Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.
Problem
Domain: IT operations and support automation
WEX needed to reduce high-volume operational support load without reducing quality. Their GTS organization handled more than 40,000 yearly support requests across complex cloud, network, security, and operations domains, including repetitive network troubleshooting and late-night EBS volume incidents that required tribal knowledge, cross-team escalation, and manual runbook execution.
Constraints:
WEX operates at regulated global-commerce scale, so reliability, security, governance, and cost optimization are core requirements.
Support workflows span hundreds of AWS accounts, Azure, Google Cloud, eight regions, and multiple on-premises data centers.
Google Chat has no native AWS collaboration integration, so WEX needed a custom secure front door.
Agents need authorization based on enterprise entitlements, not just chat identity.
Operational agents must use trusted existing automations, apply guardrails, log decisions, and escalate when policy conditions do not match.
Architecture
WEX built ChatGTS as a production agent platform rather than just a chatbot. Google Chat requests enter through Imperva/WAF and API Gateway, a Lambda router validates signed tokens, filters noise and oversized prompts, acknowledges quickly to absorb model latency, and Step Functions orchestrates specialized agents. State and conversations are stored in DynamoDB; reasoning traces land in S3; Bedrock hosts agents and knowledge. Active Directory entitlements are fetched from Google identity and cached in DynamoDB, then injected as immutable context tags. Guardrails run at the orchestrator edge to sanitize text, block topics, and redact PII. Agents use a shared Kendra Gen AI index over Confluence, GitHub, Google Drive, source code, diagrams, policies, and runbooks. For actions, Bedrock Action Groups invoke tightly scoped Lambdas in VPCs and MCP servers expose trusted playbooks such as SSM documents, Jira, SNS, and infrastructure drift remediation.
Diagram confidence: reconstructed (reconstructed from prose).
Investigates VPC connectivity by using Reachability Analyzer, flow logs, recent changes, and known issues.
Shared Kendra knowledge base, MCP and action tools
EBS triage agent
Processes CloudWatch alerts, examines OS, owner, criticality, and expansion history, and decides whether to escalate.
Jira via AgentCore Gateway, EBS maintenance agent
EBS maintenance agent
Executes trusted diagnostics, backup, cleanup, and expansion runbooks exposed as SSM/MCP tools.
MCP and action tools, SNS status publisher, Drift remediation loop
Shared Kendra knowledge base
Hybrid keyword/vector/multimodal search layer over enterprise runbooks, docs, code, diagrams, and policies.
Knowledge source connectors
Knowledge source connectors
Confluence, GitHub, Google Drive, and other automatically synced content sources.
Shared Kendra knowledge base
MCP and action tools
MCP servers, Bedrock Action Groups, SSM documents, and scoped Lambdas used for operational actions.
Bedrock AgentCore Runtime
Bedrock AgentCore Runtime
Runtime for Docker containers and MCP servers used by the agent platform.
MCP and action tools
Jira via AgentCore Gateway
Ticketing integration for incidents, logging, and later RCA context.
EBS triage agent, Drift remediation loop
SNS status publisher
Publishes status to on-call systems, dashboards, and downstream consumers.
—
Drift remediation loop
Traces resource changes back to Terraform or CloudFormation and opens PRs, issues, or emails.
—
DynamoDB state store
Stores users, chat spaces, sessions, messages, cached entitlements, and relationships keyed by session/trace ID.
Splunk visibility
S3 reasoning traces
Flight-recorder storage for reasoning traces and platform decisions.
Splunk visibility
Splunk visibility
Compliance, InfoSec, risk, and legal view of logs, redactions, policies, user messages, and agent messages via Kinesis.
Operational insight dashboard
Operational insight dashboard
Custom dashboard for replaying conversations and finding agent or knowledge-base improvement areas.
—
Data flow
A user or system event enters through Google Chat or CloudWatch. Chat requests pass through WAF/Imperva, API Gateway, and the Lambda router; alerts can be sent directly to an agent with metrics and context. The router validates identity, gets entitlements from Active Directory, caches them in DynamoDB, wraps the prompt with immutable context tags, and starts a Step Functions workflow. The orchestrator routes to specialized agents, which use shared Kendra knowledge and scoped MCP/action tools. Results, messages, sessions, and traces are persisted in DynamoDB and S3; logs flow through Kinesis into Splunk; user responses are formatted for Google Chat with citations; operational actions update Jira/SNS and may open PRs or issues against Terraform/CloudFormation origins.
