Finding out why customers still call, with generative AI
A major U.S. airline had invested heavily in its website and app, yet customers kept calling for things they could do themselves, and maintenance teams had recurring aircraft issues buried in thousands of call transcripts. Keyword reporting could not say why. We built a generative AI pipeline on Amazon Bedrock that reads every Amazon Connect transcript. It sorts each call into categories, pulls out flight numbers, stations and aircraft components, finds the root cause and recommends a next action. Each line of business gets its own dashboard.
- −19%
- calls for self-service-eligible tasks
- 94%
- extraction accuracy, up from ~70%
- <30 min
- maintenance reporting, down from 1–2 weeks
- 3 months
- to production
- An airline used Amazon Bedrock (Nova Pro) to find out why customers still call, from every Amazon Connect transcript.
- 19% fewer calls for self-service-eligible tasks.
- Maintenance issues reported in under 30 minutes instead of 1–2 weeks.
- Avoidable calls down 19%, by fixing what made customers call.
- Maintenance trends visible in minutes, not weeks.
- Live in three months.
- A dashboard per line of business with categories, root causes and next actions.
- Business teams can tune the AI categories themselves.
- Website and app wording fixed from real conversations.
- Connect and conversational analytics to Kinesis, S3, EventBridge and Lambda, then Bedrock Nova Pro.
- Nova Pro benchmarked against Nova Micro and Claude on 20,000 transcripts.
- Results in Aurora PostgreSQL and Snowflake; per-LOB prompts, PII redaction and guardrails.
From a call to an insight, in minutes
On the left, a call transcript. On the right, what Amazon Bedrock (Nova Pro) finds in it: category, root cause, flights or aircraft parts, and the next action, plus how the day’s call drivers move. Pick a call to see each one.
- Call ends
- Transcript sent
- AI reads it
- Saved
- On the dashboard
- Topic
- —
- Detail
- —
- Why they called
- —
- Key details
- —
- What to fix
- —
- Calls analyzed
- 1,284
- Time to report
- —
- Details correct
- 94%
- Sent to
- …
Illustrative transcripts. Callers, tail numbers and counts are fictional, but the flow follows the production design. Transcripts from Amazon Connect and conversational analytics stream through Kinesis and S3, Amazon Bedrock (Nova Pro) analyzes them, and each line of business sees the results on its own dashboard.
The challenge
Self-service was there, but the calls kept coming. “Book a reservation” was a common intent, yet nobody could see that people abandoned the website for large group bookings, or because travel credit rules were confusing. Conversational analytics gave intents and sentiment, not root causes or what to do about them.
On the operations side, maintenance control calls kept mentioning the same aircraft and components, but the pattern sat in unstructured transcripts. Reviewing them by hand took one to two weeks, long after the moment to act.
The airline wanted to know why people call, fix what drives avoidable calls, and spot recurring operational issues while they still matter, across every line of business.
What made this different
Conversational analytics tells you what was said. This tells you why, and what to change.
- 20,000
real call transcripts used to choose the AI model.
Nova Pro was chosen over Nova Micro and Claude for accuracy at the best cost per transcript, measured on the airline’s own calls.
- 94%
of flight and aircraft-part details extracted correctly.
Up from about 70% by hand, so maintenance teams can trust the trend lines.
Separate AI instructions for each team, instead of one for everyone.
Maintenance, crew, baggage and customer service talk differently. A single prompt gave inconsistent categories; per-team prompts fixed it.
Success targets agreed before the first sprint.
At least 15% fewer repeat calls, 90% extraction accuracy and reporting within an hour, agreed before the first sprint.
What changed
- Why people callBefore: Intents and sentiment onlyAfter: Categories, subcategories and root causes for every call
- What to doBefore: Teams guessed from summariesAfter: A recommended next action for digital and operations teams
- Avoidable callsBefore: Customers called for self-service tasksAfter: 19% fewer calls for self-service-eligible tasks
- Aircraft and flightsBefore: Pulled out by hand, about 70% accurateAfter: Extracted by AI, 94% accurate
- Maintenance reportingBefore: 1–2 weeks behindAfter: Under 30 minutes
- DashboardsBefore: Standard Connect reportsAfter: A dashboard per line of business, plus AI configuration management
- Model choiceBefore: No benchmarkAfter: Nova Pro chosen over Nova Micro and Claude on 20,000 real transcripts
- Responsible AIBefore: —After: Guardrails, PII redaction and human review
Hover or tap a tile to see what it replaced.
