How Superhuman Mail Identifies Frustrated Customers and Emerging Bugs Before They Escalate
Jessica Jess
Content Strategist, Voice of Customer
Summary
With around 100 new support tickets arriving each day, Superhuman Mail's Delight team needed a faster way to identify frustrated customers, urgent escalations, and emerging product issues.
Using Enterpret automations, the team now starts every weekday morning with a prioritized view of the customers and issues that need attention most. Three automated reports surface high-frustration conversations, flag sensitive requests for escalation, and identify emerging bug patterns — helping the team focus on responding to customers instead of manually searching through tickets.
The challenge
Finding the most urgent customer issues in a growing support queue
With around 100 new tickets arriving each day, Superhuman Mail's Delight team faced a familiar challenge: the customers who needed attention most weren't always the easiest to identify.
Frustrated customers could get buried among routine support requests. Billing disputes and privacy-related requests required a different escalation path, but arrived in the same queue. And when multiple customers encountered the same product issue, the broader pattern wasn't always obvious from individual tickets.
The team needed a way to surface these signals sooner, prioritize the right conversations, and recognize emerging issues without manually reviewing every ticket.
The solution
Turning daily support conversations into actionable signals
Every weekday at 8 a.m. Mountain Time, Enterpret reviews Superhuman Mail's new and open support tickets from the previous 24 hours and delivers three automated reports directly to the team's Slack channel.
Customer Escalations Queue: Identifies and ranks customers showing the strongest frustration signals, regardless of account size. Each flagged conversation includes a brief explanation, while legal and privacy-related notices are surfaced for separate escalation.
Lightning Bug Report: Identifies emerging bug clusters when five or more customers report the same issue within 24 hours.
Mass Lightning Bug Report: Detects larger spikes of 20 or more reports of the same issue within an hour, triggering the team's incident response process.
Together, these automations help the Delight team understand which customers need attention first and which product issues may require a broader response.
Brianna McGrew manages the Superhuman Mail setup, working alongside Anthony Aerni at Enterpret and Lorilyn McCue, Superhuman's Head of AI and Core Product for Mail, to continually refine the workflows.
How it works
Explore each automated report to see how Enterpret identifies customer signals, prioritizes issues, and surfaces the information the Delight team needs to take action.
The automations identify the signals. The team decides what happens next.
Runs 8 am Mountain
Customer Escalations Queue
Who is most frustrated right now? Ranks the last 24 hours of new and open Mail tickets by frustration, not by account size.
WindowLast 24 hoursOutputRanked top 20Delivers toSlack
Step 1 of 6 · Trigger
The morning run starts the queue
Start each day with one ranked list instead of opening tickets one at a time.
Surface new and open Mail tickets from the past 24 hours from especially frustrated or angry customers
Rank only on anger and frustration. Not severity, not account value
Never stop to ask a question. No human in the loop
Autonomous · scheduled
Step 2 of 6 · Population
Mail tickets only, from the last 24 hours
Keep other brands out of the queue without losing real Mail tickets.
Who is in scope
Ticket brand is Superhuman MailInclude
Brand tags for Grammarly or CodaExclude
Tickets that only mention Grammarly or CodaStays in
Not new or open, or older than 24 hoursExclude
Guard If no recent ticket carries a Mail brand signal, the run flags a likely upstream mapping break instead of posting an empty all-clear.
Gate · unbranded tickets are excluded by design
Step 3 of 6 · Score
Frustration is scored from the customer's words
Rank by how upset the customer is, so the team reaches them first.
Signals and points
Very negative sentiment+32
Dispute, chargeback, fraud or unauthorized-charge theme+30
Profanity+25
Bank, lawyer or legal-action language+25
Threatens to leave or warns others off+18
Several contacts without resolution+15
Heavy ALL-CAPS+15
Feels ignored, misled or very disappointed+12
Repeated punctuation like !! or ??+10
High · 35+Moderate · 15–34Low · under 15
False-positive guard If profanity or an insult is the only signal, the score is capped at 12 and labeled for a manual check. It often comes from quoted or forwarded text.
Autonomous · scores 0 to 100
Step 4 of 6 · Report
One branded page, ranked
Give the team the full picture in one link, with every row checkable.
