 on Azure. And remember: every extra email costs a little inbox rate.
- 21+ daysWarm-up
Three weeks minimum before the first cold email, 30+ days for Google. Never shortened.
- 4 poolsRotate warm-up
Several warm-up networks, rotated. Never only the pool built into your sending tool, never one pool forever.
- weeklyPlacement tests
Seed tests at least once a week, separately for Google and Microsoft recipients. We run them daily.
- G → GProvider matching
Google senders write to Google recipients, Azure senders to Microsoft and secure gateways.
- 2× / weekRamp
Never more than double the volume in a week. Come back from any pause in two steps: 50%, then 100%.
- 20%Spare capacity
Keep a pre-warmed rotation pool, so a resting domain never stops your sending.
How to make the call: signal → action
- Warm-up inbox rate slides from 99% to 95–98%
- Step volume down now. The test failure comes in about two weeks.
- 1 spam hit, warm-up still perfect
- Watch. On its own it is mostly noise.
- 1 spam hit and warm-up already sagging
- Cut. Together they are a real warning.
- 2+ spam hits on a domain
- Cut the whole domain to 50–75% for three weeks. Move volume to the spare pool.
- Bounces over 3% on the last 400 leads
- Fix the list. Over 5%: cut hard. Over 10%: stop cold sending, keep warm-up.
- Mailbox younger than 21 days
- Zero cold email, whatever the dashboard says.
- 1 ≠ enoughNever one source
Every database is blind somewhere. One source means one blind spot in your whole market.
- 40+Use dozens of sources
Databases, directories, review sites, job boards, tech detection. Then merge and de-duplicate.
- 20Filter variations per source
Each source filters differently. Map them separately instead of copying one filter everywhere.
- ~7 pagesRead the website
Category labels lie. Read home, services, case studies, pricing, team, careers. Several passes when a site blocks you.
- LLMQualify with a reason
An LLM decides what the company really does, with a verdict, a confidence score and a written reason you can check.
- wideSearch people wide
Title filters miss buyers. Search wide, then let an LLM decide which titles really buy.
- liveLinkedIn check
Before sending: still there, same title, same company. Databases keep people for months after they leave.
- 3 stepsEmail verification, always
Three independent checks plus catch-all validation. One rejection removes the address.
- 90 daysFresh signals
Hiring, posts, leadership changes, events. Dated and sourced, or they don't count.
- $Cheapest source first
Ask sources in price order and stop when you have enough. Pay for the result, not five invoices.
- ✕Exclude before sending
Customers, open deals, opt-outs and competitors come out of every list, every time.
- sampleRead the output
Before scaling, a person reads a sample of verdicts. If the reasons are wrong, the brief is wrong.
- 3+ stepsSequence length
At least two or three emails. Each follow-up restates the main idea in one line and adds something new.
- 90–120Words, at most
Short enough to read on a phone in one screen. If it needs more, it's two emails.
- 0 linksFirst email
No links, no images, no attachments in the first touch. Plain text, like a person writes.
- 1 askOne idea, one question
One reason to reply, easy to answer. Not a calendar link in the first email.
- factsPersonalize with facts
Something true about this person this month, from the dossier. Not "{company} is growing fast".
- swapSwap test
Send Rachel's email to Daniel on paper. If it still works, it's a template. Rewrite it.
- proofNo claim without proof
Every number and promise in the email has a source on file.
- {a|b}Vary the wording
Variations across mailboxes, so a hundred senders don't send a hundred identical emails.
- renderCheck the inbox view
Preview on real rows: shortest and longest names, empty fields, Gmail, Outlook, phone.
- ✍︎Signature on every step
A real name and a real signature, on every email of the sequence.
- 1 / 250Decide by a rule
Set the winning rule before launch. At least one positive per 250 emails earns the next batch.
- ↺Write it back
What got replies and what didn't goes into your learnings. The next brief starts from it.
Want this playbook running for you?
We build it with you in 90 days, then run it together.
Free · your outbound strategy in 24 hours
Answer 8 questions.
Get your outbound strategy.
Your market, your buyers, the signals to watch and the emails to send, built for your type of company on top of what we've learned from 350+ teams. The same agents from this page build it, we check it, it lands in your inbox within 24 hours.
What do you sell?
Sets the playbook: product companies and service businesses buy differently.
Questions · what people ask us
Agentic outbound,
in plain words.
What is agentic outbound?
How is it different from an AI SDR?
Do I still need SDRs?
What does the system run on?
Can I build it myself?
How long until the first emails go out?
Next step · pick your starting point
Switch your outbound to
<agentic mode/>
- Flagship · 90 days
Agentic Outbound
The whole system on this page: Second Brain, six agents in your Slack, infrastructure underneath. Built with you, then run together.
growth.band/agentic-outbound → - Infrastructure only
Email Infrastructure
Your process stays yours. We build, warm, connect and check the sending layer every day.
growth.band/email-infra → - Who we are
Growth.Band
An agentic GTM agency. Ex-lemlist and ex-Expandi operators who run their own outbound on this system first.
growth.band →
Growth.Band · The Agentic Outbound Playbook · updated October 2026Slack examples use demo data · fleet figures come from our production system