How to build an AI outbound system for a B2B agency

The typical starting point is an agency with solid delivery, referral-dependent growth, and no acquisition system at all. The constraint is rarely delivery. It is that nobody outside the founder's network is hearing from them. The fix is the same shape every time: a list built from filters instead of guesses, sent from infrastructure that will not get burned, with AI doing the qualification and the first line, and a CRM that catches every reply. What differs is the tools, and in 2026 there are three or four credible options for every job. This guide gives you the framework, the shortlist per job at the vendors' September 2026 list prices, the deliverability rules, the sizing math, and the pitfalls that end most first campaigns in week three.

What does an AI outbound system for an agency actually consist of?

An AI outbound system is five connected jobs: a written ICP that produces a verified lead list, warmed sending infrastructure on secondary domains, a sending platform, an AI layer that qualifies leads and writes the first line, and a CRM that every reply lands in with a next step attached.

The word system matters. A cold email tool with a list uploaded into it is a campaign, and campaigns end. What you want is a machine where a new batch of leads enters at the top every week and produces replies a person only has to answer. The setup that holds up at volume: Apify scrapers and intent signals for the list, a finder waterfall plus verification, AirMail or Mailforge inboxes, Instantly for email, HeyReach and Kondo for LinkedIn, a model inside Make or n8n, and whichever CRM your team will use. Each job has a credible alternative, priced below.

How do you define the ICP and where do the leads come from?

You define the ICP as a set of filters a data tool can apply, not as a persona paragraph, and only then choose the source, because the source follows the filters.

"B2B software companies that need content" is a persona. "Founder or head of marketing at a B2B software company with 20 to 200 employees in the UK, DACH or US, whose site has a demo-request page and a blog silent for 90 days" is a filter set. Every clause maps to a field a database or scraper can check, and a model checks the last two against the website. Write the disqualifiers with the same precision: consumer apps, self-serve tools with no sales team. Someone who is not a fit does not reply politely. They hit the spam button, and every mailbox on the domain pays for it. Then write the offer as a specific problem and result. "We write content for SaaS" gets no replies. "Four comparison pages a month that rank for your competitors' names and turn into demo requests within 90 days" gives a reader something to say yes or no to.

The list is four passes: source, enrich, find the email, verify. The default runs each as its own workflow step; Clay does all four in one table.

Job	Default	Alternative, and when to pick it
Source the raw list	Apify scrapers (Starter $19 a month; the Apollo scraper runs $1.50 per 1,000 leads) plus intent signals: hiring, LinkedIn engagement, job changes	LinkedIn Sales Navigator (Core $119.99 a month) when titles and posts matter more than emails; Apollo (free tier, paid from about $49 a user) for a plain firmographic filter
Find the work email	A waterfall: several finders in sequence, each run only on the rows the last one missed, such as AnyMailFinder (EUR 69 for 2,000 credits, charged only on a verified hit) and FullEnrich (Pro EUR 55 for 1,000 credits, 25+ sources, pay only on a hit)	Clay (Launch $167 a month, $150 billed annually) with its own work email waterfall when you want sourcing, AI columns and the waterfall in one tool
Verify before sending	MillionVerifier (50,000 credits $89, never expire) as the last step, on the whole list, within days of sending	ZeroBounce (from $39 for 2,000, 100 free a month) for small lists

Pull 2 to 3 times the leads you want to end with, because a third to a half of any raw list falls out at qualification or verification. Hold the list to more than 60 percent of rows with a verified email after all four passes. Below 50 percent, the source or the filters are wrong, and more volume on a bad source only makes the bounce problem bigger.

Intent signals are the part of the default most self-builds skip. A company that posted a job for a marketing manager or a sales development rep has just approved budget for the problem you solve. Apify's Hiring Intent Lead Scraper turns Google Jobs postings into hiring companies with a verified contact from $50 per 1,000 leads. People who engage with LinkedIn posts about your problem, and buyers who just changed jobs, are the other two signals worth a scraper. Give signal-led leads their own campaign with a first line that names the signal, and send those first.

