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

The Challenge: From Lead to Order in Six Days

A mid-sized manufacturing company in Dammam's Industrial City was losing opportunities they knew they should be winning. Their sales team of 12 handled an average of 85 qualified leads per month, but the average time from initial inquiry to sending a formal quotation was 6.2 days. By then, nearly one in three prospects had either chosen a competitor or lost urgency.

The founder put it bluntly: “We have a production line that runs 24/7, but our sales pipeline flows at the pace of a single human with a spreadsheet and a phone.” The bottleneck wasn't demand — it was response. Every lead that entered the system required the same laborious sequence: manual qualification, internal pricing lookup, email ping-pong for technical specifications, and a final quotation assembled from four different sources.

This is the story of how we built an autonomous AI sales agent that collapsed that six-day pipeline to under four minutes, and what we learned building it for a real Saudi manufacturing operation.

Inside the Traditional Sales Workflow

Before the AI agent, every lead followed this manual gauntlet:

• Lead triage (45 min): A junior sales rep manually reviewed each inbound inquiry — WhatsApp, email, website form — and decided whether it belonged to their division or should be forwarded elsewhere.
• Product matching (1.5 hours): The rep cross-referenced customer requirements against a 2,800-product catalog stored across three separate Excel sheets and a legacy ERP system. Product codes didn't always match, requiring manual verification.
• Pricing lookup (2 hours): Tiered pricing, bulk discounts, and customer-specific agreements were maintained in yet another spreadsheet. Every quotation required a senior manager to manually approve any deviation from standard rates.
• Quotation assembly (1.5 hours): With pricing and product data gathered, the rep assembled a PDF quotation using a Word template — copying, pasting, formatting, and re-checking for errors.
• Internal review (4 hours average): Every quotation over SAR 50,000 required sign-off from the sales manager, who was often in meetings or on the factory floor. Quotations sat in inboxes for half a day waiting for approval.
• Delivery & follow-up (30 min): The quotation was emailed or WhatsApp'd to the customer, and a manual reminder was set for three days later — if the rep remembered.

At 85 leads per month, this workflow consumed roughly 720 person-hours of sales team capacity — equivalent to employing two full-time people just to produce quotations. And that was before any actual selling, relationship building, or account management.

How We Built the AI Sales Agent

The core architecture is a multi-agent system running on a lightweight orchestration layer. Instead of building one monolithic model to “do sales,” we created a team of specialized sub-agents that collaborate on each lead:

The Inbound Classifier receives every message — WhatsApp text, email, or web form submission — and extracts structured intent: is this a new lead, a follow-up on an existing quotation, a technical question, or a complaint? It routes accordingly. For new leads, it extracts company name, contact details, product interest, and urgency level.

The Product Matcher queries the digitized product catalog using vector search. When a customer asks for “industrial shelving, 3-meter height, galvanized,” the matcher doesn't do keyword lookup — it understands the semantic meaning and returns the closest matching SKUs ranked by relevance. It also flags ambiguities (e.g., “galvanized steel or aluminum?”) and can proactively clarify via WhatsApp before proceeding.

The Pricing Engine applies customer tier, volume discounts, and any active promotions from a central pricing rules database. It calculates the full quotation total, including VAT and delivery, in under three seconds. Quotations outside standard parameters are flagged for human review but still sent with a “pending approval” note — so the customer receives a response immediately rather than waiting.

The Quotation Builder generates a professional Arabic/English bilingual PDF quotation with the company logo, customer details, line-item breakdown, payment terms, and a QR code linking to a secure approval portal. It uses the same brand template the company was assembling manually — but with zero formatting errors.

The Delivery Coordinator sends the quotation via the customer's preferred channel (WhatsApp or email), logs the interaction to the CRM, and schedules automated follow-ups at 3, 7, and 14 days if no response is received. Each follow-up is contextual — referencing the specific products quoted and offering to clarify pricing or arrange a site visit.

