What happens to your chatbot when an AI provider has an outage, and how automatic failover keeps website, WhatsApp and Messenger replies flowing.
Your support bot depends on an AI model that someone else runs. When that provider has a bad hour, your website chat stops answering, WhatsApp messages sit unread and Messenger customers wonder whether anyone is there. This article explains what provider failover is, why it matters for a support team, and how it works in AI SNS CHAT, including the limits you should know about.
Why a single AI provider is a single point of failure
Most chatbots are wired to one model from one vendor. That works until the vendor has an outage, slows down, retires a model name or limits your traffic. From the customer's side, none of that is visible. They just see a widget that spins, a WhatsApp thread that stays quiet, or a Messenger reply that never comes.
The damage is rarely dramatic. It is usually small and cumulative:
- A customer asking about delivery on a busy evening gets no answer and leaves.
- Your human team, who assumed the bot was handling routine questions, gets a sudden pile of messages.
- Someone has to notice the problem, work out whether it is your setup or the vendor, and decide what to do.
Support is also a channel where silence is read as rudeness. A slow reply is forgiven. No reply is not.
What failover actually means
Failover is a simple idea. If the first choice cannot answer, the system tries another one, automatically, without waiting for a person to change a setting.
For AI replies, that means the conversation continues even when the provider you picked is unavailable. The customer still gets an answer from the same agent, with the same persona, the same knowledge base and the same guardrails. The only thing that changes is which model produces the text behind the scenes.
It helps to separate three things that are often confused:
- Redundancy: having more than one provider available at all.
- Failover: switching to the next provider when the current one cannot respond.
- Routing by preference: choosing a model for quality, cost or tone, which is a different decision made ahead of time.
Failover is the safety net under the other two. You still choose the model you want. Failover only matters when that choice is not available.
The order of fallback matters
Not all failover is equal. A good setup is predictable, so you can reason about it before something goes wrong. Questions worth asking of any platform:
- Does it try your preferred provider first, or does it shuffle?
- Is the fallback order fixed and documented, or hidden?
- Can you keep replies on one vendor if your policies require it?
- Does it apply to every plan, or only some?
Predictability is what lets a support lead write an honest answer to a security questionnaire or a customer who asks, "Where does my message go?"
- 01Customer message arrives
- 02Agent's own provider tried first
- 03Next provider in the chain if unavailable
- 04Reply sent in the same conversation
Same persona, different engine: what stays consistent
A common worry is that a switch of model will make the bot sound like a different person halfway through the day. It is a fair concern, and it is worth being honest about it. Different models phrase things differently, so a fallback reply can read slightly differently from your usual one.
What failover does not change is everything you configured around the model:
- The agent's persona and guardrails.
- The knowledge base the agent answers from.
- The actions and integrations attached to the agent.
- The channel the customer is on.
So the bot's job description stays the same even if the engine changes. You can reduce the visible difference further by writing the persona clearly, with explicit instructions on tone, length and what never to say. A well-written persona does more to keep replies consistent than the choice of model does.
Test the alternatives before you need them
The worst time to discover that a fallback model handles your product questions badly is during an outage. Run a few of your real customer questions through different models in advance and compare the answers. If one model consistently struggles with your refund policy or your sizing guide, you want to know that on a quiet Tuesday, not during a sale.
Keeping everything on one vendor, when you must
Some organisations have a rule that customer messages may only be processed by one AI vendor. Others simply want the least surprising behaviour. For them, automatic switching to another vendor is the opposite of what they want.
This is a legitimate trade-off:
| Approach | Strength | Cost |
|---|---|---|
| Fall back across vendors | Replies keep flowing when one vendor is down | Your messages may be processed by more than one vendor |
| Stay on one vendor | One data path, one set of vendor terms to review | A vendor outage can pause AI replies |
Neither is correct for everyone. A small online shop answering delivery questions on WhatsApp probably values continuity. A company with strict vendor-approval rules may prefer to stay on one vendor and rely on a human handoff when the AI is unavailable. Decide deliberately, write the decision down, and make sure whoever answers vendor questions knows which one you chose.
What failover cannot fix
Failover is not magic, and it is better to be clear about the limits than to oversell it.
- It does not fix a weak knowledge base. If the agent has no good answer in its sources, a different model will not invent a correct one. Fix the content.
- It does not replace human handoff. If every provider were unavailable at once, or a customer needs a person, you still want a clear route to your team.
- It does not cover every feature equally. Some features depend on a specific vendor. Real-time speech-to-speech voice, for example, is tied to a single provider in AI SNS CHAT and has no fallback. Text chat is where the multi-provider chain applies.
- It does not remove the need to watch your usage. Credits are still consumed per reply, and you should keep an eye on them.