Which deliverability rules are non-negotiable in 2026?

The non-negotiable rules are authentication on every sending domain (SPF, DKIM and DMARC), a spam complaint rate under 0.10 percent, secondary domains instead of your main one, and a warmup period of at least 14 days before a mailbox sends a campaign email.

These are the mailbox providers' rules. Google's sender guidelines require SPF and DKIM for the domain, DMARC for the sending domain, valid forward and reverse DNS, and TLS in transit. On complaints, Google says to keep the spam rate in Postmaster Tools below 0.10 percent and to avoid ever reaching 0.30 percent. Senders of 5,000 or more messages a day to Gmail count as bulk senders and also need one-click unsubscribe. Yahoo's requirements mirror them: SPF or DKIM at minimum for everyone, both plus DMARC at p=none or stricter for bulk senders, spam rate below 0.3 percent. Microsoft joined in 2025: since May 5, senders of more than 5,000 messages a day to Outlook.com, Hotmail and Live addresses need SPF, DKIM and an aligned DMARC policy of at least p=none, and non-compliant mail is rejected with a 550 5.7.515 error rather than junked.

An agency setup sending 300 emails a day is nowhere near bulk; the authentication and complaint thresholds apply anyway, because the receiving server scores every message the same way. Set the records in this order: SPF and DKIM first, then wait. Google's DMARC guide says to allow 48 hours after SPF and DKIM before adding DMARC, to start at p=none, and to tighten later once you can see how receivers treat your mail. Skipping the wait is the most common DNS mistake in a failed setup, and the symptom is a domain that was in spam from its first day.

Then the rule founders resist: cold email never goes out from the domain your clients email you on. Instantly's own strategy guide says it plainly: buy secondary domains so you do not damage the reputation of your main domain, and put a maximum of 3 to 5 mailboxes on each. Register close variants of your brand, redirect them to your real site, and accept that one will eventually get burned.

Which sending infrastructure, and how many mailboxes?

Instantly AirMail, or an outbound-specialized inbox provider such as Mailforge once you run many mailboxes, is the default; Google Workspace and Microsoft 365 are the alternatives; and the mailbox count comes from a formula, not a feeling.

Instantly AirMail is $4 per inbox a month, hosted on Instantly's own infrastructure, with SPF, DKIM and DMARC configured automatically. Pick it when you send from Instantly and want the inboxes in the same account. At volume, Mailforge sells mailboxes from $3 a month billed yearly, in blocks of 10, with domains at $14 a year and the DNS set for you. Pick it when the mailbox count drives the bill or you send from another platform. Google Workspace Business Starter is EUR 6.80 per user a month billed monthly, and Microsoft 365 works the same way. Pick those when you want mailboxes on the provider most of your list uses, a hypothesis to test in your placement data, not a rule. Instantly's pre-warmed accounts skip the wait at $10 a month each, but cap you at 20 to 25 sends a day on its platform.

Instantly's warmup guide says to wait at least 2 weeks before using an account in a campaign, 3 weeks for its AirMail accounts, to hold new accounts at 10 warmup emails a day, never to disable slow warmup on a new account, and to launch only with a health score above 90 percent. Smartlead and lemlist bundle warmup too. If you want inbox placement tests on top, MailReach charges $19.50 per mailbox a month and includes 20 spam test credits.

Now the sizing formula, and this is an illustrative calculation, not anyone's client data. Instantly's account limits recommend 30 campaign emails per mailbox per day; Google Workspace allows 2,000 a day per paid user, but that is the ceiling before Google cuts you off, not the volume a cold mailbox survives. Over roughly 22 sending days a mailbox at 30 a day sends 660 emails. Divide your monthly sends by 660 and you have the mailbox count; divide by 3 and you have the domains. A 1,500-lead list in a 4-step sequence, with early replies and bounces dropping out, is about 6,000 sends in month one: 9 mailboxes, 3 domains. A 5,000-lead list is about 30 mailboxes on 10 domains, lead gen agency volume and more than most agencies need. Double the mailboxes and you double cost and complaint surface; move the reply rate from 3 to 5 percent and you get 67 percent more meetings from the same infrastructure. Size once, then spend your attention on the list and the copy.

