Sorbet, a MENA cross-border payment platform, needed an outbound machine that could find, qualify, sequence, and reply-handle accounts at scale. I designed a 10-step system with three LLM touchpoints, all running off a single Clay table.
Sorbet is a B2B fintech moving cross-border payments on USDC stablecoin rails. ADGM-licensed, serving 3,000+ businesses across UAE, KSA, and Egypt. Pricing is sharp: $0/month, 1% flat per transaction. The value is structural: same-day settlement, MENA-native, no FX margin games.
MENA SMBs routing international flows through Wise, Payoneer, or SWIFT absorb 3–5% in FX fees and 3–5 day settlement delays. The market is large, fragmented, and underserved. The objective: design an outbound system that finds these accounts, qualifies them, and reaches the right person — automatically, with an LLM doing the judgment work at every decision point.
A single Clay table drives the whole thing. Every column is either a data source, an enrichment, an LLM decision, or a sync. Each row exits as a sequenced account.
Campaign type is determined per account by data availability. The ICP layer drives everything downstream: list filters, LLM prompts, copy angle, and channel priority.
Digital marketing agencies, SaaS, consulting, e-commerce exporters billing internationally
10–150 employees · $500K–$10M estimated revenue
Wise, Payoneer, or SWIFT bank wires — active competitor displacement targets
UAE, Saudi Arabia, Egypt, Jordan, Lebanon, Morocco
Campaign type is a column, not a workspace. All three plays share the same enrichment, scoring, and CRM mapping — they diverge only at the copy prompt, where they should.
Broad top-of-funnel targeting MENA businesses fitting the financial profile: industry, headcount, revenue, and location. Used when no tech signal or intent data is available.
Competitor displacement targeting companies using Wise, Payoneer, Deel, or Stripe in MENA. These accounts have already validated the pain — they just haven't found a MENA-native fix.
Account-fit signals: hiring a Finance Manager, raising a Series A, announcing international expansion. An LLM reads the raw source to extract why Sorbet is relevant right now.
A standard workflow labels a signal “hiring” and moves on. Here, an LLM reads the raw source text — the job post, the Crunchbase entry, the news article — and extracts a one-sentence synthesis: exactly why this account should hear from Sorbet today.
| Raw signal input | LLM-generated signal_summary output |
|---|---|
| LinkedIn post: “We’re hiring a Treasury Analyst in Dubai to manage our international vendor payments.” | “Scaling vendor payment volume across borders — actively building treasury function, strong fit for Sorbet.” |
| Crunchbase: “Raised $4.2M Series A, expanding into EU market.” | “Series A with EU expansion signals new cross-border payment flows, likely still on bank wires.” |
| Job post: “Finance Manager — Riyadh, must have Wise or Payoneer experience.” | “Explicitly using Wise/Payoneer — active competitor user, high displacement potential.” |
The prompt pattern used in the Clay AI column:
"You are a B2B sales researcher. Read this LinkedIn job post / Crunchbase entry / news article and explain in one sentence why it suggests this company has a growing cross-border payment need relevant to a stablecoin payment platform. Be specific. If there is no meaningful signal, say 'no signal'."
The signal_summary column feeds directly into the AI copywriting prompt in Step 06, so each email opens on a sentence the model has earned the right to write.
For each qualified account, Apollo and Sales Navigator surface three contacts by title. Persona type is assigned as a column — it drives a different message angle, offer framing, and CTA downstream.
Controls budget and platform decisions. At companies under 50 employees, this is the founder. Cares about cost savings, compliance, and ROI. CTA: 15-min call to run the numbers.
“Our SWIFT fees are eating 3–5% of every international transfer.”
Will advocate for Sorbet internally. Cares about workflow efficiency, fewer tools, and reliable reconciliation. Reached primarily via LinkedIn. CTA: full product demo.
“I’m reconciling three payment platforms manually every month.”
Executes payments day-to-day. Cares about speed, clarity on fees, and ease of use. At small teams, this is the founder doing it themselves. CTA: free account signup, 5-min setup.
“Transfers take 3–5 days and the fees only show up after.”
A GPT-4 column in Clay reads five enriched fields per row — company name, industry, payment tool in use, persona type, and the signal summary — then generates a personalized cold email and LinkedIn DM. No templates. No manual writing.
Hi Rania,
Saw Shift Creative is billing European brands regularly. If you’re routing those through Wise, you’re likely absorbing 1.5–2% FX margin on every invoice, plus the 3–5 day lag before funds clear.
Sorbet handles MENA cross-border on USDC rails: same-day settlement, 1% flat, ADGM-licensed. Built specifically for agencies in your position.
Worth a 15-minute call to run the numbers on your volume?
