How much does it cost to build an AI Sales Agent that answers factory inquiries, quotes by rules and hands off to humans?
The cost usually has seven parts: knowledge base and product data, quoting rules, FAQ and reply scripts, channel deployment, human handoff, ongoing tuning, and model usage fees. WhatsApp per-message fees are a separate channel cost paid to Meta or a Business Solution Provider (BSP). Public vendor selection guides in China (2025, for reference only) put SME SaaS at about RMB 5,000–30,000/year, advanced SaaS at RMB 100,000–300,000/year, and private deployment or deep customization at RMB 100,000–500,000+, plus 10%–20% annual maintenance.
How are service fees calculated when you connect product knowledge, quoting rules and FAQ, and deploy on WhatsApp and your website?
Usually in two parts—a one-time build and ongoing service—with channel fees billed separately. SaaS tools charge a subscription and leave the build to you; custom projects charge development plus annual maintenance; done-for-you services bundle build and operation into an annual fee. For pricing models (per seat, per message, done-for-you) see how DeepSales is priced; this article breaks down what the work actually is.
How much work is the knowledge base?
It depends on data quality and SKU count. A clean spec sheet is quick; specs scattered across PDFs, old websites and reps’ chats take time to reconcile. The work covers documents (website, catalogs, FAQ), structured product fields (SKU, specs, price, MOQ, certifications) and, most valuable, how your best reps actually answer. Clean data matters because retrieval reduces but does not remove hallucination: a NAACL 2024 paper (ServiceNow) found up to 21% of generation steps hallucinated without retrieval and under 7.5% with it, while Stanford RegLab (2024) found retrieval-based legal AI tools still hallucinated 17%–33% of the time. Connecting your data usually means building a retrieval index, not training a custom model, so be wary of a large “model training” line item.
What does encoding quoting rules involve?
- Tiered pricing by quantity, market or customer tier.
- Incoterms: what EXW, FOB and CIF include and which is the default.
- Currency and how often rates update.
- Freight: lookup by port, or a range confirmed by a rep.
- Floor prices and approval: what the AI may quote directly and what must go to a human first.
The last item carries the risk. In Moffatt v. Air Canada (BC Civil Resolution Tribunal, February 2024), the airline was held liable for its chatbot’s wrong fare policy; the argument that the chatbot was a separate entity was rejected. The safe pattern is AI-drafted quotes approved by a person. See automating RFQ quoting.
Are FAQ and reply scripts charged separately?
Usually they are part of the build. Ask who writes them, whose past conversations they are based on, and who updates them after launch.
What does deploying on WhatsApp and your website cost?
- WhatsApp API number: via Meta’s Embedded Signup your own number can join the Cloud API; in Coexistence mode it still works in the WhatsApp Business App with synced chats (Meta developer docs).
- Meta message fees: per-message pricing since 2025-07-01 (Meta developer docs). Several BSPs including respond.io and Gupshup disclosed in September 2026 that from 2026-10-01 service messages inside the window are charged, with the first 1,000 per number per month free—check Meta’s official page. This is paid to Meta or the BSP, not the AI vendor.
- Free entry window: 72 hours for buyers arriving via Click-to-WhatsApp ads or a Facebook Page button.
- Website chat: usually a snippet of code; little technical work.
AI that serves your own customers remains allowed under Meta’s updated terms, which ban general-purpose AI assistants for all users from 2026-01-15 (TechCrunch, 2025-10-18).
Why is human handoff a cost?
You need to define triggers, routing and what reps see. China’s recommended standard GB/T 47746—2026, effective 2026-09-01, requires a clear human entry point, priority handoff for complex issues, and passing identity and history so customers don’t repeat themselves (Xinhua, 2026).
What about ongoing tuning and model usage fees?
Tuning is the largest recurring cost: weekly reviews, filling gaps, adjusting rules as products and prices change. For custom projects it appears as 10%–20% annual maintenance; for SaaS it is your own staff time. Model usage fees depend on inquiry volume, conversation length, languages and image reading; ask whether they are included and how overages are billed.
How do the four approaches compare?
| Build in-house | SaaS tool | Custom project | Done-for-you | |
|---|---|---|---|---|
| Upfront | Engineering time | Low | High, ~RMB 100k–500k+ (ref.) | Included in annual service |
| Ongoing | Model fees, servers, staff | Subscription, ~RMB 5k–30k or 100k–300k/yr (ref.) | 10%–20% maintenance/yr | Annual service fee |
| WhatsApp fees | Extra | Extra | Extra | Extra |
| Who staffs it | You | You | Vendor builds, you run | Vendor builds and runs; you supply data and approvals |
| Best for | Teams with engineers | Low volume, hands-on | ERP integration, private deployment | Factories without an AI operator |
Ranges are from public vendor guides (2025), not industry statistics. Some SaaS tools list lower prices, e.g. SaleSmartly Pro from about $15.9/month and Max from about $199/month (vendor site).
What hidden costs get missed?
- Your team’s time for data prep, quote approval and training.
- Per-message WhatsApp fees that grow with volume.
- Change requests when products or prices update.
- The cost of fixing wrong answers.
- Currency swings on dollar contracts.
- Exit costs: can you export chats and data?
- Extra numbers and websites.
How does it compare with the cost of a sales rep?
An experienced export sales rep in China costs about RMB 120,000–200,000 a year plus commission and social insurance, covering one time zone for eight hours a day (common industry level). Use this as a reference, not a replacement: negotiation and closing deals stay with people. Speed matters: HBR (2011) found firms contacting leads within an hour were nearly 7 times as likely to qualify them; the average first response was 42 hours and 23% never replied.
How does DeepSales price and what is included?
DeepSales does not publish prices. It is an annual service priced by scope (SKUs, languages, channels, whether quote drafts are needed), covering build, delivery and operation—not per seat. Our team builds it from your materials and best reps’ conversations and keeps tuning it. It runs on your own website chat, WhatsApp, Messenger, Instagram and LINE accounts; email inquiries can be managed in the platform but are handled by people. Higher tiers draft quotes for human approval; negotiation stays with people. WhatsApp fees are paid to Meta or the BSP. If you get few inquiries and just want a chat widget, a low-cost SaaS fits better; if you need private deployment with ERP integration, choose a custom project. Otherwise, book a demo and we will run a sample conversation on your catalog and quote by scope.