Which sending platform should you pick?

Instantly is the default when email is the channel and volume is the game, Smartlead is the alternative that runs many mailboxes for less, and lemlist fits when LinkedIn and calls belong in the same sequence. None of them is the system of record for replies; that is the CRM's job.

Platform	List price (vendor page, September 2026)	What it is good at	Pick it when
Instantly	Growth $47 a month (1,000 contacts, 5,000 emails); Scale $194 (25,000 contacts, 100,000 emails)	Unlimited mailbox rotation, bundled warmup, the most documentation, AirMail inboxes in the same account	The default: email-first, and you want the most documented path
Smartlead	Base $39 (2,000 active leads, 6,000 sends); Pro $94 (30,000 leads, 90,000 sends); unlimited mailboxes on every plan	Cheapest way to run many mailboxes, strong API, agency sub-accounts	Email-first at higher mailbox counts, or you also run outbound for clients
lemlist	Email $69 a month (50,000 emails); Multichannel $109 with LinkedIn, calls and unlimited emails, 5 senders per user	Email and LinkedIn steps in one sequence, built-in warmup, its own lead database	You want one tool for email plus LinkedIn and fewer than a dozen senders

Instantly's 2026 benchmark, built on platform data from January to mid-December 2025, is the reference for what any of these produces: an average reply rate of 3.43 percent, the top quartile above 5.5 percent, the top 10 percent above 10.7 percent, 58 percent of all replies on the first email, 4 to 7 steps as the productive sequence length, and first emails under 80 words doing best. Run 1,500 leads at a 4 percent reply rate and you get 60 replies; a quarter positive is 15 yeses; half of those held is 7 or 8 meetings from the first list. The platform does not move those numbers much. The list and the copy do.

Where does the AI go, and how do you test the copy?

The AI goes in three places: qualifying every row against the written ICP, writing one true first line per lead, and triaging replies. The copy is tested with two or three variants that change one thing, run for two weeks against equal slices of the same list.

Qualification pays for itself fastest. A Claude or OpenAI call reads the company website, the description and the headcount and returns pass or fail against the ICP with a one-line reason. In the default setup it is a step in a Make or n8n scenario, or a custom agentic workflow when the call needs judgement across several pages. In Clay it is an AI column, and most models there cost a flat number of data credits per row. The rows that fail never get an email, which protects your complaint rate more than any DNS record.

Personalization is where founders overdo it. The job of the first line is to prove a human looked, in one sentence, using something true and recent: a job post, a service page, a client they just announced. Generate that line from the enriched fields and let the rest be the same tested copy for everyone. AI-written whole emails read like AI-written whole emails. One specific line and a plain offer beats three paragraphs of generated flattery, and keeps the email under the 80 words the benchmark favors.

Write two or three variants that differ in one thing, usually the hook: problem-led ("your blog gets traffic and no demo requests"), result-led ("comparison pages that turn into demo requests within 90 days"), insight-led (something the reader does not know about their own market). Same offer, same call to action, same length, split at random. Keep the register peer-to-peer: a first line that admits it is a cold email, one proof point, one soft question ("worth a look?") rather than a meeting request. Plain text, no images, no tracking pixels on the first send. With 500 leads per variant and 58 percent of replies on the first email, the winner is visible by day 10. Kill the losers, move their leads to the winner, then add the next batch.

How do you add LinkedIn without burning the account?

You add LinkedIn as a second touch on the same list, a connection request with no pitch a day or two after the first email, at low daily volume, with every reply answered by a person, and you run it through a cloud-based tool rather than a browser extension.

Multichannel is the current standard for a simple reason: the second channel moves a name from unfamiliar to seen-that-before, and a founder who has seen your name replies to the email more often. Stay well under LinkedIn's weekly invitation cap and ramp new accounts slowly. Send the request without a note or with a one-line human note, never the pitch. Personalize the first message from the same enriched fields the email used. Never automate replies. Run it from a cloud tool on a dedicated IP rather than a browser extension, because extensions are what LinkedIn detects.