7 supporting quotes
“we use Google Chat, and there is no native integration between Google and AWS for collaboration. So on the left, left right, we have our users chatting with us in our workspace domain. Now those requests come over the internet. So in the middle we have our”transcript, stated
“with us in our workspace domain. Now those requests come over the internet. So in the middle we have our WAF, with Imperva, that secures our inbound traffic. And on the right we have our AWS environment fronted by API gateway. And from there we use a Lambda to route and acknowledge”transcript, stated
“we use a Lambda to route and acknowledge messages. And then we use Step Functions to orchestrate all of our agents.”transcript, stated
“and acknowledge messages. And then we use Step Functions to orchestrate all of our agents. We store state and conversations in Dynamo, reasoning traces land in S3. And of course Bedrock hosts our agents and knowledge. On paper it looks pretty simple. It's neat, it's easy, it's serverless,”transcript, stated
“agents. We store state and conversations in Dynamo, reasoning traces land in S3. And of course Bedrock hosts our agents and knowledge. On paper it looks pretty simple. It's neat, it's easy, it's serverless, it's exactly what we wanted. Now our chat application”transcript, stated
“you're allowed to talk to us. So we take that identity and we reach out to active directory and we fetch your entitlements and cash them in Dynamo. Kept it pretty simple. It's all based on group memberships and OU. And that way we're not beating up on systems that were”transcript, stated
“we wrap the entire prompt in context tags, and we include your identity and your entitlements and send that downstream to the agent. So now whatever the user claims, the agent only trusts that context because it's immutable. And here, if something goes”transcript, stated
Why this architecture
WEX deliberately reused service-oriented and serverless patterns because the agent system had to be production-safe. Step Functions disciplines non-deterministic agents with retries, fallbacks, and state transitions. Specialized agents keep reasoning focused and handoffs clean. Guardrails act as contracts. Kendra centralizes enterprise knowledge as managed infrastructure. DynamoDB, S3, Kinesis, and Splunk provide persistent observability and compliance trails. MCP and action groups let agents act through trusted existing automations instead of inventing new operational paths.
6 supporting quotes
“wanted to build more than a chat bot and a rag pipeline, right? We wanted to create a platform. We wanted something sustainable, something extensible, something that would inspire all of the other teams to come and collaborate and help us expand this, because operations”transcript, stated
“we built agents with boundaries, we gave them clear responsibilities, and we let them all work independently. And we saw guardrails become the contracts for what those agents could and couldn't do. And events, it became less about, something happened, and”transcript, stated
“Bedrock gave us the intelligence, the parts that can think and create and make decisions. Step Functions gave us the discipline we needed to keep it all in check. We used the retries and fallbacks and state transitions, and that's where it really clicked. You know, like Andrew said,”transcript, stated
“and make decisions. Step Functions gave us the discipline we needed to keep it all in check. We used the retries and fallbacks and state transitions, and that's where it really clicked. You know, like Andrew said, these agents, they can be autonomous, but they don't have”transcript, stated
“exposed them as tools to the agent on an MCP server. So now whatever the agent decides to do, it's using the same automations that we already trust, and we're not waking somebody up to push that run button. We've eliminated the chats, the texts, the pages, the cross team”transcript, stated
“This is us learning how to use what we already know, and we're reusing what already works. It's not magic, it's just engineering and architecture. And that's kind of the point. So I'm gonna zoom out, I'm gonna talk a little bit about how we got there, and then we can”transcript, stated
Models used: AWS Bedrock agents · Bedrock-hosted models · Strands SDK · Amazon Kendra Gen AI index · Cursor · Claude · Copilot
Scale
WEX operates in more than 200 countries.
“for organizations in more than 200 countries.”transcript, stated
WEX processed over $230 billion in transactions in more than 20 currencies.
“Last year we processed over $230 billion in transactions in more than 20 different currencies.”transcript, stated
Global Technology Services saw more than 40,000 support requests in the prior year.
“last year Global Technology saw more than 40,000 support requests.”transcript, stated
ChatGTS moved from pilot to production in under three months.
“helped us go from pilot to production in under three months.”transcript, stated
The platform serves well over 2,000 internal users.
