DEV Community

2pizza.team
2pizza.team

Posted on Originally published at 2pizza.team

AI Sales Automation: What Actually Works in Outbound

TL;DR: a full outbound system costs $4,000-10,000 to build and $500-1,500/month to run, against $50,000-80,000 a year for one SDR. It works for high-volume, low-touch B2B. It does not work for enterprise deals or small account counts, and it is regulated differently in the US, EU and Australia. All three matter more than the tooling.

Sales automation used to mean an email sequence. It now means something larger: software that builds the list, enriches it, writes the first touch, makes the qualification call, and hands a human only the conversations worth having. The whole top of the funnel, without a top-of-funnel team.

We build these. We also tell roughly a third of the people who ask that their business is the wrong shape for one, and the section on that is near the end. Read it before the cost section, not after.

Stage 1: building the list

Define the target profile precisely, which usually means industry, headcount band, geography, and some technology or behavioural signal. Tools like Clay scrape and enrich against that definition and return structured company records with a named contact, verified contact details, and the contextual facts you will personalise on.

What used to take an SDR a week takes a few hours. The trap is that the speed tempts you to widen the profile, and a wide list is how a domain gets burned. A tight list of 300 genuinely matched companies outperforms 3,000 loosely matched ones on every metric that matters, including how long your sending domain survives.

The unglamorous half: verification

Enrichment returns addresses that look right and are not. Bounce rate is the single biggest input to whether your mail lands at all, so every list goes through verification before it goes anywhere, and anything unverified gets dropped rather than tried. This is boring, it is where the system earns its money, and skipping it is the most common reason a pipeline underperforms.

Stage 2: personalisation that is not a mail merge

Generic cold email gets very low single-digit response and deserves to. Genuine personalisation does better, and the honest range is wide because it depends almost entirely on how well-matched the list is: we see low single digits on a mediocre list and into the low teens on a tight one in a market with a real, urgent problem. Anyone quoting you a fixed number without asking about your list is quoting a brochure.

Mechanically: the enrichment data feeds a model prompt that writes an opening line grounded in something specific and true about that company. Make or n8n handles sending. The quality bar is simple and strict, and it is the one thing we do not compromise on: if the opening line could be sent to any other company on the list, it is not personalisation and it will not perform.

What separates openings that work from openings that do not:

  • Reference something that would take a human two minutes to find, not something scraped from the homepage banner

  • Be specific about why you are writing to them rather than to their competitor

  • Say the thing in the first two lines, because that is all a phone preview shows

  • Write as a person, from a person, with a real signature and a real reply address

  • Never claim a relationship or a referral that does not exist

Stage 3: AI voice qualification

Voice agents make the first qualification call: they ask the questions, handle the obvious objections, score the lead, and book time with a human only when the answers justify it. A well-configured agent works through a hundred leads in the time a human rep gets through five, and it does the hundredth call exactly as well as the first.

What that call needs to establish:

  • Is this the person who can decide, or who can get it decided

  • Does the company actually have the problem you solve

  • Rough timeline and rough budget band

  • Whether they want a real conversation

  • How and when to follow up

Two hard rules we apply. The agent identifies itself as an automated assistant when asked and never pretends otherwise, and any request to stop is honoured immediately and permanently. Both are the right thing to do, and in several jurisdictions the second is also the law.

Stage 4: the handover, which is where most systems fail

Everything above produces warm leads. If the handover to a human is loose, the system is a lead-destruction machine with excellent metrics. What the human receives has to include the full transcript, the score and why it was scored that way, and the specific thing the prospect said they cared about.

And the response has to be fast. A lead that was warm on a Tuesday call is cold by Friday. If your team cannot respond inside a day, fix that before you build the machine that generates more of them, because otherwise you are paying to manufacture disappointment at scale.

Compliance, which is not optional and differs by market

This is the part vendors skip and it is the part that can cost you more than the build. None of this is legal advice, and if you are operating at scale you should get some.

United States

CAN-SPAM permits cold B2B email with conditions: accurate headers and sender identity, no deceptive subject lines, a physical postal address, and a working opt-out that you honour promptly. Call regulation is separate and stricter, and automated calling has its own rules.

European Union and UK

Materially harder. GDPR requires a lawful basis, and legitimate interest for B2B outreach is arguable but must be documented and balanced, not assumed. Several member states are stricter than the baseline. Treat EU outbound as a thing you do deliberately with advice, not a checkbox in a tool.

Australia

The Spam Act requires consent, which for B2B can be inferred in narrow circumstances such as a business address published without a statement refusing commercial email. Every message needs accurate sender identification and a functional unsubscribe. The inferred-consent path is narrower than most outbound tooling assumes, so this is a market where the list definition does the compliance work.

Practical consequence: the system needs a suppression list that is global, permanent and respected by every channel including the voice agent. Build that on day one. Retrofitting it after someone complains is both expensive and a bad look.

What it costs and what it replaces

A full system covering enrichment, sequencing, AI-written personalisation and voice qualification runs $4,000-10,000 to build and $500-1,500/month to operate. The running cost is mostly data and voice minutes, both of which scale with volume, so model your bill at the volume you intend to reach rather than the volume you start at.

Against that, one SDR costs $50,000-80,000 a year before the 25-35% on top for taxes and benefits, plus recruitment and three months of ramp. A working system covers the output of two to three of them on the mechanical parts of the job.

B2B Pallets is the version of this we shipped. An Australian pallet and packaging supplier with a good product, one salesperson, and no system: manual outreach, and hours a week burned on qualification calls that went nowhere. We built list building and enrichment, personalised sequencing, and a voice agent taking the first call and scoring it. Outreach volume rose roughly fivefold and coverage went from a single city to nationwide, with no new sales hires.

When not to build this

The honest list, and it disqualifies more businesses than it qualifies.

  • Your total addressable market is a few hundred companies. Automating outreach to a list you could work by hand wastes the list, and you only get one first impression per account.

  • Your deals are large, long and relationship-led. Nobody signs a six-figure contract off an AI qualification call, and trying can cost you the account.

  • You have no idea why your existing customers bought. Automation multiplies your message, so if the message is wrong you will find out faster and louder.

  • Nobody can follow up within a day. Build the follow-up capacity first.

  • You are in a market where your buyers already receive fifty of these a week. Being the fifty-first is not a strategy.

How to know if it is working

Measure the pipeline, not the activity. Volume sent is not a result. The numbers worth a dashboard are deliverability and bounce rate first, because everything downstream is meaningless if mail is not landing; then reply rate split into positive and negative, because a rising reply rate made of annoyed replies is a warning; then qualified conversations booked, and finally closed revenue per thousand contacts. That last one is the only number that decides whether the system pays for itself.

Want to know whether your market is the right shape for this before spending anything? The audit at 2pizza.team/audit takes two minutes, no call. If the answer is that your list is too small to automate, it will tell you that.


Originally published at 2pizza.team. We build AI and automation systems for small teams - fixed price, two to six weeks. See the work.

Top comments (0)