Émile
Hi Tariq,
Noticed Nana is building out a treasury function in Dubai. That usually means international vendor payments are scaling faster than current tools can handle.
We work with MENA SaaS teams who moved off bank wires to Sorbet. One platform to collect from US/EU clients, pay global vendors, and issue multi-currency invoices. Your new Treasury Analyst won’t need to touch three tools on day one.
Quick demo when they’re onboarded?
Émile
You are a senior B2B copywriter for Sorbet — MENA cross-border payments. 1% per transaction. No monthly fee. Same-day settlement. ADGM-licensed. Write a cold outreach email for this contact. Use the inputs below. PERSONA: {persona_type} — Decision Maker (CFO/CEO): lead with cost savings + ROI. CTA = 15-min call. — Champion (Head of Ops/Finance Mgr): lead with workflow efficiency. CTA = demo. — User (AP/Payroll): lead with ease of use and speed. CTA = free account signup. SIGNAL (why reaching out now): {signal_summary} PAYMENT TOOL: {payment_tool} COMPANY: {company_name} · {country} · {industry} RULES: - Subject: under 8 words. Specific to signal or persona. - Body: 3 short paragraphs, under 120 words total. - P1: Reference signal_summary as the reason for reaching out. - P2: One sentence on what Sorbet does, framed for this persona. - P3: Single, low-friction CTA. - No em-dashes. No exclamation marks. Tone: peer-to-peer.
Email, LinkedIn, and calls run simultaneously. Each channel plays to its strength: email for Decision Makers on cost and ROI, LinkedIn for Champions on workflow, calls reserved for accounts showing hard engagement signals.
No message. Let it land. Connection acceptance itself signals warm intent before the email hits.
AI-personalized per persona and payment tool. Subject line A/B tested across the cohort.
Brief and curiosity-driven. References company context, not the email thread.
Proof over pitch. Luna Studio reference. Softer CTA to reduce friction.
High-intent accounts only: 3+ email opens, link click, or LinkedIn profile view registered.
Low-pressure close or breakup. Voice note for highest-priority accounts only.
| Sending domains | 3–5 per campaign |
| Mailboxes per domain | 2–3 accounts |
| Warmup period | 4 weeks min. |
| Daily limit / mailbox | 30–50 emails |
| Send window | 8am–5pm recipient TZ |
| Gap between sends | 4–8 min randomized |
| Stop on reply | ON — critical |
Every reply triggers a webhook. An LLM classifies intent and outputs the next action. No manual triage, no inbox archaeology. A missed positive reply costs the full pipeline value of that account. A per-reply classification costs fractions of a cent.
"Classify this email reply as one of: Positive Interest, Soft Objection, Not Now, Wrong Person, Out of Office, Negative. Then write one sentence describing the ideal next sales action, specific to a cross-border payment platform pitch."
Prospect is engaged or asks for more information.
Auto-notify SDR via Slack, create deal in HubSpot, trigger pre-call research workflow.
Price, timing, or competitor concern raised.
LLM drafts a rebuttal email for SDR review — pre-filled with a counterpoint specific to the objection.
Genuine interest, wrong timing.
Pause sequence. Re-enroll in 30 days with an updated trigger email based on any signal changes.
Reply redirects to a different contact.
LLM extracts the new name and email from reply text, enriches in Clay, adds to active sequence.
Auto-reply detected.
LLM extracts return date, resumes sequence on that exact date automatically.
Prospect declines or asks to stop.
Immediate suppression across all channels. CRM status updated to Closed Lost.
Designed to clear 3–8% reply rate and 1–3% positive reply against MENA SMB targets, with 5–15 meetings booked per 1,000 contacts. At a steady 1,000 contacts a month, the system projects $48K–$192K of new ARR pipeline monthly, with the LLM doing the qualification, signal interpretation, and copywriting work that would otherwise require a two-SDR team.
Volume, Displacement, and Intent run off the same Clay table. Campaign type is a column, not a workspace. Same enrichment, same scoring, same CRM mapping — plays diverge only at the copy column, where they should.
WhyA separate campaign per angle orphans the data. One table compounds it.
A “hiring signal” tag tells you nothing. A one-sentence interpretation tells the copy prompt exactly why this account should hear from Sorbet today, so each email opens with a sentence the model has earned the right to write.
WhyCheaper than a fine-tune, faster than a human researcher, sharper than a tag.
Most outbound stacks stop being clever at the inbox. Here, a reply classifier is the third LLM in the system — routing the next action so positive interest never sits unread and objections come with a draft response ready.
WhyA missed positive reply costs the entire pipeline value of that account. A per-reply classification costs fractions of a cent.