What the page contains
KPI cards for New, Open, High and Moderate counts
A short methodology: scope, scoring, known false positives, any legal or privacy notice
A ranked top 20 with ticket link, status, opened time, account, tier, score and why it was flagged
A closing read on repeat bug patterns and likely false positives
Every row is verifiable Each row shows its own opened timestamp, so a reader can check the 24-hour window ticket by ticket. Legal and privacy notices are flagged for escalation, not routine triage.
Autonomous · exported as a branded page
Step 5 of 6 · Deliver
A short Slack post, with the full queue one click away
Put the headline where the team already looks.
The message carries
A one-sentence headline finding
New, Open, High and Moderate counts
The two or three most urgent tickets, each with a one-line reason
Any legal or privacy notice, called out separately
A link to the full ranked queue
Autonomous · posts straight to the channel
Step 6 of 6 · Human
The Delight team reads the top and replies
The queue finds the customer. A person decides what to say.
Next stepThe team reviews the top-ranked tickets and drafts the response
Legal or privacy noticeEscalated separately, not treated as routine triage
Customer-facing writesNone. The automation never replies to a customer
Profanity-only hitsCapped and labeled so a person can verify them
Human · reviews and responds
Runs 10 am Mountain
Lightning Bug Report
Are several customers reporting the same bug? Flags five or more tickets on one underlying problem inside 24 hours.
Threshold5+ ticketsWindow24 hoursDelivers toSlack
Step 1 of 6 · Trigger
Lightning reads the rulebook first
Hold every run to one definition of a lightning bug.
Loads the lightning bug definitions before anything else
Its only job is to detect and report clusters of five or more
No root-cause analysis, no frustration scoring, no customer replies
Autonomous · scheduled
Step 2 of 6 · Population
The same 24-hour Mail slice
Look only at fresh tickets, so a cluster means something happened now.
Who is in scope
Brand is exactly Superhuman MailInclude
Status is new or openInclude
Created in the 24 hours before the runInclude
Superhuman and Superhuman Docs brandsExclude
Coda and Grammarly supportExclude
Solved or pending, or open but older than 24 hoursExclude
Brand, not subdomain The run filters on the brand field, never the subdomain, which also carries some Superhuman Docs tickets.
Autonomous · applied before clustering
Step 3 of 6 · Cluster
Group tickets that sound like one bug
Catch the same problem described in different words.
How tickets are grouped
Shared theme and subtheme on the same product areaPrimary
Complaint, problem, bug or help category. Praise and feature requests are outRequired
Shared keywords or error terms in the messageSharpens
Negative sentimentCorroborates
Wording doesn't have to match Customers describe one bug many ways, so this step leans toward recall. The next step enforces precision.
Autonomous · candidate clusters only
Step 4 of 6 · Confirm
Nothing is reported until it passes five checks
Keep a false alarm from reaching the team.
Re-verify each ticket's brand, status, category and created time. Drop any that fail
Recount the survivors and require five or more, exactly
Recompute the span from real timestamps. It must fit inside 24 hours
Confirm a genuine shared symptom, not keyword coincidence
Never round a borderline count or window up to confirmed
When in doubt, it is not confirmed A cluster that fails any check is listed as near-threshold. Never dropped silently, never reported as a bug.
Gate · fail means near-threshold, not a bug
Step 5 of 6 · Report
One report, one summary
Show each confirmed bug with the tickets behind it.
The report
KPI cards for tickets reviewed, confirmed clusters and tickets in clusters
A card per cluster: the bug in one line, its size, its time window and every ticket behind it
Patterns such as a shared org, platform or time of day
A separate near-threshold section
In Slack A short summary, such as 2 lightning bugs confirmed, 14 tickets affected, with a link to the full report. If nothing is confirmed, it says so.
Autonomous · no approval step
Step 6 of 6 · Human
The Delight team reviews each cluster
The report spots the pattern. A person owns the response.
Next stepThe team reviews each flagged cluster and drafts the customer-facing response
Customer-facing writesNone
Pauses for a decisionNone. The run never stops to ask
Human · reviews and responds
Runs 10 am Mountain
Mass Lightning Bug Report
Is a spike turning into an incident? Flags 20 or more tickets on one underlying bug inside a single hour.
Threshold20+ ticketsWindow1 hourDelivers toSlack
Step 1 of 6 · Trigger
Same rulebook, tighter trigger
Treat a sharp spike as a possible service incident, and flag it urgently.