“now we've got well over 2000 users internally”transcript, stated
The network environment spans hundreds of AWS accounts, eight regions, Azure, Google, and multiple on-prem data centers.
“we operate hundreds of AWS accounts, and we've got Azure and Google all spanning eight regions and multiple on-prem data centers”transcript, stated
The EBS remediation use case targets 2 a.m. operational incidents.
“nobody wants to do that at 2:00 a.m.”transcript, stated
The pace of the build made four or five weeks feel like ten years to the speaker.
“it feels like it, but it was only four or five weeks.”transcript, stated
Alternatives considered (roads not taken)
Build only a chatbot or RAG pipeline.
Rejected because: WEX wanted a sustainable platform that could support collaboration, operational actions, and future expansion.
3 supporting quotes
“look at some diagrams, we'll see what it looks like under the hood. So when we started this, we knew we wanted to build more than a chat bot and a rag pipeline, right? We wanted to create a platform. We wanted something sustainable, something extensible, something that would inspire”transcript, stated
“we started this, we knew we wanted to build more than a chat bot and a rag pipeline, right? We wanted to create a platform. We wanted something sustainable, something extensible, something that would inspire all of the other teams to come and collaborate”transcript, stated
“more than a chat bot and a rag pipeline, right? We wanted to create a platform. We wanted something sustainable, something extensible, something that would inspire all of the other teams to come and collaborate and help us expand this, because operations”transcript, stated
Let humans handle every repetitive support request.
Rejected because: WEX had 40,000 support requests, many repetitive and time-consuming, and wanted to reduce volume without reducing quality.
3 supporting quotes
“trouble. My team's job at WEX, sorry, is to make cloud feel simple even when it isn't. And last year Global Technology saw more than 40,000 support requests. Now that's a lot, right? But if you think about it, every last one of them is critical to somebody, even if they are repetitive”transcript, stated
“that's a lot, right? But if you think about it, every last one of them is critical to somebody, even if they are repetitive and time consuming for us. And we've got operations and we've got SRE and support at all of those tiers, but we're always looking for ways to reduce”transcript, stated
“for us. And we've got operations and we've got SRE and support at all of those tiers, but we're always looking for ways to reduce that number without reducing quality. So like everyone else, we started looking at AI, but we didn't want to do anything flashy or complicated. We just wanted”transcript, stated
Use one monolithic agent for all operations work.
Rejected because: WEX uses specialized agents with one job each so reasoning stays focused and handoffs are cleaner.
3 supporting quotes
“it came to agents, I mean we were inspired by SOA, right? Think about what it taught us to break up the monolith, right? We took these big systems and we turned them into smaller pieces with a clear purpose. That same discipline applies”transcript, stated
“purpose. That same discipline applies to agents, and we built specialized agents and gave them one job that they have to do really well. And that keeps the focus really clean and the reasoning sharp, and then the handoffs between agents is even cleaner. But instead of having one giant orchestrator”transcript, stated
“and we built specialized agents and gave them one job that they have to do really well. And that keeps the focus really clean and the reasoning sharp, and then the handoffs between agents is even cleaner. But instead of having one giant orchestrator pulling all the strings,”transcript, stated
Give agents uncontrolled autonomy.
Rejected because: WEX combines autonomy with Step Functions, guardrails, retries, fallbacks, and escalation so agents stay inside operational boundaries.
3 supporting quotes
“fallbacks and state transitions, and that's where it really clicked. You know, like Andrew said, these agents, they can be autonomous, but they don't have to be. You can give them as much freedom or as much control as you need to, and that's what made them fit so well in these operational”transcript, stated
“You know, like Andrew said, these agents, they can be autonomous, but they don't have to be. You can give them as much freedom or as much control as you need to, and that's what made them fit so well in these operational workflows. So Google gives us a trusted identity, but it has”transcript, stated
“needed to have guardrails. We needed some way that we could enforce policy and compliance, so we applied guardrails at the edge with the orchestrator, and this gives us defense in depth for every decision that happens in the platform. Now, we're sanitizing text, we're blocking”transcript, stated
Let agents invent new operational remediations.
Rejected because: The EBS maintenance agent uses existing trusted SSM documents and runbooks already used by operations engineers.
