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

Sales teams in the manufacturing sector face a fundamental problem: their best reps spend less than 35% of their time actually selling. The rest goes to data entry, lead research, quote generation, and follow-up reminders. When a mid-sized Saudi manufacturer of industrial steel components approached us with declining conversion rates and a sales team drowning in administrative overhead, we knew the solution had to go beyond a simple CRM overhaul. We built them an AI Sales Agent — an autonomous system that handles the full sales cycle from lead qualification to quote delivery, operating in both English and Arabic, and trained on their specific product catalog.

The Challenge: Manufacturing Sales Is Different

Unlike retail or SaaS, manufacturing sales involves complex product configurations, custom specifications, and long decision cycles. A single inquiry for fabricated steel beams might require checking raw material availability, calculating load-bearing specifications, coordinating with the engineering team, and generating a multi-line quote — all before the prospect loses interest.

Our client faced three specific pain points:

• Lead response time averaged 8+ hours: Inbound WhatsApp and website inquiries sat unanswered overnight, and by morning many prospects had already contacted competitors.
• Quote generation took 45 minutes per request:Sales reps manually looked up pricing tiers, applied volume discounts, calculated delivery fees, and formatted quotes in Excel.
• Follow-up was inconsistent: Only 23% of qualified leads received a follow-up within 48 hours because reps were too busy with administrative tasks to prioritize outreach.

The Architecture: An Autonomous Sales Pipeline

We designed the AI Sales Agent as a multi-stage pipelinethat mirrors how a top-performing human sales rep works — but operates 24/7 across all channels simultaneously. The system sits on top of their existing Odoo ERP instance and integrates with WhatsApp Business API, their website contact forms, and email.

The pipeline breaks down into four autonomous stages:

• Stage 1 — Lead Reception & Triage: Incoming messages are classified by intent (new inquiry, existing order status, complaint, or spam). High-value leads are flagged for immediate engagement.
• Stage 2 — Qualification & Discovery: The agent asks targeted questions to determine project scope, quantities, material specifications, and delivery timeline — all in natural, conversational Arabic or English.
• Stage 3 — Product Matching & Proposal:Using a vector search over the product catalog, the agent finds the best-matching SKUs, checks inventory, and generates a preliminary quote through the ERP integration.
• Stage 4 — Follow-up & Handoff: If the prospect goes silent, the agent sends intelligent follow-ups. When a deal reaches a predefined confidence threshold, it hands off to a human sales rep for closing.

Bilingual Understanding of Technical Specifications

One of the hardest parts of this project was teaching the AI to understand technical manufacturing language in two languages. A prospect might say "I need 200 meters of 12mm rebar, grade 60, with epoxy coating" in Arabic, or "أحتاج 200 متر من حديد التسليح 12 ملم درجة 60 مع طلاء إيبوكسي" — the agent needs to extract the same structured data from both.

We built a domain-specific entity extraction layerthat recognizes over 120 distinct product attributes: material grades, dimensions, coating types, tensile strength ratings, certification standards (SASO, ASTM, BS), and packaging preferences. The extraction runs through a fine-tuned LLM pipeline with RAG over the company's technical datasheets, ensuring every specification is checked against the actual product catalog.

An interesting challenge was handling the Arabic numeral variants — the same quantity might be written as "200", "٢٠٠", or "200." in different messages. We added a preprocessing normalization layer that standardizes all numeric inputs before they hit the extraction model.

Seamless Odoo ERP and WhatsApp Integration

The AI Sales Agent doesn't exist in isolation — it reads from and writes to the company's Odoo ERP in real time. Every interaction updates the CRM record, every quote is generated as an Odoo quotation, and inventory levels are checked before promises are made.

The technical integration stack:

• Odoo XML-RPC API: For reading product data, creating leads, opportunities, and quotations, and updating pipeline stages.
• WhatsApp Cloud API: Bidirectional messaging with template-based rich responses including product images, PDF quotes, and payment links.
• Redis-backed session management: Each conversation maintains state across multi-day sales cycles — the agent remembers context even if a prospect returns after a week of silence.
• Webhook event bus: Real-time notifications to human reps when the agent flags a high-value opportunity, encounters an ambiguous request, or completes a successful quote.

Results: What Happened After Deployment

We deployed the AI Sales Agent in phases over six weeks. The first month covered WhatsApp lead handling only; the second added proactive outbound follow-up. Here are the numbers after 90 days of production use:

• Lead response time dropped from 8 hours to 23 seconds:The agent answers every inquiry within thirty seconds, 24/7, including weekends and holidays.
• Quote generation time fell from 45 minutes to under 2 minutes:Automated SKU matching, pricing, and discount application reduced manual work by 96%.
• Lead-to-quote conversion rate increased by 47%:Instant responses and consistent follow-up kept prospects engaged through the sales cycle.
• Sales team capacity grew by 3x: With administrative tasks automated, the same four-person sales team now handles three times the volume without hiring additional staff.
• 97% Arabic intent recognition accuracy: The bilingual NLP pipeline correctly classifies and processes Arabic-language inquiries with Saudi dialect variations.

Lessons Learned: What We'd Do Differently

No real-world deployment is perfect, and this one taught us several important lessons:

• Start with a narrow scope: Our initial design tried to handle too many edge cases at once (returns, complaints, order modifications). In retrospect, launching with just lead qualification and quote generation — then expanding — would have been faster and less risky.
• Human-in-the-loop is non-negotiable: Even with 97% accuracy, ambiguous requests need human judgment. We built escalation paths into every stage of the pipeline, and the sales team trusts the system much more because they know they can step in at any point.
• Arabic NLP needs dialect-specific training:Standard Arabic (MSA) models perform poorly on Saudi business conversations, which mix Najdi and Hejazi dialect terms with industry jargon. Fine-tuning on a corpus of the client's actual WhatsApp transcripts was the single most impactful optimization we made.
• Pricing logic varies wildly by customer: Long- standing customers get different discount tiers than new ones, and some products have volume break points. Storing per-customer pricing rules in a structured database — rather than trying to encode them in the LLM prompt — was the right architectural decision.

What's Next for AI-Powered Sales in Manufacturing

This project proved that agentic AI is ready for industrial applications — not just customer support chatbots, but revenue-generating sales engines that operate autonomously. We're already building the next version with:

• Multi-agent negotiation: An agent that can negotiate pricing within predefined bands, escalating only when the discount request exceeds authority limits.
• Voice channel support: Extending to voice calls with real-time transcription and response, so the agent can handle phone inquiries with the same intelligence as chat.
• Predictive lead scoring: ML models that analyze historical conversion patterns to prioritize leads most likely to close, so human reps focus exclusively on high-probability opportunities.
• Cross-platform consistency: Unified customer profiles across WhatsApp, email, website, and phone — so a prospect can start an inquiry on WhatsApp and continue on email without repeating themselves.

For the Saudi manufacturing sector, which contributes over 10% of the kingdom's GDP and is a cornerstone of Vision 2030, AI-powered sales automation isn't just a competitive advantage — it's becoming a business necessity. Companies that adopt these tools now will have a significant lead in efficiency, response time, and customer experience as the market becomes increasingly digital-first.