When the first-choice AI model becomes slow or temporarily unavailable, your Hey Button teammate can fall back to another capable model, so work continues. The switch occurs behind the scenes, and your configuration remains exactly as you established it during setup.
Why do AI models slow down or go offline at all?
AI models operate on large computing infrastructure maintained by the companies that develop and operate them. Those systems process an enormous volume of requests simultaneously, and that concentrated demand is the root of the difficulty. Consequently, they can become congested, respond more slowly than anticipated, or stop responding entirely for a period.
A busy afternoon can lengthen a reply considerably, and a brief outage can suspend all responses until normal service is restored. This behavior is an ordinary characteristic of online services, and it does not indicate that a teammate was poorly engineered or carelessly assembled. Any small business that depends on an external service must anticipate the day that service becomes unresponsive. Experienced owners already practice this kind of contingency planning instinctively. They keep a paper pad beside the register in case the card reader fails, and they know whom to contact when the internet drops.
The same logic applies to an AI teammate, and it deserves to be stated plainly. The meaningful question is therefore not whether a model will ever slow down, because every model eventually will. The meaningful question is how your teammate behaves during a slow period. Does outstanding work wait in an orderly manner, or does it proceed safely under the rules you approved in advance? A well-designed arrangement answers that question before any disruption occurs.
What does a fallback actually mean for your AI teammate?
A fallback is a deliberate backup arrangement that anticipates operational interruptions before they can affect a customer. When the first-choice model cannot respond with adequate speed, the system can route the same task to another capable model. That second model can complete the work because it draws on the same context, including the instructions and facts your teammate received during setup.
Hey Button runs on OpenClaw, an open automation framework, and models such as Claude and GPT answer through that framework. Hey Button does not operate a proprietary artificial intelligence model built from scratch, and it does not represent otherwise to any prospective client. A fallback is therefore a selection among existing capable models that the framework can use. It is not a new intelligence invented for your business or your particular industry.
Consider a substitute driver who receives the identical route sheet as the regular driver. The trip follows the same directions and the same stops, even though a different person sits behind the wheel for the day. The destination, the delivery schedule, and the rules of the road all remain unchanged throughout the journey. That principle describes the fallback arrangement accurately, and it is the most useful way to picture it.
Model routing is the broader concept underneath this behavior, and it deserves separate attention. Hey Button's model routing aims to select the least expensive model that still performs a given task correctly and dependably. A fallback becomes relevant only when the first choice is slow or unavailable at a particular moment. In that situation, the system transfers the work to another capable model rather than abandoning the task or leaving a customer waiting indefinitely.
Does switching models change what your teammate is allowed to do?
No, and this point deserves considerable emphasis. Access is determined by you, not by whichever model happens to be answering a particular request at a particular moment. Hey Button offers three access levels, read-only, draft-and-approve, and full access, and those levels remain fully in force regardless of which model responds. A backup model therefore inherits precisely the same boundaries as the model it temporarily replaces.
These boundaries matter most at the edges of the arrangement, where mistakes would become visible to customers. If your teammate is set to draft-and-approve, the backup model still prepares a draft and waits for your approval before anything leaves the system. If your teammate is configured as read-only, the backup model still reads information and takes no action on its own initiative. In short, a model switch leaves the access boundaries you defined exactly where they were.
Some owners worry that a slow day could become the moment a teammate quietly acquires additional freedom. That concern is understandable, but it does not describe how the arrangement actually functions. Access is a decision you make during setup, and you retain the authority to revise it whenever your circumstances change. A fallback does not expand that access in any respect, and it cannot do so independently.
What stays the same for your customers during a slow period?
Customers should notice very little during a slow period. In most cases, the only visible difference is that a reply takes somewhat longer to arrive than it normally would. The teammate keeps following the approval rules you selected, and any customer-facing matter needing professional judgment still comes to you first.
Customer messages, spending, and sensitive changes stay behind the approval rules you establish, whichever model is answering that afternoon. If a message would ordinarily await your review, it still awaits your review under a backup model. The approval step does not depend on the identity of the model that drafted the response, which is precisely why it remains dependable.
Consider an illustrative workflow, labeled as such for clarity. A homeowner in the Fort Wayne area sends a message requesting a quote for a spring cleanup. During a slow stretch, the first-choice model cannot respond quickly enough, so the system falls back to another capable model. That model organizes the request and identifies that the property address and several photographs are still missing. It then drafts a courteous reply requesting both, and you review that draft before anything reaches the homeowner.
The example illustrates the central principle of the entire arrangement. The backup model performed the job the first model would have performed, within the same limits and approval rules. The owner retained complete control over what the customer ultimately receives, and for a small business, that outcome matters most of all.
How can you verify that a fallback happened?
You can verify a fallback in two practical ways. First, the client portal shows what your teammate is doing and why it is doing it. Whether a portal entry identifies the model behind a response is a detail to confirm with your setup contact before relying on it. Do not assume the answer, and inquire directly instead.