3 supporting quotes
“RCA document. So now that we know what's happening, it passes on to a maintenance agent, and this agent can choose from a library of preexisting SSM documents. Now, keep in mind, these are the same documents that our ops engineers are using at 2:00 a.m. We've got Playbooks to run”transcript, stated
“and this agent can choose from a library of preexisting SSM documents. Now, keep in mind, these are the same documents that our ops engineers are using at 2:00 a.m. We've got Playbooks to run diagnostics and backup and cleanup and expansion. It's all the usual runbooks, but we've exposed”transcript, stated
“exposed them as tools to the agent on an MCP server. So now whatever the agent decides to do, it's using the same automations that we already trust, and we're not waking somebody up to push that run button. We've eliminated the chats, the texts, the pages, the cross team”transcript, stated
Store enterprise knowledge as a loose document dump.
“sure. So we wanted to build something that could grow and evolve with the organization, and we needed something more than just a a dumping ground for files in S3. So we chose Kendra, with the Gen AI index because it gives us this hybrid. We get keyword and vector search with multimodal”transcript, stated
“more than just a a dumping ground for files in S3. So we chose Kendra, with the Gen AI index because it gives us this hybrid. We get keyword and vector search with multimodal embeddings, and with all of the built-in connectors for things like Confluence and GitHub and Google Drive, we're”transcript, stated
“can deploy their changes through GitHub's pipelines. Now, it's still enterprise knowledge, we're just treating it like infrastructure now. I'm not gonna lie, observability was a bit of a puzzle. Like we have this third party chat up on the front end, this hybrid”transcript, stated
Patterns
multi-agent-orchestration(agent_architecture)
WEX decomposes operations automation into a conductor plus specialized agents for network troubleshooting, EBS triage, and EBS maintenance.
3 supporting quotes
“we built specialized agents and gave them one job that they have to do really well.”transcript, stated
“handoffs between agents is even cleaner. But instead of having one giant orchestrator pulling all the strings, ours acts more like a conductor, it sits in the middle, it can interpret intent, and then it could figure out what's happening, where is it happening, and then it connects that”transcript, stated
“it connects that problem to the right expert or the right team of experts.”transcript, stated
durable-execution(orchestration)
Step Functions supplies retries, fallbacks, circuit breakers, and state transitions around non-deterministic agent behavior.
3 supporting quotes
“Step Functions gave us the discipline we needed to keep it all in check.”transcript, stated
“We used the retries and fallbacks and state transitions, and that's where it really clicked. You know, like Andrew said, these agents, they can be autonomous, but they don't have to be.”transcript, stated
“There's retries and fallbacks and circuit breakers, and it's baked into every state transition.”transcript, stated
tool-use-orchestration(agent_tools)
Agents call MCP servers, Bedrock Action Groups, scoped Lambdas, SSM documents, Jira, SNS, and infrastructure tooling as network services.
3 supporting quotes
“They're calling APIs and CP servers. They're executing the tools we've given them to do their jobs.”transcript, stated
“we can run these Lambdas inside of our VPCs with tightly scoped permissions”transcript, stated
“It's all the usual runbooks, but we've exposed them as tools to the agent on an MCP server. So now whatever the agent decides to do, it's using the same automations that we already trust, and we're not waking somebody up to push that run button.”transcript, stated
rag-retrieval(knowledge)
All agents use a shared Kendra knowledge base over runbooks, reference architectures, source code, diagrams, policies, and living documentation.
3 supporting quotes
“Now, all of our agents tap into the same knowledge base. So that means when we have a Q and A agent answering a question about connectivity, it's pulling from the same material an ops agent would if it were troubleshooting.”transcript, stated
“pulling from the same material an ops agent would if it were troubleshooting.”transcript, stated
“It's like giving every application in your service layer, this unified data plane, only now it's made up of runbooks and reference architectures and living documentation. But thinking isn't enough, right?”transcript, stated
hybrid-search(retrieval)
WEX uses Kendra Gen AI index for keyword and vector search with multimodal embeddings across enterprise sources.
