Tray AI Demonstrates Live Automation Agent for IT Workflows

Tray AI 演示面向 IT 工作流的自动化 Agent 通过对本地知识源进行向量化并执行实时 Schema 查询，企业级 AI Agent 无需手动字段映射即可跨平台即时解决支持工单。 Stephen Stouffer 演示了如何使用 Tray.ai 平台从零构建并配置自定义 AI Agent。他首先展示了如何通过向量化的 Google Drive 文档为未配置的 Agent 提供知识支撑，从而对会议详情和展位位置实现快速且准确的语义检索。随后，他进行了一场 IT 服务管理（ITSM）实时演示，现场观众通过短信成功提交并更新了 Jira 支持工单。最后，在问答环节中，Stouffer 详细阐述了该平台如何动态处理自定义系统模式（Schema）、利用用户元数据，并与各类企业应用程序实现无缝连接。 Stephen Stouffer 现场演示在 Tray AI 中从零构建名为“AI4 agent”的自定义 Agent，展示了涵盖角色设定、系统指令、企业防护栏以及 IT 安全批准或原生模型选择的 Agent Details 工作区。 (0:04) 通过将 Google Drive 文档进行本地分块与向量化嵌入，Agent 成功实现秒级语义检索并精准返回展位与团队信息，有效避免了直接调用实时 API 导致的高延迟与配额消耗。 (1:54) 在现场短信至 Jira 的 ITSM 演示中，观众扫码发短信创建工单并回复问题已解决后，Agent 自动将工单 141 从“In Progress”流转至“Done”，展现了日常 IT 运维的自动化提效能力。 (5:04) 面对跨企业系统定制需求，Stouffer 说明 Agent 可利用 JQL 动态感知 Jira 新建字段而无需手动重新映射，并支持借助 Tray AI 的 REST 或 SOAP 连接器无缝集成任意第三方平台。 (8:36)

Stephen Stouffer demonstrates that making enterprise AI agents operational requires pairing language models with vectorized internal documentation and dynamic, API-driven execution across platforms like Jira without manual schema remapping. Agent Configuration and Vectorized Knowledge Retrieval (0:04) Building an agent from scratch in Tray.ai begins in the Agent Details workspace, where administrators define the agent's role, corporate guardrails, branding guidelines, and model selection. Organizations can choose IT security-approved models or Tray's native model at no additional cost. To enable operational utility, the workspace coordinates data ingestion, executable enterprise tools, interaction channels (such as Slack, Microsoft Teams, or SMS), and diagnostic execution logs that record reasoning steps and tool selection in real time. To prevent hallucinations and supply domain context, agents must be grounded in specialized data. In an unconfigured state, Stouffer's demonstration agent cannot answer where the company's booth is located. Once connected to a Google Drive data source containing an event schedule PDF, the platform automatically chunks the text and builds local vector embeddings. Stouffer notes that this local vector search avoids the severe latency and heavy API consumption associated with calling Google Drive's live APIs for every user query. When re-queried, the grounded agent immediately identifies Booth 223 and provides contact information for speakers Stephen Stouffer and Adan Potyondy. Real-Time IT Support Ticket Automation (5:04) To demonstrate autonomous action execution, Stouffer connects an IT Service Management (ITSM) agent directly to Jira and opens an SMS intake channel via a displayed QR code. Audience members simulate common IT support scenarios—such as hardware failures, software access needs, or password resets—directly from their mobile phones. Stouffer explains that the primary objective of an ITSM agent is automating routine ticket remediation and preliminary troubleshooting, sparing support staff from repetitive tasks like reminding employees to reboot machines. The demonstration confirms bidirectional synchronization between the conversational front end and the Jira board. When an audience member texts the agent stating that their reported issue is resolved, the agent processes the message and automatically transitions Ticket 141 from the In Progress column to Done. The underlying logic handles status updates, automated commenting, and SOP-based diagnostic workflows, while retaining the ability to escalate unresolved issues to human engineers. Dynamic Schema Handling and Enterprise Integrations (8:36) Addressing audience questions regarding custom Jira setups, Stouffer explains that the agent evaluates user metadata—such as department, manager, and email address—alongside AI classification to route tickets to the appropriate boards or Epics. The system queries Jira using Jira Query Language (JQL) and dynamically inspects system schemas, allowing it to read or populate newly added custom fields without requiring manual re-mapping by administrators. Stouffer emphasizes that the platform remains application-agnostic, supporting integrations through native connectors as well as custom REST or SOAP endpoints. To ensure operational safety, administrators can enforce security guardrails by restricting destructive capabilities, such as delete permissions. Furthermore, the agent conducts conversational requirements gathering: when faced with vague requests like an unpowered computer, it interrogates the user for model numbers and error codes to construct a complete, SOP-compliant ticket before submission. Adan Potyondy notes that while SMS provided an accessible demonstration interface for a live conference audience, enterprise deployments typically sit directly inside collaboration environments such as Slack and Microsoft Teams to resolve workplace support requests.

