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

In early 2026, a mid-sized manufacturing company in Dammam, Saudi Arabia, came to us with a familiar but urgent problem: their sales team was drowning. Inbound inquiries from WhatsApp, email, and their website were piling up faster than their five-person sales team could handle. Quotes were delayed, follow-ups were missed, and they estimated they were losing 35% of potential leads simply because no one got back to them fast enough.

They had tried chatbots before. Off-the-shelf solutions from Zendesk and Tidio had been deployed, but their generic responses frustrated customers and couldn't handle the technical depth required for industrial manufacturing inquiries. The client needed something fundamentally different: an AI sales agent that could genuinely understand their products, their pricing, and their customers — and act on that understanding without human hand-holding.

This is the story of how we built it.

The Challenge: Sales at Scale in Manufacturing

Manufacturing sales is fundamentally different from e-commerce or SaaS sales. A single inquiry might involve custom specifications, material grades, delivery timelines, and bulk pricing tiers — none of which fit neatly into a product catalog. The client's team was spending 60% of their time answering the same 20-30 questions repeatedly, and the remaining 40% on manual quotation generation.

Key pain points included:

• Response time: Average first reply took 4+ hours during working hours, and inquiries outside 8 AM-6 PM would wait until the next day.
• Knowledge fragmentation: Pricing rules, product specs, and shipping policies were scattered across PDFs, Excel sheets, and senior sales reps' memories.
• Language barriers: The client handled inquiries in both Arabic and English, and their existing chatbot only supported English.
• Quotation complexity: Each quote required checking current raw material prices, applying customer-specific discounts, and calculating shipping — a process that took 30-45 minutes per inquiry.

Our Approach: Why Agentic AI Was the Answer

Traditional chatbots operate on a retrieval-augmented generation (RAG) pattern: they look up information and regurgitate it. But a sales agent needs to do more than answer questions — it needs to act. It needs to cross-reference inventory availability against order quantities, calculate dynamic pricing based on volume tiers, create quotations, and escalate to humans at the right moment.

This is where Agentic AI enters the picture. Instead of a single chatbot, we designed a multi-agent system where specialized AI agents collaborate like a real sales team:

• The Receptionist Agent: Greets the customer, identifies their language (Arabic or English), and routes the inquiry to the right specialist agent.
• The Product Knowledge Agent: Maintains deep knowledge of 1,200+ SKUs, their specifications, technical datasheets, and compatible alternatives.
• The Pricing Agent: Calculates real-time quotes using the latest raw material indices, customer discount tiers, and volume breaks.
• The Quotation Agent: Generates professional PDF quotations, stores them in the CRM, and sends them to the customer via their preferred channel.
• The Orchestrator: Coordinates all the above, decides when to hand off to a human, and maintains conversation context across the entire interaction.

Building the Knowledge Foundation

The single most important factor determining the success of any AI agent is the quality of the knowledge it operates on. We spent three weeks working shoulder-to-shoulder with the client's senior sales team, extracting their expertise and structuring it into a machine-readable knowledge graph.

We ingested:

• Product catalog: 1,200+ SKUs with full technical specifications, material properties, and dimensional data.
• Pricing engine: A custom rules engine that factors in base costs, market indices, customer tier, and order volume.
• Policy documents: Shipping terms, return policies, warranty conditions, and payment terms in both Arabic and English.
• Historical conversations: 15,000+ past WhatsApp and email threads used to fine-tune response patterns and identify common edge cases.
• CRM integration: Real-time access to customer history, past orders, and open support tickets via API.

Multi-Channel Deployment: WhatsApp, Web, and Email

Saudi Arabia has one of the highest WhatsApp penetration rates in the world — over 90% of adults use it daily. For a B2B manufacturing company, WhatsApp was not optional; it was the primary sales channel. We built the AI agent to operate seamlessly across three channels simultaneously:

• Whatsapp Business API: The primary channel. Customers send a message, and the AI agent responds within seconds. Supports product inquiries, quote requests, order status checks, and payment follow-ups.
• Website live chat: Embedded on the company website with the same agent operating behind it, with context carried seamlessly if a customer switches from web to WhatsApp.
• Email: For formal quotation requests and documentation, the agent monitors a dedicated sales inbox and responds with generated quotes attached as PDFs.

Cross-channel context was critical. If a customer started a conversation on WhatsApp about a steel pipe quotation, then emailed to follow up with additional specifications, the AI agent recognized it was the same customer and the same conversation — no repetition, no frustration.

The Language Challenge: Arabic B2B Sales

Arabic poses unique challenges for AI agents. Dialectal variation, right-to-left formatting in generated PDFs, and culturally specific communication norms all had to be addressed. A customer might send a message in Gulf Arabic about "مواسير حديد" (iron pipes) mixed with English technical terms like "Schedule 40" — the agent needed to handle both seamlessly.

We approached this by:

• Fine-tuning on Arabic sales data: Using 8,000+ Arabic-language sales conversations to improve the model's understanding of Gulf dialect, industry terminology, and polite business Arabic.
• Bilingual knowledge retrieval: Storing all product information in both Arabic and English, with the agent automatically serving the right language.
• RTL PDF generation: Building a custom PDF generator that produces professional Arabic quotations with proper right-to-left formatting, Arabic numerals, and Hijri date support.
• Cultural sensitivity: Training the agent on appropriate greeting and closing conventions, the use of honorifics, and the importance of religious greetings like "Assalamu Alaykum" during appropriate times.

Results: What Happened After Launch

The AI sales agent went live on March 1, 2026. After three months of production use, the numbers speak for themselves:

Perhaps most importantly, the human sales team's job satisfaction improved dramatically. Instead of spending their days answering the same repetitive questions, they focused on high-value activities: negotiating major contracts, building relationships with key accounts, and strategic planning. The AI agent handled the 80% of inquiries that were routine, while the human team stepped in for the 20% that required judgment, negotiation, or creative problem-solving.

Behind the Scenes: Technical Architecture

For teams considering a similar build, here's a high-level view of the architecture:

• LLM Foundation: DeepSeek V3 as the core reasoning engine, chosen for its strong Arabic language performance and cost efficiency at scale.
• Agent Framework: Built on the Hermes Agent orchestration layer, giving us multi-agent orchestration, tool use, and conversation memory out of the box.
• Channel Gateways: Custom middleware integrating WhatsApp Business API (via WATI), website chat widget, and IMAP/SMTP for email.
• Quotation Engine: A lightweight Python service that generates dynamic pricing using real-time raw material feeds from regional exchanges.
• Escalation Protocol: When the AI agent detects uncertainty (confidence below 0.7), it automatically creates a Slack ticket for the human sales team with full conversation context and suggested next steps.

Key Takeaways for Manufacturing Leaders

Based on this project, here are the lessons we believe apply to any manufacturing company considering AI-powered sales automation:

1. Start with knowledge, not technology. The AI model is only as good as the data it has access to. Invest the time upfront to structure your product data, pricing rules, and policies — it will pay for itself ten times over in agent performance.
2. Design for escalation, not replacement. The goal is not to eliminate your sales team but to amplify them. Every routine inquiry the AI handles frees up a human for a conversation that actually needs human judgment.
3. Lead with your best channel. For this client, WhatsApp was the priority. For another, it might be email or a web portal. Identify where your customers already are and meet them there.
4. Localize everything. Arabic B2B communication has distinct conventions. A translation layer is not enough — the agent must understand the cultural and linguistic context of the region it operates in.
5. Measure what matters. Response time, escalation rate, and quotation velocity are the metrics that directly impact revenue. Track them from day one.