Pipeline in a market where nobody knows your name.
Entering the US market from Europe is a cold start in the most literal sense: no brand recognition, no referrals, no warm anything. Our client had the product and the pricing, and needed the part that cannot be improvised, a steady stream of qualified conversations with the right American companies. We built an agent that runs that stream end to end: sourcing from the Apollo B2B database, researching every prospect individually, sequencing compliant outreach across US time zones, and handing the sales team LinkedIn touchpoints ready to send. The client's team wakes up in Europe to a pipeline that worked the American afternoon.
Hermes runtime Apollo database LinkedIn touchpoints Sequencing engine Supabase
Selling into the US from Europe has a brutal asymmetry: the market is enormous, and it does not care. Manual prospecting meant a salesperson searching for companies, guessing at the right contact, writing an email at 4 PM Central European time that lands in a Chicago inbox at 9 AM, and getting through maybe fifteen genuine attempts a day. At that pace, a market of hundreds of thousands of potential accounts might as well be closed.
The second problem was the working day itself. US business hours barely overlap with European ones. Every reply that arrived overnight sat unanswered until the next European morning, which in US terms is the middle of the night. Momentum died in the timezone gap.
The client did not need more effort. They needed a system that treats US market entry as an engineering problem: define who to reach, find them at scale, contact them properly, and never let a reply go cold.
> define
The ideal customer profile is a written contract: industry, company size, roles, US regions, and explicit exclusions. The agent prospects inside those lines and nowhere else.
> source
The agent pulls matching prospects from the Apollo B2B database, a licensed source built for exactly this. No scraping, no gray-zone lists. Contacts are verified before they enter the pipeline; bounced domains and catch-all traps are filtered out.
> research
Every sourced prospect gets an individual profile: what the company does, recent signals, and where the client's offer plausibly fits. The research feeds the writing, nothing generic survives.
> sequence
Email sequences run on US time zones, a prospect in Denver is contacted on Denver's morning, not Belgrade's. Every message carries a working opt-out and the client's physical address, CAN-SPAM is a design input, not an afterthought.
> touchpoint
In parallel, the agent prepares LinkedIn touchpoints, personalized connection notes and follow-up messages, queued for the client's salespeople to review and send from their own profiles. The agent writes, humans press send on LinkedIn.
> handoff
A reply ends automation for that prospect instantly. The thread moves to the sales team with the full research profile and interaction history attached, so the human enters the conversation already informed.
Layer I · Visual Architecture
One diagram: ICP in, Apollo sourcing, verification, research, two parallel channels (email sequences and LinkedIn touchpoint queue), reply detection, human handoff. The handoff is drawn as the goal of the system, everything upstream exists to produce it.
Layer II · Contracts
The ICP is a contract, the messaging is a contract, and the compliance rules are a contract: what a message may claim, what it must contain, when sending stops. The client's sales leadership signed all three before the first prospect was sourced.
Layer III · Technical Diagrams
Sourcing filters, verification flow, timezone-aware scheduling, deliverability protection (warm-up, volume ramps, domain separation), reply detection, and the LinkedIn queue, specified before implementation. Deliverability got its own diagram, because a burned domain ends the project regardless of how good the copy is.
Layer IV · Implementation
Hermes runtime on a dedicated isolated instance. Apollo as the licensed prospect source, a sequencing engine with per-timezone scheduling, Supabase for prospect profiles, sequence state, and the full interaction log.
runtime Hermes sourcing Apollo B2B database, verified contacts only channels email sequences + LinkedIn touchpoint queue scheduling per-prospect US timezone alignment state Supabase (profiles, sequence state, interaction log) compliance CAN-SPAM by design, instant opt-out, reply kill-switch
Burned sender domain
Outreach runs on separate sending domains with gradual volume ramps and warm-up, the client's primary domain is never exposed. Deliverability metrics are monitored continuously, and volume drops automatically when they dip.
Stale or wrong data
Database records age. Every contact is verified before sending, and bounces feed back into the filter. A high bounce rate is a deliverability wound that takes months to heal, so it is prevented, not treated.
The timezone trap
A sequence scheduled in European time lands at night and reads as automation. Scheduling is anchored to each prospect's local business hours, down to the region.
Sequencing past a reply
Nothing destroys a warm lead like an automated follow-up arriving after they already answered. Reply detection halts the sequence within minutes, across both channels.
LinkedIn account risk
Automating LinkedIn actions risks the salespeople's own accounts, which are assets the client spent years building. That is why the agent prepares LinkedIn messages but never sends them, the human's account acts only through the human's hand.
~300
verified prospects sourced and researched per week
2
channels per prospect, sequenced in coordination
< 15 min
from inbound reply to sequence stop and human handoff
5.2%
positive reply rate on the primary sequence
0
spam complaints, 0 burned domains since launch
Figures from the first two months in production, measured across the primary ICP segment.
Market entry is a scheduling problem in disguise. The single biggest lift did not come from better copy, it came from sending at the prospect's 9 AM instead of ours. The same message, moved eight hours, stopped reading as European automation and started reading as a normal business email.
The channel split is the safety model. Email can be automated responsibly, LinkedIn cannot, so the system does not pretend otherwise. The agent does the research and the writing for both, and the boundary of what it sends follows the boundary of what each platform's rules and risks allow. Architecture should encode judgment, not just capability.
An agent that stops is worth more than an agent that persists. The reply kill-switch, the opt-out handling, the automatic volume drops, every mechanism in this system that halts something is the reason the numbers hold month after month. In outreach, restraint is not a limitation of the system. It is the system.
A market that does not know you yet is just a market you have not systematized yet.
> ../book_a_call.sh