Ticket triage · Production

Miro uses Amazon Bedrock to boost bug routing accuracy and reduce time-to-resolution from days to hours

The problem

Miro's engineering organization suffered from high rates of bug misrouting, causing repeated reassignments, SLA misses, and an estimated 42 years of cumulative lost productivity annually. Traditional NLP classifiers degraded quickly when organizational structures changed.

First attempt

An existing fine-tuned GPT model showed quickly degrading performance and required retraining whenever teams merged or responsibilities changed, making it impractical for Miro's dynamic engineering organization.

Workflow diagram · grounded in source
1
User submits bug in Slack
trigger
“A user posts a feedback report (for example, of a bug) into a dedicated Slack channel. The report might contain text and media attachments. Media attachments might include a video (typically a screen recording) that describes the process…”
2
Parse media attachments
ai_action
“We use the image understanding capabilities of Amazon Nova Pro to parse the media attachment description as text. One challenge with this approach is that the LLM isn't context-aware by default; it lacks information about the type of ima…”
3
Enrich bug with RAG context
ai_action
“We indexed the following data sources in the knowledge base: Confluence documentation, Miro help center articles, resolved Jira tickets, GitHub READMEs, and Backstage documents (technical documentation and the software catalog)”
4
Route bug to responsible team
routing
“Using Anthropic's Claude Sonnet 4 on Amazon Bedrock, the system combines the enriched bug descriptions along with detailed textual information on each team and their responsibilities into a single, optimized classification prompt that pe…”
5
Generate root cause analysis
ai_action
“BugManager can optionally generate a root cause analysis of the bug. Again, we provide the necessary context to run such an analysis using Amazon Bedrock Knowledge Bases, this time drawing on the entire Miro GitHub code base for referenc…”
6
Human review and override
human_review
“By default, the bug is routed to the most likely team, but users can manually overwrite this selection”
7
Jira ticket created and assigned
output
“a Jira ticket with the original bug description and supporting documentation retrieved from the knowledge bases as well as the results for the root cause analysis is cut and assigned to the selected team”
Reported outcome

BugManager achieved a six-fold reduction in team reassignments and a five-fold improvement in median time-to-resolution, transforming what once took days into an hours-long process.
Top-1 routing accuracy exceeded 75%—a 70% increase over the prior solution—with top-3 accuracy reaching 95%.

Reported metrics
Team reassignmentssix-fold reduction
Median time-to-resolutionfive-fold improvement
Top-1 bug routing accuracyover 75%
routing accuracy improvement vs prior NLP solution70% increase
Show all 10 reported metrics
team reassignmentssix-fold reduction
median time-to-resolutionfive-fold improvement
top-1 bug routing accuracyover 75%
routing accuracy improvement vs prior NLP solution70% increase
top-3 routing accuracy95%
extended thinking accuracy gains7–9%
average classification latency53 seconds
cumulative lost productivity from misrouting (baseline)estimated 42 years of cumulative lost productivity annually
engineering teams in scopenearly 100
bugs routed in productionthousands of bugs and support requests
Reported stack
Amazon BedrockAmazon Bedrock Knowledge BasesAmazon Nova ProAnthropic's Claude Sonnet 4Amazon OpenSearch ServerlessAmazon EKSAmazon S3ConfluenceGitHubBackstageSlackJira
Source
https://aws.amazon.com/blogs/machine-learning/how-miro-uses-amazon-bedrock-to-boost-software-bug-routing-accuracy-and-improve-time-to-resolution-from-days-to-hours/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

BugManager achieved a six-fold reduction in team reassignments and a five-fold improvement in median time-to-resolution, transforming what once took days into an hours-long process.

What tools did this team use?

Amazon Bedrock, Amazon Bedrock Knowledge Bases, Amazon Nova Pro, Anthropic's Claude Sonnet 4, Amazon OpenSearch Serverless, Amazon EKS, Amazon S3, Confluence, GitHub, Backstage.

What results were reported?

Team reassignments: six-fold reduction; Median time-to-resolution: five-fold improvement; Top-1 bug routing accuracy: over 75%; routing accuracy improvement vs prior NLP solution: 70% increase (source-reported, not independently verified).

What failed first in this deployment?

An existing fine-tuned GPT model showed quickly degrading performance and required retraining whenever teams merged or responsibilities changed, making it impractical for Miro's dynamic engineering organization.

How is this ticket triage AI workflow structured?

User submits bug in Slack → Parse media attachments → Enrich bug with RAG context → Route bug to responsible team → Generate root cause analysis → Human review and override → Jira ticket created and assigned.

Related ticket triage cases