How We Built an AI Sales Agent for a Saudi manufacturing Company

In early 2026, a mid-sized Saudi manufacturing company came to us with a familiar pain point: their sales team was spending 60% of their time on cold outreach, lead qualification, and follow-up — leaving only 40% for actual closing. With 15 sales reps covering three product lines across five regions, they were losing leads in the cracks and struggling to scale. Their ask was simple but audacious: "Build us an AI sales agent that can do the grunt work so our people can focus on selling."

This is the story of how we designed, built, and deployed a fully autonomous AI sales agent for a real B2B manufacturing business — the architecture, the challenges, the results, and the lessons learned along the way.

The Challenge: Industrial B2B Sales at Scale

Manufacturing B2B sales in Saudi Arabia comes with unique complexity. The client — let's call them "GulfMetals" — manufactures steel components, aluminum extrusions, and specialized fasteners. Their buyers range from small construction contractors to large government infrastructure projects. Every segment requires different outreach, different technical knowledge, and different pricing models.

The core problems we needed to solve:

• Cold outreach inefficiency: Reps manually researched prospects, drafted emails, and made calls — producing 5-10 qualified leads per week per rep at best.
• Multilingual complexity: The sales cycle spans Arabic and English, and some technical documents are only available in Arabic. Reps needed translation support for every interaction.
• Long follow-up cycles: B2B deals in manufacturing take 45-90 days with multiple touches required. Human reps often forgot to follow up at the right intervals.
• Inconsistent qualification: Without a standardized lead scoring framework, reps qualified leads based on gut feel rather than data.
• CRM data decay: Contact information, company details, and past interactions lived in spreadsheets or outdated CRM entries.

The client wanted a system that could autonomously research prospects, initiate outreach in both Arabic and English, qualify leads against a scoring model, schedule meetings, and hand off only high-intent opportunities to human reps.

The Architecture: Building the Agentic Sales Stack

We designed the AI sales agent as a multi-agent orchestration system with five specialized sub-agents, each responsible for a distinct part of the sales process. Here's how the stack is laid out:

Technical Deep Dive: How the Agent Works

Under the hood, the system runs on a modular agentic framework built on top of Hermes Agent as the runtime orchestrator. Each sub-agent is a self-contained process with its own tools, memory, and prompt instructions.

Tool-Augmented Reasoning

Every sub-agent uses function-calling LLMs (we used DeepSeek V4 and GPT-4o) to decide which tools to invoke. The Research Agent has tools for web scraping, LinkedIn API queries, and CRM reads. The Communication Agent has tools for send_email, send_whatsapp, and schedule_meeting. This tool-augmented approach means the LLM never guesses — it always consults real data before acting.

Persistent Memory & State Management

Each lead's journey is tracked as a state machine: New → Researched → Contacted → Engaged → Qualified → Hot (escalated). The orchestrator persists state in a PostgreSQL database (using Shayntech TimeTravel for audit trail, naturally) and uses a Redis-backed conversation cache for real-time context. If the system crashes mid-call, it picks up exactly where it left off.

Multilingual Natural Language Pipeline

For Arabic-language interactions, we built a pipeline that: (1) detects language from the incoming message, (2) uses a dedicated Arabic LLM fine-tune for understanding nuanced Saudi dialect and business Arabic, (3) generates responses in the detected language with culturally appropriate phrasing (Islamic greetings, respect titles, local business idioms), and (4) maintains a bilingual thread so a conversation that starts in Arabic and switches to English remains coherent. This was critical — our client's buyers commonly switch between the two languages mid-conversation.

Results: What Happened After 90 Days

After three months of production deployment, the numbers speak for themselves:

Lessons Learned: What Worked & What Didn't

Building an AI sales agent for a real B2B environment taught us hard-won lessons that shaped our approach to agentic systems:

• Human-in-the-loop is non-negotiable: We started with full autonomy but quickly learned that certain interactions — price negotiations, complex technical questions, sensitive account situations — need a human. The escalation mechanism must be fast and seamless.
• Arabic language quality is hard: Off-the-shelf LLMs handle Modern Standard Arabic (MSA) well, but Saudi dialect, business Arabic, and code-switching are significantly harder. Fine-tuning on domain-specific Saudi business conversations was essential for natural-sounding interactions.
• Data quality matters more than AI quality: Garbage in, garbage out. Clean CRM data, accurate contact information, and up-to-date company profiles had a bigger impact on agent performance than any model upgrade.
• Voice calls are the hardest channel: Email and WhatsApp work reliably with current LLM quality, but voice calls (especially cold calls) are still challenging. Background noise, accents, interruptions, and unexpected questions push the limits of current voice-to-voice pipelines.
• Monitoring and observability are essential: Every agent action needs to be logged, traceable, and reviewable. We built a dashboard showing all agent decisions, conversation transcripts, and escalation logs so the sales manager could audit any interaction.

What's Next: The Roadmap

We're actively extending this platform for GulfMetals and other manufacturing clients. The roadmap includes:

• WhatsApp Business API deep integration: Full product catalog sharing, quotation delivery via WhatsApp, and payment link embedding within the conversation thread.
• Multi-channel conversation merging: Seamlessly connect email, WhatsApp, phone, and even in-person meeting notes into a single lead timeline.
• Predictive lead scoring with ML: Beyond rules-based scoring, training a machine learning model on historical conversion data to predict which leads are most likely to close and recommend optimal outreach sequences.
• Arabic-first document understanding: The agent will be able to read and extract information from Arabic PDFs (BOQs, contracts, technical specs) to automatically populate CRM fields and generate customized proposals.
• Expansion to other Gulf markets: Adapting the agent for UAE, Qatar, and Kuwait, with market-specific compliance, business culture, and dialect handling.