Buy an AI agent or build your own on a language model API? What a build actually involves, what it costs to run, and when each is right
With language model APIs cheap and capable, a technical founder's first thought is "I could build this in a weekend". The model call, yes. The rest, no. A WhatsApp AI agent is a WhatsApp connection, a retrieval system, safety rules, an inbox, booking, monitoring and a maintenance burden that never ends. Here is the honest list, and when building is still right.
Quick answer
Building your own agent means: connecting to WhatsApp (official API with approval and fees, or a linked-device library with restriction risk), splitting and indexing your knowledge for retrieval, writing prompts and guardrails, adding handover, an inbox for people, booking logic, per-number rate limits, spam handling, logging, backups, and keeping up with model and WhatsApp changes. Running costs are model tokens plus hosting plus, above all, your time. Build when the agent must live inside your own systems or you are a software business. Buy when the job is answering customers and booking them; a product like ReplyKit provides the whole list for $29 to $99 a month.
What a build actually includes
- WhatsApp connection. Official API (Meta business verification, a provider, templates, per-message fees) or a linked-device library (instant, fragile, not sanctioned and a restriction risk). See Why ReplyKit uses the official platform.
- Knowledge and retrieval. Crawling, chunking, indexing, searching, and keeping it current when a price changes. See how retrieval works.
- Prompting and guardrails. Scope, tone, refusals for regulated advice, the not-sure rule, disclosure, loop guards.
- Handover and inbox. Flagging, takeover by a person, hand-back, notes, search, several people at once.
- Booking. Slots, durations, staff, reminders, reschedules, waitlists.
- Safety limits. The 24-hour window, template approval, quality rating, STOP handling, spam rules.
- Operations. Logging, backups, uptime, alerts, crash recovery, cost monitoring.
- Maintenance. Model updates, WhatsApp library changes, policy changes such as Meta's 2026 AI rules, security patches.
What it costs to run
| Item | Build | Buy (ReplyKit) |
|---|---|---|
| Model tokens | Cents per conversation, variable | Included |
| Hosting and storage | Tens of dollars a month | Included |
| WhatsApp fees | Meta's per-message fee on templates | Meta's per-message fee on templates; replies in the 24-hour window free |
| Development | Weeks initially | None |
| Maintenance | Hours every month, forever | None |
| Monthly total | Low in cash, high in time | $29 to $99 |
The token cost is real and small; ReplyKit's measured cost per reply is a fraction of a cent, which is why plans can include thousands. The time cost is the one that matters for anyone whose business is not software.
When building wins
- The agent must act inside your own systems: order databases, ERPs, custom workflows no product supports.
- You are a software company and the agent is part of your product.
- Volume is very large and per-reply economics justify engineering time.
- You have specific compliance requirements on hosting and data residency.
In those cases, use the official API through a provider such as Twilio; see ReplyKit vs Twilio.
When buying wins
- The job is answering customers and booking them.
- Nobody on the team wants to be on call for a WhatsApp library breaking at 2am.
- You want the inbox, CRM, booking and website assistant without building them.
- You want someone else to track policy changes and model updates.
Where the connection matters: ReplyKit runs on Meta's official WhatsApp Business Platform, so there is no linked-device pairing and no restriction risk from automation. Meta's per-message fees apply only to reminder and follow-up templates sent outside the 24-hour window; replies to customers who message you carry no Meta fee inside it. The trade-offs are explained in why ReplyKit runs on the official platform.
ReplyKit's plans are $29, $49 and $99 with AI included and a 7-day trial; details in what it costs. If you build, our mistakes to avoid list is a decent checklist.
Frequently asked questions
Can I build a WhatsApp AI agent myself with an LLM API?
Yes, if you are technical. The model call is easy; the WhatsApp connection, retrieval, safety, inbox, booking and maintenance are the real work.
How much does it cost to run a self-built agent?
Model tokens and hosting are small. Development and ongoing maintenance time are the real cost.
When should I build rather than buy?
When the agent must live inside your own systems, you are a software business, or volume and compliance justify engineering.
What does a bought product provide?
The whole list: connection, retrieval, guardrails, inbox, booking, limits, operations and maintenance, for a monthly fee.
Which WhatsApp route should a build use?
The official API for a sanctioned, durable connection, accepting fees and setup; linked-device libraries are fragile and not sanctioned.
The whole list, for $29
Connection, retrieval, guardrails, inbox, booking and maintenance, done. 7-day free trial.
Sources and further reading
- Meta, WhatsApp Business Platform pricing · per-message pricing for the official API, for comparison
- WhatsApp Terms of Service
- Meta for Developers, Policy for AI Providers on the WhatsApp Business Platform
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.