Detects only clusters of 20 or more inside a rolling one-hour window
A confirmed cluster is framed as a potential service incident
No root-cause analysis, no frustration scoring, no customer replies
Autonomous · scheduled
Step 2 of 6 · Population
The same 24-hour Mail slice
Use the exact population the Lightning Bug report uses.
Who is in scope
Brand is exactly Superhuman MailInclude
Status is new or openInclude
Created in the 24 hours before the runInclude
Superhuman and Superhuman Docs brandsExclude
Coda and Grammarly supportExclude
Solved or pending, or open but older than 24 hoursExclude
Brand, not subdomain The run filters on the brand field, never the subdomain, which also carries some Superhuman Docs tickets.
Autonomous · identical to Lightning Bug
Step 3 of 6 · Cluster
Group tickets that sound like one bug
Catch a real incident even when customers use different words.
How tickets are grouped
Shared theme and subtheme on the same product areaPrimary
Complaint, problem, bug or help category. Praise and feature requests are outRequired
Shared keywords or error terms in the messageSharpens
Negative sentimentCorroborates
Wording doesn't have to match Customers describe one bug many ways, so this step leans toward recall. The next step enforces precision.
Autonomous · candidate clusters only
Step 4 of 6 · Confirm
A stricter gate, because a false alarm is costly
Never raise an incident alert on a near miss.
Re-verify each ticket's brand, status, category and created time. Drop any that fail
Recount and require 20 or more, exactly. Never borrow the Lightning Bug threshold
Recompute the span. It must fit inside one hour
Confirm a genuine shared symptom, not keyword coincidence
Never round a borderline count or window up to confirmed
More conservative than Lightning A cluster that fails is listed as near-threshold and never raises the urgency flag.
Gate · when in doubt, do not confirm
Step 5 of 6 · Report
An alert only when it is earned
Make a real incident impossible to miss, and a near miss impossible to mistake for one.
The report
An alert banner above the masthead, shown only if a cluster is confirmed
A card per confirmed cluster with its one-hour window and signs of an active incident
A near-threshold section and a closing assessment
In Slack If confirmed, the message opens with an urgency flag for a potential service incident, then one line per cluster and the total tickets affected. Near-threshold clusters never raise it.
Autonomous · no approval step
Step 6 of 6 · Human
The Delight team takes the incident from here
The report raises the flag. A person runs the response.
Next stepThe team reviews the incident and drafts the customer response or status update
Customer-facing writesNone
Pauses for a decisionNone
Human · reviews and responds
Results
17 emerging bug clusters identified
Over the same period, the Lightning Bug Report identified 17 bug clusters involving five or more customers reporting the same issue within 24 hours.
By connecting related reports across individual support conversations, Enterpret helps the team recognize emerging patterns that may be difficult to spot when reviewing tickets individually.
Each ■ = 1 confirmed bug cluster · Lightning Bug Report, Aug 17–Oct 2
Week
Bug clusters caught
Total
Aug 17–21
0
Aug 24–28
2
Aug 31–Sep 4
5
Sep 7–11
1
Sep 14–18
4
Sep 21–25
2
Sep 28–Oct 2
3
Up to 19 hours of potential manual triage time saved each month
The automations review approximately 110 new tickets each weekday.
At an estimated 30 seconds to manually review and triage each ticket, that's roughly an hour of manual work per day — or approximately 19 hours per month potentially freed up for higher-value customer support.
Time savings are estimated based on average daily ticket volume and manual review assumptions, rather than a measured time study.
A real example: identifying an emerging issue across customer conversations
On September 17, Enterpret identified eight customers reporting the same account-connection issue within approximately 16 hours.
The following morning, the Lightning Bug Report surfaced nine additional reports of a closely related problem.
Rather than relying on individual tickets to reveal a broader issue, the team had visibility into a recurring pattern across customer conversations and was able to solve it quickly.
What's next
Superhuman Mail continues to refine its automations to make customer signals more accurate and actionable.
The team is focused on improving how highly frustrated customers are identified, reducing noise from automated or quoted messages, and connecting emerging bug reports with issues already being tracked.
As these workflows evolve, the goal remains the same: help the Delight team identify what matters most and act on customer feedback sooner.