3 supporting quotes
“we chose Kendra, with the Gen AI index because it gives us this hybrid.”transcript, stated
“We get keyword and vector search with multimodal embeddings, and with all of the built-in connectors for things like Confluence and GitHub and Google Drive, we're able to keep all of our information in sync automatically with Chron schedules.”transcript, stated
“We get keyword and vector search with multimodal embeddings, and with all of the built-in connectors for things like Confluence and GitHub and Google Drive, we're able to keep all of our information in sync automatically with Chron schedules.”transcript, stated
“all of those messages come with a signed token that we can validate”transcript, stated
“So now whatever the user claims, the agent only trusts that context because it's immutable. And here, if something goes sideways, say the agent fails or times out or can't produce a response, we log the error and we can send a safe response back to the user without blowing up the whole workflow.”transcript, stated
“we're sanitizing text, we're blocking topics and redacting PII, both on the way in and on the way out.”transcript, stated
observability-tracing(operations)
WEX stores messages and traces, pushes logs through Kinesis to Splunk, and uses a dashboard to replay conversations and inspect reasoning.
4 supporting quotes
“Because every item here keys back to a session and a trace ID. And those traces land in S3, and this became our black box.”transcript, stated
“This is like the flight recorder that captures every decision that's made across the platform.”transcript, stated
“So we pushed all of our logs into Splunk through Kinesis, and now even info security and compliance and risk and legal, they've all got a real time view of what's happening in the platform.”transcript, stated
“But literally with this dashboard, we can replay a conversation step by step, and then we can see the story sort of unfold.”transcript, stated
agent-memory(memory)
Operational investigations are logged so agents can recognize recurring issues and build memory over time.
3 supporting quotes
“since we're logging all of these investigations, we're able to spot recurring issues”transcript, stated
“And as they learn, they're building memory. And over time they can start to recognize patterns. Now we can learn where we're starting to, we're cleaning up on occasion where we have to escalate and maybe they start to spot issues at the application layer.”transcript, stated
“And over time they can start to recognize patterns. Now we can learn where we're starting to, we're cleaning up on occasion where we have to escalate and maybe they start to spot issues at the application layer.”transcript, stated
human-in-the-loop-gate(control)
The EBS workflow escalates to humans when expansion history, criticality, platform ownership, or other triage signals look problematic.
“if anything looks off the agent steps out of the way and we escalate back to a human”transcript, stated
“So at this point, if anything looks off the agent steps out of the way and we escalate back to a human and then we're back to the way we do it now. But if not, our agent can connect into Jira through AgentCore Gateway and it can open a ticket and start logging the incident.”transcript, stated
Tradeoffs
Start with high-volume, high-friction, well-understood support work.
Chose Known runbooks and repetitive operational tasks over Flashy or novel agent use cases.
Rationale: WEX looked at support data and existing processes to find work where agents could make a practical dent safely.
3 supporting quotes
“We didn't want to just build agents because we could, or get stuck in that cycle of chasing shiny objects whenever something new came out.”transcript, stated
“we looked at our data, our support history, and those 40,000 tickets”transcript, stated
“that high volume, high friction, well understood work, that's where we knew we could make a dent with AI.”transcript, stated
Use Step Functions around Bedrock agents.
Chose Deterministic workflow discipline plus autonomous reasoning over Pure autonomous agent coordination.
Rationale: Step Functions gave retries, fallbacks, and state transitions to keep operational agents accountable.
3 supporting quotes
“Bedrock gave us the intelligence, the parts that can think and create and make decisions.”transcript, stated
“Step Functions gave us the discipline we needed to keep it all in check.”transcript, stated
“You can give them as much freedom or as much control as you need to”transcript, stated
Use existing trusted runbooks as agent tools.
Chose SSM documents and production automations exposed through MCP over New agent-written remediation procedures.
Rationale: This lets the agent automate what operations already trusts while reducing night work and manual button-pushing.
3 supporting quotes
“So now that we know what's happening, it passes on to a maintenance agent, and this agent can choose from a library of preexisting SSM documents. Now, keep in mind, these are the same documents that our ops engineers are using at 2:”transcript, stated
“Now, keep in mind, these are the same documents that our ops engineers are using at 2:00 a.m. We've got Playbooks to run diagnostics and backup and cleanup and expansion.”transcript, stated
“MCP server. So now whatever the agent decides to do, it's using the same automations that we already trust, and we're not waking somebody up to push that run button. We've eliminated the chats, the texts, the pages, the cross team escalations, and now these systems, they're starting to learn”transcript, stated
Centralize knowledge for all specialized agents.
