Lessons Learned Scaling a Multi-Tenant AI Product as a Meta Tech Provider.

Faigen is a multi-tenant SaaS that deploys AI agents on WhatsApp and Instagram for small and medium businesses in Kerala. A business signs up, configures their AI agent — product catalog, tone, FAQs, language — and their customers start getting instant AI replies on WhatsApp within 48 hours.

The AI handles order taking, appointment booking, product queries, and broadcast campaigns. It speaks Malayalam, Manglish, and English — auto-detecting what the customer writes in and replying in the same language.

I built the entire thing solo — backend, frontend, admin console, WhatsApp webhook infrastructure, AI integration, and the landing page. Getting registered as an official Meta Tech Provider was the first major milestone. Keeping it running reliably for multiple businesses simultaneously was the hard part.

To send messages via the WhatsApp Business Cloud API at scale — not just to verified test numbers — you need to be registered as a Tech Provider (ISV) on Meta's platform or work through a BSP (Business Solution Provider).

The Tech Provider route means you build directly on the Cloud API, manage your own WABA (WhatsApp Business Account), and are responsible for your own compliance. The process involves:

I learned the WABA webhook subscription lesson the hard way — after migrating to a new phone number on a different WABA and spending an hour debugging why messages were arriving at Meta but producing zero logs on my server.

The first version of Faigen had a fundamental multi-tenancy design flaw. I stored the AI system prompt and business configuration in environment variables and a single config object — assuming I would only ever run one AI agent at a time.

The correct design — which I moved to before onboarding any clients — stores all business configuration per company in MongoDB, scoped by a companyId on every document. The webhook handler looks up the company from the incoming phone number ID, loads their config, and runs their AI agent in isolation from every other company on the platform.

This pattern — company lookup from the incoming phone number ID, then everything scoped to that company — is what makes the platform genuinely multi-tenant. Adding a new client means creating a new Company document in MongoDB. Zero code changes.

WhatsApp sends webhook events asynchronously. If a customer sends three messages quickly — "Hi", "I want to order", "2 coir mats" — all three webhooks arrive at the server within milliseconds.

Without a queue, three concurrent AI requests fire simultaneously. The AI sees each message in isolation, without the context of the previous message. The third message ("2 coir mats") arrives with no context that the customer already said they want to order. The AI responds with a greeting — completely wrong.

The solution is a per-phone message queue. Messages from the same customer are processed sequentially — each one waits for the previous to complete before the AI runs.

The hardest part of building an AI agent that serves real businesses is not the infrastructure — it is getting the AI to behave consistently across thousands of conversations with customers who write in unpredictable ways.

The Menu Loop Problem

Early versions of the system prompt produced a consistent bug: the AI would show the main menu (Our Products / Place Order / Contact Us) on every single message — even mid-conversation. The customer would ask "how much is the coir mat?" and get a menu instead of an answer.

The fix was explicit instruction in the system prompt:

Getting Malayalam to Work Without a Malayalam Model

Gemini and Groq do not have dedicated Malayalam language models. But they understand Malayalam reasonably well because it appears in their training data. The trick is making the language detection and response explicit in the prompt rather than hoping the model figures it out.

Keeping the AI On-Topic

Without explicit restrictions, the AI would answer anything — geography questions, recipe requests, general knowledge. For a business AI agent this is wrong. The agent should only discuss the business.

Every AI response costs money. Gemini Flash is cheap — roughly ₹0.02-0.05 per response — but at scale with multiple businesses and active customers, it adds up. More importantly, a single abusive user could generate hundreds of messages per day and run up significant costs.

24-Hour Message Limit

Every conversation has a maximum number of customer messages per 24-hour rolling window. The limit is configurable per company — defaulting to 15 for the Faigen demo and higher for paid clients.

When the limit is hit, instead of going silent (which confuses the customer) or throwing an error, the system sends a single warm handoff message with contact details — converting the limit hit into a sales touchpoint.

Broadcast Credit System

Broadcast campaigns — sending WhatsApp messages to hundreds of customers at once — cost real money. Meta charges per message: ₹0.15 for Authentication templates, ₹0.88 for Marketing templates.

Every company has a broadcastCredits balance in MongoDB. Credits are deducted before sending — if the balance is insufficient for the full campaign, the broadcast is rejected with a clear error rather than sending partial messages and leaving the company confused.

Every outbound WhatsApp message that is not a reply within a 24-hour customer-initiated window must use a pre-approved message template. Templates are categorised by Meta into three types — Authentication, Utility, and Marketing — each with different pricing and approval criteria.

Authentication templates (for OTPs) are approved fastest — usually within minutes. The format is fixed by Meta:

Marketing templates get the most scrutiny. They cannot contain misleading claims, cannot promise guaranteed returns, and must clearly identify the business sending them. I have had templates rejected for being too generic ("Hello, check out our offers") and approved for being specific ("Your Onam offer from [Business]: 20% off all products this week").

The practical lesson: be specific in template content, use the business name prominently, and avoid superlatives like "best" or "guaranteed." Approval times range from minutes to 24 hours.