AI is Not Responding or Has Reduced Capability

Summary

Torqueflow is designed so that AI never fails hard. If the system cannot reach its full AI capability - because of a webhook problem, an outage, or an exhausted credit balance - it drops to a simpler mode rather than stopping. This article explains the degraded modes you might see, how to tell which one is active, and what to do about it.

Prerequisites

• You are signed in as an owner, service advisor, or technician.
• To view voice call logs you need the voice.calls.view capability.

Steps

Understand the fallback modes

Torqueflow has three layers of fallback. Which one you see depends on what is wrong:

Mode	When it triggers	What customers experience
Full capability	All systems healthy, credits available.	AI answers, knows the caller, reads the live job, talks through when you have room, takes what the caller needs and escalates to staff. It does not write the appointment itself.
Message-taker (Level 3)	Credits exhausted. This applies to trial accounts too: a trial runs on its trial credit balance.	AI still answers in your garage voice, but only takes a name, number, and reason for calling. No availability, no status lookups, no tool calls.
Static fallback	The assistant-request webhook is unreachable or times out (more than 7.5 seconds).	A generic message reads: “Thank you for calling. I’m unable to access the full system right now. Can I take your name and number and have someone call you back?”

For WhatsApp and SMS, the equivalent of message-taker mode is an automatic canned reply followed by routing to the human inbox.

Identify which mode is active

1. Voice calls - go to Voice AI > Call Logs and open a recent call. The metadata panel shows which config was used. Calls handled by the static fallback are flagged with isFallback: true and appear with a warning icon. Calls handled in message-taker mode show a “Level 3” tag and the reason the call was degraded (credit exhaustion).
2. WhatsApp and SMS - go to Messages and open the conversation. A degraded AI reply shows as a system message: “Hi! We’ve received your message and our team will get back to you shortly.” The conversation is marked for staff attention and routed to the inbox automatically.
3. Credit balance - go to Settings > Billing and check the credit balance card. Orange or empty means you are in credit-exhaustion degradation.

Resolve credit-exhaustion degradation

1. Go to Settings > Billing.
2. Scroll to the Credit Packs section.
3. Buy a credit pack, or enable auto top-up so it never happens again.
4. Within a few seconds of the purchase, the balance cache refreshes and AI capability is restored on all channels automatically. No restart, no reload, no support ticket.
5. Voice calls that were already in progress when credits ran out complete normally - degradation only applies to the next interaction.

See ai-credits/what-happens-when-credits-run-out for the full explanation of credit exhaustion behaviour.

Resolve credit exhaustion on a trial account

1. A trial starts with a trial credit allowance. When it runs out, the AI drops to message-taker mode, the same as a paid account.
2. Credit packs cannot be bought during a trial. To restore full capability, start your subscription from Settings > Billing, or contact Torqueflow support.
3. See Your free trial for how trial credits work.

Resolve webhook / infrastructure degradation

If calls are answering with the static “I’m unable to access the full system” message, the assistant-request webhook is failing to respond in time. This is usually a short-lived infrastructure issue and resolves on its own.

1. Check Voice AI > Call Logs for recent calls flagged isFallback: true. A cluster of them in a few minutes indicates a platform issue.
2. Refresh the Voice AI > Dashboard page and check the channel status indicators.
3. If the issue persists for more than 5 minutes, contact Torqueflow support with the time of the first affected call.
4. Customers who hit the static fallback leave their name and number on the recording. Check the call log, pick up the messages, and call them back manually.

Understand what callers hear when a single step fails

A call can be at full capability and still hit one step that does not work, such as a lookup that times out or a request that will not save. The AI does not describe the fault to the caller. What it says depends on what failed:

What happened	What the caller hears
A lookup is taking a few seconds	A short holding line while it waits.
A lookup failed	That the AI cannot pull that up just now. It offers to take their details or help with something else.
A request could not be passed to your team	Plainly that it cannot pass this on right now, and a request to ring the garage back.
It is not certain whether something saved	That it cannot confirm it, and a request to ring the garage back to check. It does not say it worked, and it does not say it failed.

None of these promise that someone will ring the caller. When the AI asks a caller to ring back, it is because nothing reliable reached your team.

