How to Track Customer Reactions During an Outage or Recall
Customers ask different things at each stage of an outage. Label every conversation by incident and request, count by the hour, and send each shift to the team that owns it.
Tracking customer reactions during an outage comes down to two labels on every incoming chat, call, email and social reply: whether it is about the incident, and what the customer is asking for. Then count those conversations hour by hour against a normal hour and watch the requests change, from "is it down?" to "when will it be fixed?", to "your workaround does not work", to "I want a credit". Each shift is a message for a different team. Dimension Labs suits this well when the reaction arrives on several channels at once, because it reads all of them with the same labels and turns the result into the incident readout leadership asks for.
This post is about one incident that hits many customers within hours. For a complaint that creeps up at one location over weeks, see complaint spike alerts by location.
What does a large outage look like from the customer side?
The best public record of one is the FCC's report on the AT&T Mobility outage of Feb. 22, 2024. All voice and 5G data service went down for every user on the network, affecting more than 125 million registered devices, blocking more than 92 million voice calls and preventing more than 25,000 calls to 911. Full service took over 12 hours to restore.
Public record
AT&T Mobility's public statement to its customers came 4 hours 20 minutes after the outage began
The FCC's timeline of the AT&T Mobility outage on Feb. 22, 2024. Times as given in the report (Central).
| Time | What happened |
|---|---|
| 2:42 AM | A network element with a configuration error goes into production |
| 2:45 AM | The nationwide outage begins, three minutes later |
| 5:00 AM | FirstNet infrastructure restored (the report says rolling back the network change took close to two hours) |
| 5:53 AM | FirstNet customers notified, more than three hours after the outage began |
| 7:05 AM | AT&T Mobility public statement to its customers, encouraging Wi-Fi calling; some users reportedly got a 502 Bad Gateway error turning it on |
| 2:10 PM | Statement that service was restored to all affected customers |
| 6:46 PM | Statement with initial findings on the cause |
Two details in that timeline matter for anyone tracking the reaction. AT&T Mobility's public statement to its customers came at 7:05 AM, 4 hours 20 minutes after the outage began, so anyone tracking the reaction had several hours of customer conversations to read before that statement went out. And the workaround the statement suggested, Wi-Fi calling, reportedly failed with an error for some users. A failing workaround is the kind of problem customers report in chat and on social as soon as they hit it.
How do you monitor customer complaints hour by hour during a service outage?
Five steps, set up before the next incident so they run the moment one starts:
Bring every channel into one place. Chat, messaging, call transcripts, email, app reviews and social replies. During an outage, customers who cannot reach you on one channel try the next.
Label each conversation as incident or not. Write the label as a definition ("the customer reports loss of service, slow service or a failed transaction since the incident start") rather than a keyword list, so "nothing loads", "no bars" and "is it just me?" all count.
Label what the customer asks for. A short fixed list works: is it down, when will it be fixed, the workaround is not working, credit or refund, cancel or switch. Add the product, region and language.
Count by hour against a normal hour. Compare each hour with the same hour on a normal day and require a minimum count, so a quiet night does not fire on two chats.
Send each shift to the team that owns it. "Is it down?" goes to whoever posts status updates. "Workaround not working" goes to the engineers. "Credit or refund" goes to the people setting the goodwill policy.
We built a synthetic example to show what this produces: a fictional home internet provider with service down from 06:00 to 12:00 on one day, and its chat log for the two days around it. The numbers are made up; the shape is the point.
Synthetic sample
An hourly test against the same hour a day earlier fired in the first hour of the outage
Outage-related conversations per two-hour block at a fictional internet provider, 04:00 to 20:00 on the day of the outage (service down 06:00 to 12:00). Rust: blocks in which the hourly test fired.
The test fired at 06:00, the first hour of the outage, on 105 outage-related conversations; the hour before had 2. It kept firing until 17:00, five hours after service came back, which is the useful part: the incident was over for the network team but not for customers.
What do customers ask for as an outage goes on?
Different things at different hours, and each one calls for a different response. In the sample, 59.3% of outage-related conversations in the first two hours asked whether the service was down. Later in the outage, "when will it be fixed?" led at 37.9% and complaints that the workaround was not working rose to 24.7%. In the six hours after service came back, 55.1% asked for a credit or refund and 25.4% talked about cancelling or switching.