Stephen Stouffer 在 Ai4 大会上展示了 Tray AI 构建自主智能体的工程实践，指出通过本地向量化数据基底、动态架构自适应以及多通道集成，企业能够在无需人工字段映射的前提下实现 IT 服务管理与业务流程的自动化闭环。 智能体核心架构与企业安全配置 (0:04) Stouffer 演示了在 Tray.ai 平台中从零构建定制化智能体（AI4 agent）的完整流程。智能体的工作区被明确划分为核心身份设定、数据源接入、工具集调用、交互通道以及测试日志五个功能模块。在“智能体详情”（Agent Details）区域，企业可以为其设定专属角色、系统运行指令、品牌规范以及企业护栏，以确保模型输出符合合规要求。在底层大语言模型选择上，企业既能直接指定通过内部 IT 安全团队审查的外部模型，也可以无额外成本调用 Tray 平台内置的原生模型。 在架构协同上，数据源模块负责解析标准作业程序（SOP）等知识文档，工具模块则赋予智能体执行跨系统查询、编辑、新建记录或触发 REST/SOAP 请求的操作能力。平台配套的测试沙箱与执行日志能够完整回溯智能体的每一步推理逻辑与工具选取过程，为企业审计智能体的自主决策行为提供了透明依据。 本地向量化检索与低延迟知识接地 (1:54) 为了说明知识接地的必要性，Stouffer 首先测试了未配置上下文的空白智能体，当询问“Tray AI 展位在哪里”时，该智能体因缺乏数据而明确回复不掌握相关信息。随后，演示者接入包含会议日程与展位信息的 Google Drive PDF 文档。平台在后台自动完成文本提取、内容切块并生成本地向量嵌入，使智能体具备了语义检索能力。 Stouffer 强调，若智能体每次响应查询都向 Google Drive 发起实时原始 API 调用，会引入极高的网络延迟并迅速耗尽接口配额；而在本地构建向量化检索流程，能在确保事实准确性的同时大幅缩短响应时间。同步完成后，智能体即刻准确指明 Tray AI 位于 223 号展位，并输出了现场演讲者 Adan Potyondy 与 Stephen Stouffer 的姓名及联络方式。全链路日志同步展示了智能体结合工具与知识库的思维链演进过程。 基于短信与 Jira 的工单自动流转 (5:04) 在实际业务场景演示中，Stouffer 呈现了一个直连 Jira 的 IT 服务管理（ITSM）智能体。现场观众通过扫描屏幕上的二维码自动生成短信，模拟员工提交密码重置、权限申请或硬件故障等典型支持工单。Stouffer 表示，在服务台前置部署此类智能体旨在承揽重复性的初步故障排查与工单修复工作，避免技术人员耗费时间提醒员工执行重启等基础排错步骤，从而将人力释放至核心项目。 现场观众提交工单后获得了 141 号工单编号。随后该观众通过短信向智能体发送问题已解决的确认信息，智能体即时处理文本，并在现场大屏的实时 Jira 看板上自动将 141 号工单从“进行中”（In Progress）状态拖动至“已完成”（Done）。Stouffer 说明，选择短信是考虑到现场观众缺乏电脑操作条件的灵活性安排，该套底层逻辑同样适用于 Web 门户、Slack 或 Microsoft Teams 等日常企业通信场景。 动态模式自适应与异构系统集成 (8:36) 针对各企业在 Jira 等系统中高度差异化的看板结构、Sprint 设定与自定义工作流，Stouffer 介绍了智能体的自动化适应机制。在企业协作工具中，智能体会抓取提单用户的部门、直属主管和邮箱等上下文元数据，并运用 AI 分类机制判别请求类型（例如权限获取或入职支持），自动将任务分流至正确的看板或 Epic。 在底层架构层面，智能体利用 Jira 查询语言（JQL）动态读取系统模式（Schema）。当管理员在 Jira 中新增自定义字段时，智能体在检索阶段能实时感知该结构变化，无需工程人员重新进行手动字段映射即可直接读写新字段。在安全性方面，管理员拥有完全的权限控制力，可主动屏蔽删除等高危操作。此外，Tray AI 除预置原生连接器外，还提供通用的 REST 和 SOAP 连接器以兼容任意目标系统。面对“电脑无法开机”等模糊输入，智能体还会主动多轮追问缺失的设备型号与报错信息，将其整理为符合企业 SOP 规范的需求文档后再行归档。 Adan Potyondy 在总结中说明，现场演示虽以短信呈现，但企业实际部署中最普遍的接入点仍是 Slack 与 Microsoft Teams。该方案展示了企业级智能体自动化执行任务的技术可行性，但其实际运行效果仍取决于企业内部 SOP 文档的规范程度，且业务稳定性依赖于管理员对底层接口权限（如删除限制）的严格约束。