Chose Shared Kendra knowledge base over Separate knowledge sources per agent.
Rationale: Shared knowledge keeps Q&A and ops agents consistent even though they are separate services.
3 supporting quotes
“Now, all of our agents tap into the same knowledge base. So that means when we have a Q and A agent answering a question about connectivity, it's pulling from the same material an ops agent would if it were troubleshooting.”transcript, stated
“They're separate services, but it's consistent understanding. It's like giving every application in your service layer, this unified data plane, only now it's made up of runbooks and reference architectures and living documentation.”transcript, stated
“unified data plane, only now it's made up of runbooks and reference architectures”transcript, stated
Treat knowledge source configuration as infrastructure.
Chose Terraform, GitHub, CI/CD, and SME pull requests over Manually maintained knowledge-base configuration.
Rationale: This gives content owners self-service control while keeping enterprise knowledge governed and deployable.
3 supporting quotes
“they're configured in Terraform and they live in GitHub, and we deploy them with CICD.”transcript, stated
“So with self-service, we allow the people who own that content, the subject matter experts, to maintain it themselves. They can open a pull request and we can deploy their changes through GitHub's pipelines.”transcript, stated
“They can open a pull request and we can deploy their changes through GitHub's pipelines.”transcript, stated
Let agents remediate drift after incidents.
Chose Trace upstream to Terraform/CloudFormation and open PRs/issues/emails over Leaving manual incident changes outside infrastructure-as-code.
Rationale: This closes the loop between operations and infrastructure so emergency remediation does not become unmanaged drift.
3 supporting quotes
“follow that resource all the way upstream back to the Terraform or the cloud formation”transcript, stated
“it's gonna open a pull request, or it's gonna create a Jira issue, or send an email.”transcript, stated
“It's gonna do something that closes that loop between operations and infrastructure”transcript, stated
Gotchas (what bit them)
The hardest part was not the model, but enterprise documentation and knowledge management.
Root cause: Enterprise documentation lives across many systems, and even when available, people often do not read it.
Resolution: Treat the Kendra knowledge base and its data sources as infrastructure managed through Terraform, GitHub, CI/CD, and SME pull requests.
3 supporting quotes
“Documentation, of course. And ironically, that was the hardest part of all this”transcript, stated
“it can live anywhere. And even when people can find it, they don't like to read it”transcript, stated
“they're configured in Terraform and they live in GitHub, and we deploy them with CICD.”transcript, stated
Agent autonomy needs workflow discipline.
Root cause: Non-deterministic agents can create chaos if they are not wrapped in retries, fallbacks, state transitions, and guardrail contracts.
Resolution: Use Step Functions and guardrails to provide structure while preserving useful autonomy.
3 supporting quotes
“Step Functions gave us the discipline we needed to keep it all in check.”transcript, stated
“It gave us this structure where there was so much potential for chaos. And it's all the same patterns, right? There's retries and fallbacks and circuit breakers, and it's baked into every state transition.”transcript, stated
“we saw guardrails become the contracts for what those agents could and couldn't do.”transcript, stated
Chat identity did not equal authorization.
Root cause: Google could prove who was talking to the chat app, but not what that user was allowed to do inside WEX systems.
Resolution: Fetch Active Directory entitlements, cache them, and inject immutable identity/entitlement context into prompts.
3 supporting quotes
“Google gives us a trusted identity, but it has no concept of permissions.”transcript, stated
“we reach out to active directory and we fetch your entitlements and cash them in Dynamo.”transcript, stated
“whatever the user claims, the agent only trusts that context because it's immutable.”transcript, stated
Reasoning traces required a different observability model than ordinary request logs.
Root cause: Agent systems need to inspect behavior and reasoning, not just HTTP status codes.
Resolution: Store trace-linked messages in DynamoDB, reasoning traces in S3, and logs in Splunk, plus a replay dashboard.
3 supporting quotes
“We still need to see everything, but there's more to this than 200s and 500s now. We're looking at behavior and reasoning.”transcript, stated
“But instead of following a request, we're following a train of thought. And of course we needed transparency. So we pushed all of our logs into Splunk through Kinesis, and now even info security and compliance and risk and legal, they've all got a real time view of what's happening in the platform.”transcript, stated
“But literally with this dashboard, we can replay a conversation step by step, and then we can see the story sort of unfold.”transcript, stated
Volume expansion automation can hide underlying problems if unbounded.