Measured Results: From Days to Minutes

After a six-week build and a two-week parallel-run testing phase, the AI sales agent went live in March 2026. The results over the first 90 days exceeded every projection:

The most striking finding: the sales team didn't shrink — they re-focused. The 720 monthly person-hours previously consumed by quotation assembly were reinvested into outbound prospecting, customer visits, and upselling existing accounts. The 12-person sales team effectively became a 12-person team doing the work of 20.

Beyond Speed: Error Reduction and Customer Trust

While response time was the headline metric, the deeper value surfaced in error reduction. During the parallel-run phase, we compared 50 AI-generated quotations against their manually produced counterparts. 12 of the 50 manual quotations contained errors — wrong product codes, miscalculated discounts, or missing line items. The AI agent: zero errors across all 50.

The financial impact of errors was substantial. Before the AI agent, the company was issuing an average of 8 credit notes per month to correct billing mistakes on quotations that had already been accepted. Each correction consumed another 45 minutes of admin time and, more importantly, eroded customer trust. “When a customer sees three different prices for the same product over a month, they start shopping around,” the sales manager told us. Post-deployment, credit notes dropped to fewer than one per month.

There was also an unexpected sales uplift from non-business-hours responses. The AI agent operates 24/7. When a procurement manager in Riyadh sent a WhatsApp inquiry at 10:30 PM, they received a complete quotation by 10:34 PM. The company's sales team had never responded to a late-night inquiry before — and they discovered that roughly 18% of their highest-value leads arrive outside working hours.

Built for the Saudi Market

Several design decisions were specific to the Saudi manufacturing context:

• Bilingual quotations: Every quotation is generated in both Arabic and English, with Arabic text rendered in a proper right-to-left layout. The agent detects the customer's language preference from the initial message and prioritizes accordingly.
• WhatsApp-first design: Over 70% of the company's inbound leads arrive via WhatsApp. The agent integrates directly with the WhatsApp Business API, maintaining conversation history and sending rich media (PDF quotations, product images) inline.
• ZATCA-compliant invoicing: The quotation engine applies VAT at the correct 15% rate and generates data that feeds directly into ZATCA-compliant e-invoicing, eliminating a separate data-entry step when the order converts.
• Saudi holiday awareness: The agent adjusts delivery timelines based on Saudi public holidays (Eid, National Day) and the Islamic calendar — a detail that manual processes frequently missed, causing unrealistic delivery promises.
• Local number formatting: Prices are displayed in SAR with Arabic numeral support, and large numbers use the local grouping convention (e.g., ١٬٤٠٠٬٠٠٠ SAR) for Arabic-language quotations.

These aren't nice-to-haves. When 85% of your customer base operates in Arabic, and the entire procurement cycle runs through WhatsApp, an English-only email-based AI agent would fail at step one. The localization layer is the difference between a demo and a deployment.

What It Took: Six Weeks from Go-Ahead to Go-Live

One concern we hear from every manufacturing company is disruption fear: “We can't afford to stop our sales process while you install new software.” The answer: you don't have to. Here's the actual rollout timeline:

• Week 1 — Discovery and data audit: We mapped the full sales workflow, documented every Excel sheet and ERP table, and identified the eight data sources the agent needed to query. No code was written; this was pure process archaeology.
• Week 2 — Product catalog digitization: The 2,800 SKU catalog was cleaned, normalized, and embedded into a vector database. Duplicate entries were removed, missing specifications were filled in, and pricing tiers were rationalized.
• Week 3-4 — Agent development: The five sub-agents were built, integrated, and connected to the WhatsApp API and email. The orchestrator logic for agent handoffs, error recovery, and human escalation was implemented.
• Week 5 — Parallel run: The AI agent ran alongside the human team. Every inbound lead was handled by both, and outputs were compared side by side. This is where we discovered the 24% error rate in manual quotations and tuned the agent's pricing edge cases.
• Week 6 — Cutover and training: The AI agent became the primary handler. The sales team spent two half-day sessions learning how to review AI-generated quotations, handle escalations, and use the agent dashboard. Zero downtime.

Total time from initial agreement to live production: six weeks. Total new software the client needed to install: zero — everything runs as a cloud service integrated with their existing WhatsApp and email infrastructure.