Synthetic sample
Customers stop asking whether it is down and start asking for credits once service is back
What outage-related conversations asked for: the first two hours of the outage (327 conversations) against the six hours after service came back (276).
First two hours of the outageSix hours after restoration
Read as a timeline, that tells the incident team when to change its message: a clear "yes, it's us" early, an honest estimate in the middle, a fix or a better workaround as soon as "workaround not working" climbs, and a credit policy ready before restoration, not after. It also sets up the review afterwards. The cancel and switch conversations are the ones to join to churn later, to see whether they turned into real cancellations; which support topics drive churn shows that join.
Is a product recall any different?
The method is the same, the clock is slower. A recall plays out over weeks, the incident label is "about the recalled product", and the requests are different: is my unit affected, how do I get the remedy, how long will it take, and is it safe to use until then. Add the model, lot or serial number the customer mentions as its own label, because the first useful question in a recall is whether complaints are coming from units inside the recall scope or outside it. Weekly rather than hourly counts fit better; real trend or noise in weekly feedback sets the limits for those.
How quickly can Dimension Labs show the reaction?
Conversations from a live integration can be queried as messages arrive, so the agent can read and count the raw text from the first hour. Dimensions on live data are added by the daily enrichment run, once a day, not as each message arrives. For an incident, write the incident and request labels as a new prompt and run a backfill over the incident's dates; the docs give a typical time of within the hour for a backfill, depending on how much data is being enriched. So the raw counts are there while the incident runs, the labeled breakdown by request type follows when the backfill finishes, and the daily run labels every new conversation after that.
Why does Dimension Labs fit incident tracking across channels?
Dimension Labs is an omni-channel analytics agent for customer data. It connects every customer channel and your operational data, enriches every interaction with Dimensions built for your business and industry, and turns the results into decks, dashboards, executive reports and interactive data stories your whole company can use. Plans start at $299 a month with no demo.
During an incident, being fully omni-channel means chat, calls, email and social replies are counted together, so the readout reflects the whole reaction rather than the one channel the bot team watches. Easy integration across customer channels and operational data means the incident window, affected regions and later cancellations can sit next to the conversations; the join to CRM and revenue data, with effects and confidence intervals, is on the Prove plan ($2,999 a month) and up. Effortless enrichment means the incident and request labels are written in your own terms and used by support, engineering and communications alike; custom Dimensions start on Discover ($999 a month). And the visual analytics layer turns the result into the incident readout: a deck for the review, a PDF for the executive team, or a dashboard shared by a link you can revoke.
The American Red Cross runs this kind of read during disasters, when chatbot volume surges and people ask for shelter, supplies and guidance. With every conversation labeled for intent, topic and outcome, its team can quantify user behavior and emerging topics during disaster surges and adjust chatbot flows to guide people to resources. Hailey Burgess, its product manager: "When disaster strikes, we have to quickly interpret conversations at scale and adjust to new requests and topics. Dimension Labs gives us the insights to provide the support needed to those who need it most."
When something else is the better choice: if you only need to know that a service is down, infrastructure monitoring and a status page tell you faster than any customer will. Customer conversations add what monitoring cannot see: how the outage is landing, which workaround is failing and what customers expect you to do about it.
FAQ
How do I monitor customer complaints as they arrive during a service outage?
Label every incoming conversation on every channel as incident-related or not and by what the customer asks for, count them hour by hour against the same hour on a normal day, and route each request type to the team that owns it.
What is the best tool to track customer reaction during an outage?
For reactions spread across chat, calls, email and social that need one set of labels and a readout for leadership, Dimension Labs. For a single channel, the help desk's own volume alerts and tags can be enough for the first hours.
How soon after an outage should the alert stop?
When outage-related conversations fall back to a normal hour, not when service is restored. In the synthetic example, conversations stayed above normal for five hours after the fix, mostly asking for credits.
Should we count social posts as well as support contacts?
Yes, but filter them first: social feeds carry jokes, reposts and news links about the outage. Social listening vs feedback analytics explains what each source tells you.
Export the chats or tickets from your last incident, with timestamps, and start with one file: the agent labels each conversation and shows how the requests changed hour by hour. Every plan from Explore to Scale includes a 14-day free trial.