ShipStation Builds AI IT Help Desk with AgentCore, Auto-Resolving 97.5% of Questions

Internal IT support has long been an efficiency black hole — employees send messages when they hit a problem, wait for replies, go back and forth, and the IT team gets buried under a flood of repetitive inquiries. ShipStation Global's engineering team changed that with an AI agent called Deckhand: it lives in Slack, is built on Amazon Bedrock AgentCore, and resolved 97.5% of channel questions in its first three months, with a 96% positive feedback rate.

What Deckhand Is: An IT Colleague in Slack

Deckhand isn't a simple FAQ bot. It's designed as an IT help desk agent that actually gets work done, with three core capabilities:

• Answering routine IT questions from a curated knowledge base: Employees post questions in the Slack channel, and Deckhand retrieves answers from a knowledge base maintained by the team rather than letting the large model improvise.
• Reading screenshots of error messages: Employees can drop a screenshot of an error popup directly into the channel, and Deckhand can recognize the image content and provide the corresponding explanation.
• Automatically creating Jira tickets: When a problem falls outside what the knowledge base covers, it creates a ticket for the employee in Jira, eliminating the manual form-filling and handoff steps.

Together, these three capabilities cover the most common scenarios in IT support: answer what can be answered directly, and automatically file a ticket for what can't.

Why Amazon Bedrock AgentCore

ShipStation Global built Deckhand on top of Amazon Bedrock AgentCore, and that choice is worth noting. AgentCore provides the underlying infrastructure needed to build, deploy, and scale AI agents, so teams don't have to build agent runtimes, tool invocation, and session management from scratch.

For engineering teams, this means they can focus on business logic — how to organize the knowledge base, how to write the Jira integration, how to design the Slack interaction — instead of reinventing the agent framework. ShipStation Global's case shows that the barrier to deploying enterprise-grade AI agents is dropping fast.

What Is Amazon Bedrock AgentCore

Amazon Bedrock is AWS's managed generative AI platform. Developers can call foundation models from multiple vendors (such as Anthropic Claude, Meta Llama, Amazon Nova, and more) through a unified API without managing model inference infrastructure themselves. AgentCore is the agent runtime layer built on top of Bedrock, specifically addressing the challenge of "turning a large model into an agent that can actually get work done." Its core components typically include: an agent runtime (a managed execution environment that handles session lifecycle and state), a tool/action invocation framework (letting agents call external APIs, retrieve from knowledge bases, and operate SaaS systems), memory and context management (preserving session state across turns), and identity and permission controls (authentication when an agent accesses enterprise systems on behalf of a user). For enterprises, AgentCore's value lies in eliminating the work of building an agent framework from scratch — teams don't have to implement common capabilities like tool scheduling, session orchestration, and failure retries themselves, and can focus directly on knowledge base organization, Jira integration, Slack interaction, and other parts that truly reflect business differentiation. This is also the key prerequisite that let ShipStation Global move Deckhand from project kickoff to launch in a relatively short time.

RAG and the "Curated Knowledge Base" Constraint

The reason Deckhand emphasizes "answering from a curated knowledge base" rather than letting the large model improvise is essentially that it adopts a RAG (Retrieval-Augmented Generation) architecture. RAG works like this: first vectorize the user's question, retrieve the several most semantically relevant passages from a pre-built knowledge base, then pass those passages as context to the large model to generate an answer. This has two direct benefits: first, it significantly reduces "hallucination," because the model's answer is anchored to real documents; second, the knowledge is maintainable — when the IT team updates a document, the agent's answers update immediately, with no need to retrain the model. The word "curated" matters just as much — if the knowledge base itself is messy, outdated, or self-contradictory, the retrieved context will mislead the model, and output quality will actually decline. For a scenario like an IT help desk, where answers are relatively standardized but accuracy requirements are high, a controllable knowledge source determines the final result more than model size does — and this is the key design tradeoff behind Deckhand's high feedback rate.

What the Numbers Say

97.5% of channel questions resolved in the first three months, and a 96% positive feedback rate. Taken together, these two figures mean more than either one alone:

A high resolution rate could simply mean problems were passed elsewhere; but a high feedback rate means employees actually approve of the resolution itself. In other words, Deckhand not only caught the volume — it caught the quality.

For internal IT teams, the implication is direct: once a large share of repetitive inquiries is automated, manpower can shift to more complex problems that require human judgment. IT support is no longer a "queue up and wait for a reply" experience, but a "send a message and it's handled" experience.

How to Interpret the 97.5% and 96% Metrics

When reading this kind of agent deployment data, pay attention to how the metrics are defined. "Resolution rate" usually refers to the share of questions that reach closure without being escalated to a human, but how "resolved" is determined (whether the employee confirms it, whether the question simply stops being asked, or whether the agent marks it itself) can significantly affect the number. More important is how it pairs with the "positive feedback rate": if the agent simply converts a large number of questions into Jira tickets across the board, the resolution rate could still be very high, but that counts as "passing them along" rather than "answering them correctly," and the employee experience wouldn't improve. A 96% positive feedback rate indicates that a substantial share of employees are satisfied with the result itself, not just that the process felt smooth. Still, be mindful of survivorship bias — employees who proactively give feedback in the Slack channel may be more inclined toward positive ratings to begin with, and the first three months tend to be a honeymoon period when the knowledge base is maintained most diligently and question types are most concentrated. So these two figures are better treated as evidence that the direction is right, rather than as promised values that can be extrapolated directly to a long-term steady state.

A Reference Sample for Enterprise AI Agents

ShipStation Global's Deckhand offers a pragmatic sample of enterprise AI agent deployment: it doesn't pursue general intelligence, but focuses on a specific scenario (IT help desk), using a curated knowledge base to constrain output quality, screenshot recognition to expand input methods, and Jira integration to close the workflow loop.

This approach can be transplanted to other internal support scenarios — HR inquiries, expense reimbursement, ops Q&A — anywhere the domain is "high-frequency, repetitive, and has standardized answers" is a place where AI agents can make a difference. And as infrastructure like AgentCore matures, the time window from project kickoff to launch for such projects is clearly shrinking.

One-line Summary

ShipStation used Deckhand, a Slack agent built on Bedrock AgentCore, to resolve 97.5% of IT questions in three months.

Paragraph Summary

ShipStation Global's engineering team built Deckhand, an AI IT help desk agent in Slack, running on Amazon Bedrock AgentCore. It has three core capabilities: retrieving answers from a curated knowledge base, recognizing error screenshots uploaded by employees, and automatically creating Jira tickets when the knowledge base doesn't cover a problem. In its first three months, Deckhand resolved 97.5% of questions in the channel and received 96% positive feedback. The article argues that the combination of a high resolution rate and a high feedback rate is especially important — the former shows it caught the volume, the latter shows it caught the quality. This case offers a pragmatic sample of enterprise AI agent deployment: not pursuing general intelligence, but focusing on a specific scenario, using a curated knowledge base to constrain output, screenshot recognition to expand input, and Jira to close the workflow loop — an approach that can be transplanted to high-frequency, repetitive scenarios in HR, finance, and ops.