Buyer guide · Practical decisions

What happens to customer enquiries when your AI service is down?

Your enquiry form works, but the AI service that reads each message and drafts a reply has stopped responding. Does the customer get a reference, does anyone see the request, and can staff take over? Those questions matter more than whether a second model is available. Design the reduced service around what the customer still needs.

Published Sources checked

Read the guide ↓
01 /

Decide what must keep working

Start with a minimum useful service: receive the message, retain it, give the customer a reference and make it visible to the team. AI classification and draft writing can happen afterwards. A confirmed receipt must mean the request has actually been saved; it must not imply that someone has reviewed it or accepted a booking.

AWS describes graceful degradation as keeping a core function available when a dependency fails. Applied here, the enquiry inbox is the core function and the AI draft is an enhancement. This is a design recommendation, not a guarantee that your existing form or automation platform already separates them.

Ask your supplier to show where the request lives before the model is called. If saving depends on a successful AI response, that dependency needs changing. If storage itself fails, show a clear failure message and an established alternative contact route; do not display a success receipt.

02 /

Illustrative example: a hire desk during a busy morning

Imagine a UK equipment hire business using AI to classify web enquiries and prepare replies for staff. This is a fictional operating example, not a Fewzen customer result. The model becomes unavailable at 9 am. Intake still works, but AI drafts pause.

The team uses a standard acknowledgement saying the enquiry has been received and is awaiting review. It does not promise stock, a price or a collection slot. Staff can open the original message and claim it for manual handling. Customers needing a time-critical answer can use the desk’s normal telephone route during its stated opening hours.

A reduced service agreed before the outage
StepDuring the AI outage
CaptureSave the original request and receipt time; return its reference.
PrepareMark the AI draft as pending. Keep the original message readable.
RespondLet an assigned colleague reply through the normal workflow.
ResumeSkip requests already completed manually; check the rest before drafting.
03 /

Stop slow model calls blocking the rest of the service

Give the AI call a time limit and decide how many attempts are useful within the customer's waiting budget. A long timeout on every request can tie up the resources needed for intake. Microsoft's circuit breaker pattern temporarily stops calls after a failure threshold, then allows limited attempts to check recovery.

Ask the implementer to distinguish a temporary service failure from a bad request, revoked credentials or a safety refusal. Sending every failure to another model is not an appropriate recovery rule. The switch to reduced service should be visible to staff, with an alert to a named owner and a way to pause AI processing manually.

04 /

Work out how long catching up will take

A queue preserves waiting work, but it does not create processing capacity. Amazon's guidance on queue backlogs explains how accumulated work can prolong an outage's business impact. Monitor the age of the oldest waiting enquiry alongside the count; a modest queue can still contain a customer who has waited too long.

Suppose this fictional desk receives 60 enquiries an hour. A 30-minute pause creates 30 waiting requests. After recovery, processing capacity is 90 an hour while new arrivals continue at 60. Only 30 an hour remains for catching up, so clearing the backlog takes another hour. This simplified calculation assumes steady arrivals, successful processing and no manual completions.

Set a waiting-time threshold that triggers staff intervention. Agree who covers breaks and what happens outside opening hours. If people cannot keep pace either, update the customer message and manage priorities explicitly. Moving everything into a queue without an owner merely hides the delay.

05 /

Treat a second model as another route to qualify

A backup model may help, but it needs the same task checks as the primary one. Can it read the attachments, produce the required fields and handle your awkward examples? Check its permitted data handling, available capacity and access before sending customer enquiries to it. A shared failing gateway or database can disable both routes.

Microsoft's AI application guidance recommends separating layers and enforcing their policies independently. For this workflow, keep permissions and business checks outside the model choice. A backup must not gain authority to confirm availability or send an unchecked quote simply because the normal draft service is unavailable.

A standard acknowledgement and a staffed inbox may be a better first fallback for a small team. Multiple providers add configuration, evaluation and maintenance work. Choose them when the continuity benefit justifies that responsibility.

06 /

Resume without replying twice or reviving old requests

When service returns, start with a small amount of waiting work and watch actual completion. Avoid releasing the whole backlog at once. Record which enquiries were claimed, answered or cancelled while AI was paused. Check current status before generating or sending anything, including any delayed result from the original attempt.

Some requests will no longer be useful to process automatically: a customer may have phoned, changed the collection date or gone elsewhere. Route these for review. Keep the original reference throughout so staff can reconcile the customer conversation with the pending job. The related guides below cover duplicate actions and stale updates in more detail.

07 /

Rehearse one outage before expanding automation

Use test enquiries in a non-production environment. Make the model unavailable while leaving intake running, then restore it. Have the person who would cover the desk follow the handover instructions. Repeat with slow responses rather than an immediate error: the customer experience can be very different.

Check that every accepted test request remains visible, acknowledgements describe the actual state, staff can complete work manually, and recovery leaves no duplicate replies. Also test an intake-storage failure, which should not produce a false receipt. Record the oldest waiting request and the time taken to clear the backlog.

The useful next step is a one-page operating note: minimum service, trigger for switching, owner, customer wording and recovery checks. Use the rehearsal to decide whether the workflow is ready for a busier day.

Practical answers

Before you take the next step.

Does a provider status page tell us whether our workflow works?

It is useful context. Your own intake, credentials, integrations and processing queue still need monitoring. Measure whether customers' requests are being accepted and handled.

Should every AI failure trigger a manual task?

Group service-wide failures into an incident and keep individual enquiries in one visible work queue. Otherwise staff may receive hundreds of alerts while the same outage continues.

Work directly with Matthew

Bring the workflow and the point where it stops.

We can discuss the intake, handover and recovery steps needed for a dependable integration.

Explore integration support ↗Discuss your project with me ↗Prefer to write? Send a message. No sensitive records needed for an initial enquiry.