Root cause: Repeatedly expanding disks can defer root-cause cleanup and keep kicking the can down the road.
Resolution: Check expansion history, maintain cap policies, and escalate when conditions look off.
3 supporting quotes
“How many times have we expanded the volume here? You know, we maintain policies to cap expansion, right? You don't want to just keep adding disc, you're kicking the can down the road, and that's never gonna work.”transcript, stated
“You know, we maintain policies to cap expansion, right? You don't want to just keep adding disc, you're kicking the can down the road, and that's never gonna work.”transcript, stated
“You don't want to just keep adding disc, you're kicking the can down the road”transcript, stated
Network troubleshooting depended on tribal knowledge across many domains.
Root cause: Engineers may lack access or expertise for transit gateways, firewalls, VPNs, flow logs, and core network accounts.
Resolution: Use an agent to fan out across Reachability Analyzer, flow logs, recent changes, and known issues, then escalate with complete context.
3 supporting quotes
“Now, Jared doesn't have access to the transit gateways, or the firewalls or the VPNs”transcript, stated
“And now what used to require all this tribal knowledge across all these different domains, it can happen in minutes and anyone can use it now, even Jared.”transcript, stated
“it fans out and it checks flow logs, it looks at any recent changes in the network”transcript, stated
Evaluation
WEX evaluates and improves the system through support-ticket volume selection, internal user adoption, workflow observability, trace replay, Splunk compliance visibility, recurring-issue detection, dashboard-driven roadmap decisions, and explicit escalation behavior when agents encounter unsafe or abnormal conditions. The talk also notes AgentCore Evaluations for non-deterministic testing as a new AWS capability.
Metrics: More than 40,000 yearly support requests as target pool · Pilot to production in under three months · Well over 2,000 internal users · Network investigations completed in minutes · CloudWatch alert response without waking an engineer · Trace IDs and session-linked persistence · Replayable conversations and agent struggle points · Redactions and policy enforcement visible to InfoSec/compliance/risk/legal
6 supporting quotes
“last year Global Technology saw more than 40,000 support requests.”transcript, stated
“And they're really what helped us go from pilot to production in under three months. And now we've got well over 2000 users internally, so it's been a good year.”transcript, stated
“And now we've got well over 2000 users internally, so it's been a good year. But first we've got the company slide.”transcript, stated
“And now what used to require all this tribal knowledge across all these different domains, it can happen in minutes and anyone can use it now, even Jared.”transcript, stated
“But literally with this dashboard, we can replay a conversation step by step, and then we can see the story sort of unfold.”transcript, stated
“This is what drives our roadmap. This is what tells us what we need to build next. I'm almost done, I promise. It's my last one, then everybody can go.”transcript, stated
Outcomes
WEX moved ChatGTS from pilot to production in under three months.
The system serves well over 2,000 internal users.
The network troubleshooting agent can perform investigations in minutes and provide escalation context.
The EBS workflow makes agents the first line of defense for some CloudWatch alerts, using trusted SSM playbooks and escalating when conditions are abnormal.
The platform logs reasoning traces, conversations, redactions, policies, and agent behavior for compliance and improvement.
Knowledge-base management became infrastructure-managed through Terraform, GitHub, CI/CD, and SME pull requests.
WEX operates in more than 200 countries, processed over $230 billion in more than 20 currencies, saw more than 40,000 GTS support requests, spans hundreds of AWS accounts plus Azure/Google/eight regions/on-prem data centers, reached production in under three months, and now serves well over 2,000 internal users.
6 supporting quotes
“And they're really what helped us go from pilot to production in under three months. And now we've got well over 2000 users internally, so it's been a good year.”transcript, stated
“And now we've got well over 2000 users internally, so it's been a good year. But first we've got the company slide.”transcript, stated
“And now what used to require all this tribal knowledge across all these different domains, it can happen in minutes and anyone can use it now, even Jared.”transcript, stated
“instead of paging an engineer, we can send these alerts to an agent with all of those metrics and context.”transcript, stated
“info security and compliance and risk and legal, they've all got a real time view”transcript, stated
“Now, it's still enterprise knowledge, we're just treating it like infrastructure now. I'm not gonna lie, observability was a bit of a puzzle.”transcript, stated