Treat failover as one layer in a stack: good content, a clear persona, sensible guardrails, a human escalation path and then provider redundancy.
A hypothetical to make it concrete
Imagine a shop that receives 300 chats a month across its website, WhatsApp and Messenger, and that most arrive in the evening after work. If its AI provider is unavailable for an evening, a meaningful share of that month's conversations could land in that window. With failover, the same questions are answered by the next provider in the chain. Without it, they queue up for the morning. The numbers here are only an illustration, but the shape of the risk is real: outages do not respect your quiet hours.
A short checklist for your own setup
Before you rely on any failover feature, run through this list:
- Pick a primary model on purpose. Choose it for answer quality on your real questions.
- Decide your vendor policy. Continuity across vendors, or one vendor only.
- Check your plan. Some model features are gated by plan.
- Compare models on real questions. Use a test environment, not live customers.
- Write a strong persona. It keeps tone steady across models.
- Set up human handoff. Make sure the right person is alerted and knows what to do.
- Review conversations after any incident. Look at what customers saw and adjust.
Doing this in AI SNS CHAT
AI SNS CHAT supports four text providers: Anthropic (Claude), OpenAI, Google (Gemini) and OpenRouter. Each agent has its own provider and model, and failover is built on top of that choice.
How the chain works
The agent's own provider is always tried first. If that provider cannot answer, AI SNS CHAT moves through the rest of the chain in this order: Anthropic, then OpenAI, then Google, then OpenRouter. So if your agent is set to Google, Google goes first, and the others follow in that fixed sequence.
Grok (xAI) is voice-only. It is never used for text chat, and any stored Grok text model is rewritten to Claude Sonnet, so it will not appear as a fallback for website, WhatsApp or Messenger replies.
Choosing your model
Open the agent, then go to Settings → AI. This is where you pick the provider and model for that agent. AI SNS CHAT discovers available models live from each provider's models endpoint and caches the list for 6 hours, so new models appear without waiting for a product update, though a very recent release may take a few hours to show up.
Plan gates
Advanced models and model fallback need the Hobby plan or higher. On a lower plan, an agent falls back to the standard model of the same provider rather than moving across the chain. If continuity matters to you, check your plan before you depend on it. Current plans and prices are on the pricing page.
Keeping replies on one vendor
A platform administrator can set the fallback mode to "same", which keeps every reply on a single vendor. This is a platform-level setting rather than something each customer toggles on a dashboard page, so if your organisation needs it, raise it with your account contact or platform admin.
Comparing models before you commit
Use the Playground to test your agent, and the Playground A/B model compare page to put two models side by side on the same question. Run your hardest real questions through it: refund edge cases, product comparisons, anything with numbers. Pick the primary model you trust, and look at how the alternatives answer too, since those are the ones that might step in.
Around the failover
- Human handoff: under Helpdesk → Scheduling, set shifts per person and timezone. Handoffs are auto-assigned to whoever is on shift, with alerts to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. Helpdesk needs the Standard plan or higher.
- Restore points: before you change a model or persona, you can roll back using restore points (up to 40 per agent).
- Usage: each reply uses 1 credit, and the Usage page shows credits per agent.
- Channels: the same agent serves your chat widget, WhatsApp, Messenger and the other channels you deploy to. See all channels and the feature overview.
Failover applies to text replies. Synthia speech-to-speech uses the OpenAI Realtime API only, with no fallback.
Conclusion
Provider failover is not glamorous, but it decides whether a quiet outage stays quiet. Choose your primary model on purpose, decide whether continuity or a single vendor matters more to your business, test the alternatives on real questions, and keep a human route open. To start building, create an account and open Settings → AI on your first agent.
Frequently asked questions
Which provider is tried first when an AI SNS CHAT agent replies?
The agent's own provider is always tried first, the one you set under Agent → Settings → AI. If it cannot answer, the chain continues in a fixed order: Anthropic, OpenAI, Google, then OpenRouter.
Is failover available on every plan?
No. Advanced models and model fallback need Hobby or higher. On a lower plan, an agent falls back to the standard model of the same provider instead of moving to another vendor.
Can I stop my messages from being processed by a second vendor?
Yes, through a platform-level setting that keeps every reply on one vendor. It is controlled by a platform admin rather than a customer dashboard toggle, so ask your admin or account contact if your policies require it. The trade-off is that a vendor outage can then pause AI replies, so plan a human handoff.
Does failover cover voice calls too?
Not in the same way. Failover applies to text replies. Synthia speech-to-speech uses the OpenAI Realtime API only, with no fallback, and Grok is used for voice only, never for text chat.
How do I check that my agent still answers well on a different model?
Use the Playground and the A/B model compare page to run the same real questions through two models side by side. Do this before an incident, then refine the persona and knowledge base if one model handles your content poorly.