Fewer avoidable calls, and issues found in minutes.
Acceptance criteria were set up front: at least 15% fewer repeat calls, at least 90% extraction accuracy, and reporting within an hour. All three were beaten.
- Calls for self-service-eligible tasksBeforeBaselineAfter−19%
- Flight and component extraction accuracyBefore~70% manualAfter94% AI
- Maintenance reporting delayBefore1–2 weeksAfter<30 min
As reported for the engagement in production; repeat calls measured on self-service-eligible tasks.
How a call becomes an insight
Every call is analyzed within minutes of ending. Business teams see the pattern, not just the call.
- 01 · CallA call endsAmazon Connect and conversational analytics produce the transcript and contact record.
- 02 · StreamInto the pipelineKinesis streams it to S3; EventBridge starts the Lambda workflow.
- 03 · AnalyzeNova Pro reads itCategory, subcategory, root cause, next action, plus flights, stations and aircraft components.
- 04 · StoreStructured resultsInto Aurora PostgreSQL for the agent desktop and Snowflake for enterprise reporting.
- Self-service fix
- Content fix
- Maintenance trend
- Process gap
- 05 · ActOn the dashboardEach line of business sees its trends and recommended actions.
How we ran it
- 1Data and pipelineSprint 1: data readiness, the Kinesis to S3 pipeline, and the first prompts.
- 2Benchmark modelsSprint 2: Nova Pro, Nova Micro and Claude compared on 20,000 transcripts for quality, cost and latency.
- 3IntegrateSprint 3: results into the agent desktop database and the line-of-business dashboards.
- 4Prove and go liveSprint 4: UAT, KPI validation against the acceptance criteria, production cutover.
- 5StabilizeNinety days of tuning prompts, categories and throughput in production.
The outcome
- 19% fewer calls for tasks customers could do themselves
- Flight, station and aircraft component extraction 94% accurate, up from about 70% by hand
- Maintenance control reporting delay cut from 1–2 weeks to under 30 minutes
- Root causes and next best actions for digital and operations teams
- A dashboard for each line of business, with AI categories business teams can configure
- Delivered in three months, then a 90-day stabilization
- Contact center
- Amazon Connect
- Conversational analytics transcripts
- Contact records
- AI
- Amazon Bedrock
- Amazon Nova Pro
- Pipeline
- Amazon Kinesis
- Amazon S3
- Amazon EventBridge
- AWS Lambda
- Data
- Amazon Aurora PostgreSQL
- Snowflake
- Insight
- Dashboards per line of business
- AI configuration management
Four lessons. One prompt for every team does not work; each line of business needs its own. Transcript analysis needs an event-driven pipeline built for bursts. General models need aviation context to read maintenance conversations well. And insights only matter if teams can act on them, which is why every line of business got its own dashboard and the ability to tune categories.
A partner AWS has validated, on exactly this kind of work.
- AWS Advanced Tier Services PartnerRecognized by AWS for proven delivery, certified people and customer outcomes.
- AWS Migration Services CompetencyValidated by AWS for moving and modernizing workloads on AWS at scale.
- Amazon Connect Delivery PartnerValidated by AWS to deliver Amazon Connect, including migrations from Avaya, Genesys, Cisco and other platforms.
- AWS AI Services CompetencyValidated by AWS for building AI into real products, including the contact center.
- AWS Experience-Based Acceleration (EBA) partnerWe run EBA engagements with AWS to speed up migrations with hands-on, time-boxed sprints.
- ISO/IEC 27001 certifiedOur information security management system is independently certified to the international standard.
- SOC 2 Type IIIndependent auditors have tested our security controls over time, not just at a single point.
More work
All case studies →- Amazon Connect
- Amazon Lex
- AWS Lambda
- Amazon Connect
- Contact flows
- AWS Lambda
- Amazon EKS
- Amazon Kinesis
- Amazon DynamoDB