Second, review the work on a regular and predictable schedule. Select three or four recent tasks each week, and read each one from beginning to end without skimming. Look for drafts that sound unnatural, facts that have changed since setup, and replies that overlooked a detail the customer had provided. Record the date, the task, what the teammate did, and what you expected it to do. Patterns tend to emerge quickly in a log like that, even when each individual entry appears minor. Compare each draft against your own standards for tone, accuracy, and completeness. A short checklist helps: did it use the correct name, state only facts you approved, and request missing details before proposing any commitment? Keep a copy of any draft you edit heavily, since those edits show where the teammate needs more guidance.
Could a backup model give a weaker answer?
Yes, that possibility is real and deserves an honest acknowledgment. Different models possess different strengths, and a backup model may phrase its response differently or miss a nuance the first choice would catch. A customer who mentions a deadline in passing might receive a reply that ignores it entirely. If one request category draws weaker replies during slow periods, the review log will show it, and you can tighten those approval rules quickly.
For that reason, systematic review and clearly defined limits form part of the setup itself rather than an optional addition made later. Hey Button does not promise that every fallback response will match the first choice in quality or accuracy. AI systems can make mistakes, and a backup model can make them too. The genuine protection comes from approval rules, access levels, and your own careful review of real examples over an extended period.
A practical rule provides considerable protection in this situation. Anything a customer will rely upon should stay on draft-and-approve, including quoted prices, appointment times, warranty statements, and timing commitments. Maintain that rule until you have developed real confidence in the teammate's handling of that particular kind of request. You can widen access later, once the results have earned your confidence through consistent and verifiable performance over time.
Hey Button also does not guarantee uptime, and no honest provider could offer such a guarantee about a service it does not control. A fallback improves the odds that work continues when the first choice is slow or unavailable. It does not eliminate every outage, and it does not make any individual answer correct.
How should owners in Northeast Indiana prepare for slow days?
Owners in Northeast Indiana, including those in the Fort Wayne area, should treat slow days as a planning problem rather than a surprise. Many local businesses depend heavily on personal relationships, prompt responses, and teams that are already stretched thin throughout every working day. A customer who sends a message on a Saturday morning expects a sensible answer, and that customer also expects the answer to arrive promptly.
A landscaping company, a cleaning service, or a general contractor can apply this planning in a straightforward and manageable manner. Decide which messages must reach a person the same day, and decide which messages can reasonably wait for attention. Busy stretches provide the most realistic test of any arrangement. Weekends during spring cleanup season are one example, and storm recovery work for lawn and tree crews is another demanding case. Evenings after staff have gone home also generate sudden accumulations of messages. Those moments expose a slow model, so plan for them as you would for a crew member who calls in sick at peak time.
Local trust also depends heavily on accuracy and consistency. A teammate that states the wrong service area, schedule, or price can damage a relationship built over many years. That damage can exceed the harm of a modestly delayed reply, so keep approval rules strict while the teammate learns your business.
What should a small business do before go-live to prepare?
Work through these four steps before any customer message reaches the teammate.
- Inventory your work by urgency, listing the messages and tasks that cannot reasonably wait several hours without causing real difficulty for a customer.
- Assign a human backup to each urgent category, like a phone line reaching a real person or a shared inbox checked twice daily.
- Establish the approval rules with real precision, naming which replies may be sent automatically and which must always return to you for final judgment.
- Confirm the access level with your setup contact, and review real examples before anything touches a customer.
Hey Button's setup starts with an intake conversation about how work arrives, which tools you already use, and where repetitive busywork accumulates. Use that conversation to name the categories that always need your personal attention, such as complaints, refunds, pricing exceptions, and job acceptance. The backup plan in step two must still function when automation fails, because it does not depend on the teammate at all.
During go-live, you observe the teammate working through real examples before it interacts with any customer. Use that observation period to test a deliberately long and complicated request, a reasonable stand-in for a difficult slow day. Send the request, then read the draft closely, confirming that the approval rules held and that the teammate asked for every detail it needed.
If you would like to understand the setup process step by step, read how Hey Button sets up your AI teammate. When you are ready to talk through your own workflows, start here and describe how work comes into your business.
Frequently Asked Questions
- How long could a slow model delay a customer reply?
- No fixed delay can be promised, because it depends on how busy the provider is at that moment. Customers usually see a reply that arrives a little later, and nothing reaches them without your approval where approval is required.
- What should I check in the client portal after a slow day?
- Look at recent tasks, drafts waiting for approval, and any reply that changed since setup. Record anything unusual in a short written log so patterns become visible over time.
- Who should handle urgent messages if the teammate is slow?
- A person or telephone line should always stay reachable for urgent matters. The teammate supplements that plan rather than replacing it, so the backup works even when automation is unavailable.
- Should I keep my teammate on draft-and-approve during slow periods?
- For anything a customer will rely on, a cautious start makes sense. Draft-and-approve keeps you in the loop until you have checked the teammate's work on real examples and trust its replies on that kind of request.
- Which AI models power Hey Button?
- Hey Button runs on OpenClaw, an open automation framework, with models such as Claude and GPT answering through it. Hey Button does not offer a proprietary model built from scratch.
Hey Button
Want a teammate that handles this for you?
Answer a few questions about your business and Hey Button will show you what an AI teammate would take off your plate, with you approving anything that goes out.




