A lot of AI-driven cold outbound is the same broken pattern as old-school spray-and-pray, dressed up in agentic clothing. It blasts thousands of templated emails, gets very few replies, and trains your future buyers to ignore your domain. Done right, AI outbound is a careful research-and-write engine that sends fewer, better emails and doesn't embarrass you. Here is the order that matters.
Step 0: Deliverability and compliance
Before a single send, set up the plumbing:
- Send from a separate domain or subdomain so your main domain's reputation is protected.
- Authenticate with SPF, DKIM and DMARC. Google's email sender guidelines require them for bulk senders and set spam-rate limits you should stay well under.
- Warm up new mailboxes gradually and keep daily volume per mailbox modest.
- Include a clear opt-out in every email. In the US, the CAN-SPAM Act requires opt-outs to be honoured within 10 business days; aim for the same day. In the EU and UK, check GDPR and the national e-privacy rules for B2B email, which differ by country.
Step 1: ICP, ruthlessly
Before sourcing a single lead, write a one-page ICP that is genuinely narrow: "VP of Engineering at Series B to D B2B SaaS companies in fintech with 50 to 250 engineers, headquartered in the US or Western Europe." Most ICPs are too broad to support quality outbound.
The narrower the ICP, the more research the AI can do per prospect, and the better the email gets. Write down disqualifiers too: industries you can't serve, companies too small to buy, and regions you can't support.
Step 2: Sourcing and enrichment
Apollo or Clay for sourcing, BuiltWith or a similar service for tech-stack enrichment, Crunchbase for funding signals, LinkedIn Sales Navigator for org-chart context. Suppress against your CRM, existing customers, open deals, anyone who has opted out and any do-not-contact list you hold. Verify addresses before sending, because bounces damage sender reputation quickly.
Step 3: Research, per prospect
This is where AI adds real value. For each prospect, an LLM-driven research agent reads:
- Recent LinkedIn posts from the last 30 days: what are they thinking about?
- Company blog posts and press from the last 90 days: what are they shipping or announcing?
- Tech-stack signals: are they using something that suggests a need for your offering?
- Funding or hiring patterns: are they scaling or restructuring?
The output is a five-line brief on what would be relevant to this specific prospect. That brief is the material the writer agent works from. Give the research agent a strict rule: if it can't find a real signal, it says so. A fabricated "I enjoyed your recent post" is worse than no personalisation at all.
Step 4: Writing
A separate writer agent drafts a first-touch email that references two specific signals from the brief. Tone matters: short, specific, no salesy language, one ask. A useful structure is one line on the signal, one on the problem it usually implies, one on how you help, and one low-friction question, in well under 100 words. A person reads every email before it is sent for the first month; after that, your error patterns are predictable and review can move to a sample.
Step 5: Sequence
Three to five touches over about three weeks, branching on prospect behaviour. Combine email with a LinkedIn touch, but keep the LinkedIn side manual: LinkedIn's user agreement prohibits automated messaging tools. Each follow-up should add something new, such as a relevant resource, a different angle or a short question, not "just bumping this to the top of your inbox".
Step 6: Reply triage
Replies route to your sales inbox with the prospect brief attached and an AI-suggested intent classification (interested, not now, wrong person, unsubscribe). A person owns the reply. Always.
What to measure
| Metric | What it tells you |
|---|---|
| Bounce rate | List quality |
| Spam complaints | Targeting and message fit, and deliverability risk |
| Positive reply rate | ICP and message fit together |
| Meetings held (not booked) | Real pipeline created |
| Opportunities opened | Whether meetings were with the right people |
Judge the programme on meetings held and pipeline, not opens. Privacy features in several mail clients make open rates unreliable.
Common mistakes
- Scaling before a person has read the first hundred emails.
- Personalisation that is obviously generated, such as praising a post the prospect never wrote.
- An ICP broad enough to fit half the market.
- Measuring opens and clicks instead of meetings.
FAQ
Is AI outbound allowed in Europe?
Rules for B2B cold email vary by country: some permit it under conditions, others are stricter, and GDPR applies to the personal data you collect either way. Document your legitimate-interest assessment, record the source of every contact, and take local advice for each country you target.
How many emails a day should we send?
Fewer than you think. Start low per mailbox, increase gradually, and watch bounces and complaints every day. If quality drops, cut volume before you change anything else.
How we run this at Velura Labs
Our AI Lead Generation service runs this playbook end to end, with a person reviewing output before anything scales. For the broader agentic architecture, see our agent framework guide. Talk to us if your outbound volume is high and your reply rate is not.
Available to businesses across the United States (Washington, California, Texas, New York), Europe (France, Italy and the wider EU), the Middle East (Dubai and the Gulf) and India. Get in touch to scope your build.