## Background 该视频由 Ai4 频道于 2025 年 8 月 27 日发布在 YouTube 平台上。内容由 Tray.ai 销售副总裁 Adan Potyondy 与自动化解决方案总监 Stephen Stouffer 共同呈现，主要展示如何在 Tray.ai 上从零现场构建 AI 智能体，并结合 Slack、Jira 和 Salesforce 等业务系统实现事件触发与企业级落地部署。 ## TL;DR 视频中，Stephen Stouffer 现场演示了从空白画布构建企业级智能体的全流程，详细解析了角色设定、合规模型选择、数据源与受控工具链等核心组件。随后通过接入 Google Drive 资料实现本地 RAG 向量化问答，并邀请现场观众通过短信触发 ITSM 智能体，实时完成 Jira 工单的创建、自动推进与状态闭环。最后，演讲者结合现场答疑深入探讨了 Agent 原生利用 JQL 进行动态 Schema 探测、智能多租户路由以及通过多轮对话主动补全工单信息的能力。 ## Atmosphere & style 整场演讲风格硬核且注重现场实操，摒弃了空洞的理论宣讲，直接采用“现场从零搭建”的快节奏演示。观众通过手机发送短信与大屏幕上的 Jira 看板形成实时联动的互动体验，让 AI 规划与工具调用的过程变得生动具体，现场问答环节更展现出直面企业落地痛点的务实与专业。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=4s) Tray.ai 智能体由六大核心模块构成：角色设定与护栏、AI 模型接入、知识数据源、外部系统工具、交互渠道以及测试与可观测性日志。 - [01:54](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=114s) 数据源接入时平台自动执行分块与本地向量化（Vectorizing）存储，避免高频实时调用外部 API 带来的接口负载与网络延迟。 - [01:54](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=114s) 智能体具备多工具自主串联调用能力，可根据任务需求在不同业务系统中自主执行查找、创建或更新记录，无需人工手动映射字段。 - [05:08](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=308s) 现场演示展示了 ITSM 智能体通过 SMS 与 Jira 看板实时联动，用户确认问题解决后工单自动从“进行中”流转至“已完成”。 - [08:51](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=531s) 智能体通过交互渠道识别用户组织画像，并结合大模型意图分类，能动态将工单路由归属到对应的看板、Epic 或多租户规则中。 - [10:05](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=605s) 智能体原生掌握 JQL（Jira 查询语言）并可动态识别实时 Schema，底层新增自定义字段时无需管理员手动重新映射元数据。 - [10:05](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=605s) 企业可通过限制智能体可调用的工具范围构建权限安全护栏，精细控制自由度以防止 AI 出现误建或误删数据的越轨行为。 - [11:16](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=676s) Tray.ai 具备通用 API 连接能力，不仅内置原生连接器，还支持 REST 和 SOAP 协议对接任意企业级历史与现有系统。 - [11:16](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=676s) 面对描述模糊的初始报障，智能体会基于企业 SOP 护栏展开多轮拟人化追问，主动补全硬件型号与报错代码等关键信息。 - [11:16](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=676s) 在实际企业落地中，智能体主要无缝嵌入 Slack 或 Microsoft Teams 等员工核心协同界面，降低使用门槛并顺畅完成业务流转。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=4s) 从零搭建 Tray.ai 智能体：核心组件与能力解析 在演示开始时，演讲者 Stephen Stouffer 展示了 Tray.ai 平台的项目构建界面。平台本身提供了诸如 ITSM 智能体、知识库智能体等预置加速器（Accelerators），可以直接拉取所需的数据并完成知识摄取；但他选择走一条“更硬核的路线”——直接从空白项目从头创建，并将其命名为 `AI4 agent`。 ```mermaid flowchart LR accTitle: Tray.ai 智能体核心架构组成 subgraph Core["智能体核心配置 (Agent Core)"] direction TB A["Agent Details<br/>(角色设定 / 护栏 / 品牌规范)"] B["AI Model<br/>(合规模型 / 原生免费模型)"] end subgraph DataAndTools["知识与能力层"] direction TB C["Data Sources<br/>(SOP / 营销资料 / 向量化)"] D["Tools<br/>(查找 / 创建 / 更新系统记录)"] end subgraph ChannelAndOps["交互与运维"] direction TB E["Interaction Channels<br/>(Slack / Teams / SMS)"] F["Logs & Observability<br/>(Prompt / 推理链路 / 工具链)"] end Core --> DataAndTools --> ChannelAndOps ``` Stephen 逐一拆解了构建一个企业级智能体所必备的核心模块： - **Agent Details（智能体详情）**：用于设定智能体的角色定位、工作任务、安全护栏（Guardrails）以及企业品牌沟通准则。 - **AI Model（模型选择）**：支持灵活切换由企业 IT 安全团队审批的特定大语言模型；平台也自带开箱即用的原生模型，无需额外调用成本。 - **Data Sources（数据源）**：用于知识注入与训练，支持直接导入内部标准作业程序（SOP）文档、营销物料等结构化与非结构化数据。 - **Tools（工具集成）**：赋予智能体“采取行动”的能力，包括在业务系统中查询信息、创建或修改数据记录。 - **Interaction Channels（交互渠道）**：支持将智能体部署至 Slack、Microsoft Teams、电子邮件等。由于现场观众多使用手机而非笔记本电脑，本次演示选择以 SMS 短信作为主要交互渠道。 - **Testing & Logs（测试与可观测性日志）**：提供沙盒测试环境及详细的调用日志，用于追踪 Prompt、推理链（Reasoning）及多工具调用过程。 --- ### [01:54](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=114s) RAG 向量化与上下文注入：让智能体掌握专属知识 为了展示智能体在缺乏知识时的状态，Stephen 首先进行了一次测试，直接提问："Where is the Trey AI booth?"（Tray.ai 的展位在哪里？）。由于此时智能体没有任何上下文信息，只能礼貌地返回："I apologize. I don't have any information"（抱歉，我没有任何相关信息）。 