Known issue: when the system is under heavy strain and a call drops to taking a voicemail, a caller can occasionally be told their request was passed on when it did not arrive. If a customer chases a callback you have no record of, find their call under Voice AI > Call Logs and ring them back.

Check if WhatsApp is replying but the 24-hour window has expired

1. WhatsApp can only send free-form messages within 24 hours of the last customer message (Meta policy).
2. If credits are exhausted and the 24-hour window has closed, Torqueflow does not send the canned reply (Meta would reject it). The message is routed silently to the inbox for staff to handle with a template or wait for the customer to message again.
3. This is not a bug - it is required by Meta’s WhatsApp Business rules.

See whatsapp-ai/whatsapp-24-hour-session-window for how the window works.

Expected Outcome

• You can tell the difference between full capability, message-taker degradation, and the static webhook fallback.
• You know which button to press to restore full capability (buy credits, start your subscription, or contact support).
• Customer-facing conversations never fail silently - every channel keeps responding in some form.

Troubleshooting

Problem: Voice AI sounds like a completely different voice than usual. Cause: The static fallback assistant is answering because the webhook timed out. It does not use your configured persona. Fix: This is a temporary infrastructure issue. Check call logs for isFallback: true. If it clears within a few minutes, no action needed. If it persists, contact support.

Problem: Voice AI is answering but not discussing availability or checking work orders. Cause: You are in message-taker (Level 3) mode, most likely because your credits have run out. Fix: Buy a credit pack from Settings > Billing. On a trial, start your subscription or contact support, because credit packs are not available during a trial. Full capability restores automatically on the next call.

Problem: WhatsApp replies stopped - customer gets no AI response. Cause: Either credits are exhausted and the 24-hour window has closed (silent routing), or AI is disabled for WhatsApp in communication settings. Fix: Check Settings > Billing for credit balance. Check Settings > Communication to confirm AI is enabled for WhatsApp. The conversation is in the inbox - reply manually or with a template.

Problem: SMS replies are coming back with “our team will get back to you shortly” instead of real answers. Cause: Credit-exhaustion degradation. The AI is intentionally not making AI service calls. Fix: Buy credits or upgrade. Conversation is already routed to the inbox - staff can reply directly.

Problem: Customers on WhatsApp or SMS are receiving “we’ve got it and a member of our team will be in touch” instead of AI answers. Cause: Your organisation has reached its daily or monthly AI query limit. This is the limit set under Settings > AI Settings, separate from credit exhaustion. Fix: Check the usage dashboard on Settings > AI Settings. Raise the limit if appropriate. Until the limit resets, reply to waiting conversations from the inbox - the holding message has promised customers a human follow-up.

Problem: A customer says the AI asked them to ring back because it could not pass their message on. Cause: One step failed during an otherwise normal call, so the request never reached your escalation queue. The AI tells the caller that plainly instead of describing the fault. Fix: Find the call under Voice AI > Call Logs and ring the customer. If it happens on several calls in a short spell, contact Torqueflow support with the call times.

Problem: I topped up credits but AI still seems degraded. Cause: Balance cache takes a few seconds to refresh. A call already in progress will complete in degraded mode. Fix: Wait 15 seconds and make a test call. If still degraded after a minute, contact support.

Problem: A trial account is in Level 3 and I cannot buy credits. Cause: The trial credit allowance has run out, and credit packs are not sold during a trial. Fix: Start your subscription from Settings > Billing, or contact Torqueflow support.

Notes

• Torqueflow’s billing layer never blocks AI. If a credit check throws an unexpected error, the system defaults to “allow the interaction” - you will not see a black hole.
• In-flight grace: credit exhaustion during an active call or conversation never cuts it off mid-sentence. The current interaction finishes at full capability; the next one starts in degraded mode.
• Latency-based degradation (slow AI service) and credit-based degradation are separate systems. Credit degradation takes priority when both would apply.
• Demo orgs are always treated as having unlimited credits and never degrade.

Permissions

• voice.calls.view - required to see call logs and identify which mode a call was handled in.
• billing.credits.purchase - required to buy credits to resolve credit-exhaustion degradation (owners only by default).
• settings.communication.manage - required to check WhatsApp/SMS AI toggles.