The setup that holds up once several people send is HeyReach for sending, $79 a month on Growth for unlimited sender accounts, and Kondo for the inbox. Kondo calls itself Superhuman for DMs. Basic is $28 a user a month billed annually, and Business at $36 adds CRM integrations and the Sales Navigator inbox. lemlist is the alternative when you want LinkedIn steps inside the email sequence, at $109 a month for Multichannel, and La Growth Machine starts at EUR 60 per identity a month. Whatever you pick, LinkedIn replies land on the same CRM record as the email replies.

What does the stack cost per month?

At September 2026 list prices a lean email-only stack costs about 80 to 100 a month, a standard 9-mailbox build 300 to 450, and a multichannel build with LinkedIn 400 to 700, mixing the vendors' dollar and euro prices one to one.

Stack	What is in it	Illustrative monthly cost
Lean: one founder, 3 mailboxes, email only	3 AirMail inboxes $12 with Instantly Growth $47, or 3 Workspace mailboxes (EUR 20) with Smartlead Base $39; Apify Starter $19 for the list; ZeroBounce $39 pack once; HubSpot free CRM or Close Solo $19	About 80 to 100, plus the one-off verification pack
Standard: 9 mailboxes, a finder waterfall	9 AirMail inboxes $36 with Instantly Scale $194, or 10 Mailforge mailboxes from $30 with Smartlead Pro $94; Apify $19; AnyMailFinder EUR 69 and FullEnrich EUR 55, or Clay Launch $167; MillionVerifier $89 once; Attio Plus EUR 39	Roughly 300 (Mailforge plus Smartlead) to 450 (AirMail, Instantly and Clay)
Multichannel: email plus LinkedIn, 2 senders	Standard stack plus HeyReach Growth $79 and Kondo Basic $28 a user, or lemlist Multichannel $109 in place of both; Sales Navigator $119.99 if you use it	Roughly 400 to 700

Each price is from the vendor's pricing page linked above, and several show euros to European visitors and dollars to everyone else, so check before you buy. Enrichment is the line people cut, and it is the wrong one: the finders and the verifier keep bounces down. Next to 30 days of founder time and what one closed client pays back, 700 a month is a rounding error. The build-or-buy question below is about attention, not budget.

Where do replies go once they arrive?

Every reply lands in the CRM as a person attached to a company, with the campaign, the variant, the reply text and a next action, and nothing stays in the sending tool's inbox.

This is the layer that separates a system from a campaign. The unified inboxes in Instantly, Smartlead and lemlist are fine for reading and a terrible place to run a pipeline: no history, no deal value, no reminders, and in 3 months you will not remember which of the 60 replies said "ask me in Q4". So the reply is pushed out into the CRM as a person, a company, a deal in the stage "replied", and a task for whoever answers. The platform's native CRM integration does it, or a webhook into a Make scenario (Core $9 a month) or an n8n workflow (EUR 20 a month billed annually, free to self-host). Zapier, from $19.99 a month billed annually, is the alternative if your team already runs on it.

Pick the CRM your team will actually use; the five that come up most all hold this pipeline. Close is Solo $19, Essentials $49 and Growth $109 a user a month billed monthly, the pick when calling is daily work. HubSpot is free for 2 users and Sales Hub Starter is $20 a seat, the place to start from nothing. Pipedrive, from EUR 14 a user billed annually, is the pipeline-first classic with nothing new to learn. Attio is EUR 39 a user on Plus and free for 3 seats, the pick when you want to shape the data model around outbound. GoHighLevel is Starter $97 or Unlimited $297 a month with unlimited users, the pick when a bigger team wants one flat bill. Whichever you choose, build 6 stages (replied, positive, meeting booked, held, proposal, won or lost) and 2 automations: a task on every positive reply, and a follow-up date on every "not now" so the lead comes back in 90 days.

