“I think we’ve reached the limits of our address book.”
That’s how Lionel Hadjadjeba, CMO of Gosta Labs, described the problem. Gosta builds an AI operating system for specialty care, leads its home market in Finland and set out to win the US and key European markets. The product was ready. The contacts were running out.
Three oncology congresses later, Gosta’s campaigns had sent 3,930 messages, received 198 replies and produced 62 qualified opportunities.
AI did a lot of the work in those campaigns, but the results started earlier. Gosta knew exactly who buys and why, and the team was already active at industry events. AI makes lead generation faster at whatever you point it at. Point it at a vague target group and it sends more messages to the wrong people, faster.
The 5 workflows below are the ones we see producing meetings, in our own sales and in client projects.
First: a target list with names on it
Every workflow in this article depends on one thing: an ideal customer profile defined down to account names.
In many companies nobody has formally decided the target group. It forms around whichever accounts a salesperson calls when the calendar is empty. The sales report looks fine, and the problems show up a year later in delivery, when the customers turn out to be price-sensitive and expensive to serve.
With AI-assisted prospecting, a precise definition stops being optional. Some examples of what that looks like in practice:
- Our own list covers about 1,450 companies. Every one of them should hear from us at least once every four months.
- Evitec, a banking and insurance software provider, used AI to identify more than 1,000 relevant decision-makers across 100+ financial organizations in Norway and Sweden. See the Evitec case.
- Woodly, a Finnish maker of wood-based plastic alternatives, defined its UK buyers first: packaging buyers, sustainability leads and procurement teams at café chains and food brands. Only then did we build the contact lists.
Workflow 1: Act on signals
Signal-based selling gets a lot of attention right now. Most signals fall into three groups.
Change. A company hires a new marketing director, raises funding or expands to a new country. Change opens budgets.
Interest. Someone visits your website, downloads something or comments on a post about your industry.
An existing relationship. A former customer moves to a new company, or you met the person at an event.
Cold lists are still the base of everything, but they’re rarely the best place to start. Existing relationships are the cheapest and fastest signals, and they’re used surprisingly little. Few teams track when their former customers change jobs, even though those are some of the best leads available.
Signals are also not equal, and the system should treat them differently. This is how we sort them:
- Replied or met us: a person takes over, and the machine leaves the contact alone.
- Raised a hand by registering for or attending an event: a message sequence starts right away.
- Moved into a relevant role: a sequence about the new role.
- Light interest, such as a reaction to a post or a website visit: no message at all. The contact warms up and moves up the queue.
One like on a LinkedIn post is not a buying signal.
How strongly to react is a design decision for each company. For Singa, an ICP-fit company visiting the website triggers a tailored first email. Claude drafts it, and a person approves it in Slack with one click before it goes out through Instantly. The Singa case shows the whole setup.
Workflow 2: Event campaigns, one event at a time
Events give you a reason to get in touch and a short window to use it.
For Gosta Labs, we built a separate target list for each oncology congress instead of one big list. AI wrote a first line for every message that referred to the doctor’s own published research. The first test ran at the Community Oncology Conference in Orlando and produced warm leads within days.
Across Community Oncology, ESTRO and ASCO, the best campaigns reached a reply rate of almost 19 percent. At ESTRO, about one reply in three turned into an interested lead. The campaigns worked well enough that we tripled the number of sender domains for the next round. The full story is in the Gosta Labs case.
We use the same model for our own events. Before Nordic Business Forum, we sent 481 messages through the event app Brella and 360 on LinkedIn, which led to 31 booked meetings.
The follow-up matters as much as the build-up, and it starts with cleaning the list rather than writing the message:
- People who registered for our speaker session or came to our after-work event get a message.
- People who registered for the after-work event but didn’t come get nothing. A no-show isn’t a signal.
- People we already had a meeting with get a personal message from a person.
- Current customers get thanked face to face.
Then two different messages go out: one with a thank-you, the other with a thank-you and a suggestion to talk more. Our first messages went out during the event, and most of the rest on the Friday and Monday after it. A week later, people no longer remember the conversation.
Workflow 3: Segment personalization at volume
Generic outreach doesn’t work in a market where nobody knows your name.
For Woodly’s UK launch, we used Clay and Apollo to identify more than 10,000 potential contacts across the target segments, enriched each one and scored it by relevance. AI did the research and the first draft of the personalization for each segment. Woodly’s team reviewed and refined the messages so they matched the company’s voice. The campaigns ran in Instantly.
In five months, Woodly generated more than 120 qualified leads in the UK, and several conversations moved to technical testing with major UK café chains. The Woodly case has the details.
Singa’s outbound program works the same way at a larger scale. The first wave covered roughly 10,000 prospects across US and UK karaoke bars, family entertainment centers and bars with karaoke nights. Each segment got its own messaging angle, built from Singa’s sales templates and call recordings. In the first two weeks, the US program reached around 2,000 prospects and produced 15 warm leads.
Workflow 4: The always-on cycle
Campaigns end, and the market doesn’t wait for the next one. Our own system runs in four-month rounds. During each round, every contact on our list hears from us once, even if they never do anything. Signal-based campaigns run on top of that.
The channel order is fixed:
- A LinkedIn connection request to everyone, with no sales pitch after it.
- A phone call to warm contacts who haven’t replied to messages. The machine builds the call list, and a person makes the call.
- A LinkedIn message to warm contacts and to those we have no email address for.
- An email to cold contacts.
The rules for when nothing goes out are just as important. No message goes to someone who has already heard from us this round, who has said no, or who is a customer, a former customer or in the middle of a buying process. Nobody gets more than one contact from us in seven days.
Every send gets logged, including the phone calls a person makes by hand.
Workflow 5: Warm the list with content before you call
When we made our first cold calls, the day’s result was 37 calls, 21 people reached and 6 meetings booked. All six meetings came from people who had liked or commented on our LinkedIn posts. The rest of the list produced none.
Marketing had done the work before the phone rang. The salesperson only had to agree on a time.
For the calls, we built a simple tool that pulls the call tasks from our CRM and puts the warmest contacts first. Each card shows the person’s name, role, an opening line and the post they reacted to. The machine sorts the list, the content warms it, and a person makes the call.
Three questions before any message
Every message our system sends has to pass three questions, in this order:
- Do we know who this person is?
- Are we allowed to contact them?
- What was the signal?
Many systems only ask the third one. That’s how a prospecting message ends up in a current customer’s inbox.
The same discipline applies to data protection. Document why you’re allowed to contact each person and tell them where you got their details. Check the rules in each market before you send.
And remember, you need the best B2B data you can get. There’s lots of different vendors offering data. Make sure that vendor has the data you need and equally that they have a headless solution for you to access that data. UI is so last season.
Build or buy
Clay is a good tool. It collects data on companies and people from several sources, puts it in a table and fills in missing phone numbers and email addresses, all without a developer. Woodly’s lists were built with it.
Claude Code can now write the same chain as code. It uses the same data sources through their APIs, without row limits or credits per row, and the data can go straight into the CRM instead of a table you need to export from.
Clay is still the better choice for a team that has nobody to maintain its own system. Custom code has to be documented and fixed every time an API changes, and that’s easy to forget while everyone is excited about the build. When the data volume is high and the same chain runs every week, your own code gets cheaper quickly.
Our own lead generation runs on a GTM machine built on Anthropic’s managed agents, and our AI-powered marketing work for clients uses the same building blocks. Once the leads start replying, the next bottleneck is handling them, which is where AI sales agents come in.
Take last month’s outbound messages and check each one against the three questions. The ones that fail show you what to fix before you add a new tool.
