Back to Blog
ProductNovember 17, 2026 5 min read

When an AI Provider Goes Down, What Do Your Customers Actually See?

ReliabilityProductInfrastructure
Dariu Dumitru
Authored by Dariu Dumitru, Co-Founder & CMO
Published Nov 17, 2026.
When an AI Provider Goes Down, What Do Your Customers Actually See?

It's a busy Tuesday afternoon. A customer calls your business to ask whether their order shipped, a visitor on your site asks about weekend hours, and someone texts to move an appointment. At that moment, the AI provider behind your customer service starts failing. Every major AI provider has bad hours, and their status pages show it. The question isn't whether it will happen. It's what those three customers experience when it does, and whether you'll even know. Let's walk through it channel by channel.

On the phone. With a tool built on a single provider, a caller dials your number and gets one of three things: dead air, a looping "sorry, I'm having trouble," or a dropped call. Callers are the least patient customers you have. They don't wait for a status update, they hang up, and a meaningful share call the next business on the list. You find out, if you find out at all, when someone mentions it days later. With a system that routes each step across vendors, the step that failed is handed to a model from another provider and the call continues. The caller hears an answer, perhaps a moment slower than usual, and hangs up with their question answered.

In web chat. A single-vendor chat usually fails politely, which is almost worse. The typing indicator spins, or a generic error appears, and the visitor, who came to your site because they were ready to buy, book or ask, closes the tab. Chat visitors rarely complain. They just leave, so the cost never shows up anywhere you'd look. With cross-vendor routing, the visitor gets their answer and never sees the problem.

By text and WhatsApp. Messaging fails silently, and that's what makes it dangerous. A customer texts your business a question and gets nothing back. There's no error message on their end, just a thread with no reply. They can't tell whether you're busy, closed, or ignoring them, and many will assume the last. The failure is invisible to them and invisible to you. When the system can switch vendors, the reply simply arrives.

By email. Email is the most forgiving channel, since nobody expects a reply in seconds. But a single-vendor outage can still leave a backlog of unanswered messages that quietly ages through the afternoon. With cross-vendor routing, email is answered on the same schedule as any other day.

Put those four pictures side by side and the business case is plain. An outage on a single-vendor tool doesn't look like an outage to your customers. It looks like your business stopped answering. They don't blame the AI provider, whose name they've never heard. They blame you.

This is the reason we route across vendors in the first place. Our full pipeline recovers from a provider outage, including an OpenAI outage: when the model handling a step fails, a model from another lab takes over that step and the conversation carries on. We wrote about the model side of this, routing each job to the best model and switching when one fails, and the architecture is described on our page on vendor-agnostic AI customer service. This post is about the other side, what it means for the person running the business.

For an owner, it comes down to continuity. You adopted AI so that customers get answered when you and your staff can't answer them yourselves: after hours, during the rush, while everyone's with another customer. If that coverage disappears whenever one company upstream has a bad afternoon, you haven't built coverage, you've added a single point of failure to the part of your business customers see first.

It's also worth being clear-eyed about what cross-vendor routing covers. It removes the most common failure, one AI provider going down. Your own internet connection, your phone carrier, or a problem with your website are separate links in the chain, and they deserve their own plan. Good continuity means knowing what happens at every link, not just the AI one.

So when you evaluate any AI tool for customer service, including ours, these are the continuity questions worth asking.

"Show me what my customer sees on each channel if your system can't answer." Not the architecture diagram, the customer's screen and the caller's ear. A thoughtful vendor will have an answer for phone, chat, text and email separately. A vague one hasn't thought about it.

"Does your failover keep the whole conversation going, or just the reply?" Understanding the question, writing the answer and checking it are separate steps. If only the reply can switch providers, an outage anywhere else still stops the conversation.

"Where does the conversation go when a person is needed?" Within the hours live support is set to cover, a good system puts a caller through to a live person, can connect a website visitor to a signed-in agent by audio or video, and points text and email customers to live support. If urgent matters must reach someone at night, ask whether live support can be set to cover those hours too. Know this path before you need it.

"Will I know it happened?" You should be able to see every conversation afterward, including the ones on the day something went wrong, and search, filter and review them. If you can't see what your customers experienced, you can't judge your continuity plan.

The best time to ask these questions is before your busiest day, not during it. Outages don't wait for slow periods, and the afternoon one lands on is, by bad luck, often the one that matters most. Choosing a system that can route around a failing provider is how you make sure that afternoon looks, to your customers, like any other.

The bigger picture

Why the next wave of customer service answers accurately and acts reliably

The same guesswork that writes a fluent answer should never be what executes your transactions. See how grounded answers and a deterministic engine change what AI customer service can be trusted to do, across 200,000+ flows in production, zero errors.

Read: The Third Wave of customer service