The CRM is also where reply triage runs. A model reads each reply and tags it, usually in a custom agentic workflow, because the call needs judgement: positive gets the task, objection gets a drafted answer for a person to send, "not now" gets the follow-up date, bounce or unsubscribe removes the lead from every campaign and flags the domain if bounces are rising. What it never does is answer a positive reply on its own. That conversation is the whole point of the system, and it is a person's job.

What goes wrong in the first 30 days?

What goes wrong is almost always one of five things: sending from the main domain, launching before warmup finishes, a list that was never verified, copy that pitches instead of asks, and nobody reading the deliverability data every morning.

The shape of a failed first campaign is consistent. A founder buys 2 domains, warms them for 4 days because the launch date was already set, uploads a database export with no second verification, and sends a 120-word email that opens with "We help SaaS companies grow". Bounces first, spam complaints next, both domains in the junk folder by week 2, and the conclusion that cold email does not work.

The failure that hides best is the last one. Deliverability is a number you read every day, never a setting you configure once. Bounce rate, spam complaints in Google Postmaster Tools, warmup health per mailbox, reply rate per variant. Those four figures take 5 minutes a morning and tell you a day early when a domain is going bad. Pause the mailbox, let warmup rebuild it, move its share of the list. Few founders do this, which is why first campaigns die in week 3.

Should you build it yourself or have it installed?

Build it yourself if you can give it 30 days of your own attention and you will keep reading the numbers after launch; have it installed if that attention is worth more on delivery or on the sales calls the system produces.

Everything above is buildable by one founder with the vendor docs open. The tools are cheap, the DNS is an afternoon, and you can write the copy about your own business. What breaks the self-build is rarely skill. Week 2 collides with a client escalation, warmup gets abandoned at day 9, and the half-finished enrichment table sits there for 3 months. If you already know that is how it will go, be honest about it now instead of on day 40.

The two alternatives are renting it or having it installed. A lead gen agency rents you the infrastructure and the list, and keeps both when the retainer ends. The install route is what Ourix does. Ourix solves B2B acquisition for founder-led businesses, agencies first: it builds a done-for-you outbound system inside your business, on your own domains and accounts, with AI-assisted SDRs that book qualified meetings every week. Campaigns go live in about 30 days. The build is a one-time fee scoped to your channels and your market, with no lock-in, you own everything, and the exact number comes on the call. Once acquisition runs, Ourix consults founders on building their agency operating system through AgencyOS, so new clients come in at a better margin. What an agency operating system is shows where outbound sits in that larger build.

Either way, the test at day 30 is the same: qualified meetings on the calendar every week, nobody prospecting by hand, and a dashboard that shows which variant and which list they came from.

What to do next

1. Write the ICP as filters plus disqualifiers, and the one-line offer with a specific problem and result, before you buy anything.
2. Register 3 secondary domains with 3 mailboxes each. AirMail sets SPF, DKIM and DMARC for you; on Workspace, set SPF and DKIM, wait 48 hours, then add DMARC at p=none. Start warmup at 10 a day and launch 14 days later at the earliest.
3. While warmup runs, pull 3,000 to 4,000 raw leads from scrapers and intent signals that match your filters, qualify them with a model, run the finder waterfall, then verify the full list with a dedicated verifier. Stop at 1,000 to 2,000 verified rows.
4. Write two or three variants under 80 words, plain text, one generated first line, one soft question. Split the list at random.
5. Build the CRM pipeline with 6 stages and 2 automations, and wire replies out of the sending tool into it before the first email goes out.
6. Launch at 30 sends per mailbox per day. Read bounce rate, Postmaster complaints, warmup health and reply rate every morning. Cut the losing variants on day 10, add the next batch, and add the LinkedIn touch (HeyReach plus Kondo, or lemlist) once the email data is clean.

If you would rather see where outbound sits among your bottlenecks first, the free growth audit is a read of your acquisition setup and what to fix first, delivered as a document you keep. If acquisition is the constraint, that document is the brief for the build above.