为了赋予其上下文，Stephen 在数据源中接入了一个包含本次 AI4 大会资讯及 Tray.ai 参展信息的 Google Drive 文件夹中的 PDF 文件。当保存配置后，系统会自动拉取文件并进行分块（Chunking）与向量化（Vectorizing）处理，生成专属的向量检索表。Stephen 强调，如果每次请求都直接向 Google Drive 发起实时 API 调用，不仅接口调用量巨大，而且网络延迟极高；通过本地向量存储进行检索，能大幅提升响应速度。 同步完成后再次发起相同的提问，智能体立即准确回答出展位号为 **223 号展位**，并附带了现场负责人 Adan Potyondy 与 Stephen 自身的联系方式。在后台日志中，系统清晰记录了智能体的 Prompt、决策推理过程以及所调用的工具组合。智能体能够根据任务需要，自主、多次、串联调用各类工具（例如在 ServiceNow 或 Salesforce 中先执行查找记录工具，再执行创建或更新工具），无需人工进行繁琐的字段映射。 --- ### [05:08](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=308s) 现场实时互动：基于 Jira 的 ITSM 自动化服务台 随后，Stephen 展示了在大会前配置好的 ITSM 智能体。该智能体与 Jira 系统打通，旨在作为企业 IT 团队的前置防御层，自动分担并解决诸如软件申请、权限获取、密码重置等常见问题，从而将 IT 人员从“拔掉电源再插上”这类琐碎的重复劳动中解脱出来。 ```mermaid flowchart TD accTitle: ITSM 智能体自动化工单生命周期 subgraph Trigger["用户接入层"] U["现场观众 / 员工"] -->|扫描二维码 / 发送短信| SMS["SMS 交互渠道"] end subgraph AgentLogic["智能体推理与知识库"] SMS --> A["ITSM 智能体"] KB[("企业 SOP / 解决手册")] <--> A end subgraph JiraActions["Jira 自动化执行"] A -->|创建新工单| T1["Jira: In Progress 状态"] SMS -.->|回复'问题已解决'| A A -->|自动更新状态| T2["Jira: Done 状态"] A -.->|复杂/无法解决| ESC["转派 IT 人工处理"] end ``` Stephen 邀请现场观众掏出手机扫描大屏幕上的二维码，通过预填或自拟的短信向系统发送 IT 求助请求。现场观众发送了一条测试请求后，工单系统立即创建了对应的 Jira 任务。 - 观众反馈其工单编号为 **141**。 - Stephen 调出 Jira 看板，当该观众在短信中向智能体回复确认“问题已解决”后，大屏幕上的 141 号工单立即实时从 **“In Progress”（进行中）** 列自动流转到了 **“Done”（已完成）** 列。 - 智能体不仅能跨生命周期阶段推进工单，还能自动添加处理备注、提供排错方案；一旦遇到超出能力范围的疑难杂症，则会自动转派给对应的人工 IT 专员。 --- ### [08:51](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=531s) 现场答疑：应对复杂与定制化业务流程的智能路由 在演示结束后，一位现场观众提问指出：各家企业的 Jira 落地实现方式大相径庭，有的使用独特的看板，有的划分不同的 Sprint 周期，Tray.ai 如何支持这种高度个性化的定制需求？ Stephen 对此给出了明确的架构落地思路： 1. **利用交互渠道获取上下文身份**：在实际生产环境中，智能体通常嵌入在 Slack 等内部办公软件中。当用户触发智能体时，系统可以直接获取其部门归属、汇报上级、企业邮箱等基础画像信息。 2. **AI 智能意图分类**：利用大模型对工单内容进行语义分类，判断其属于 IT 权限请求、内容物料需求还是新员工入职流程。 3. **动态路由与关联**：结合提问者的部门身份与分类结果，自动将工单归属到对应的 Board、Epic 或具体项目中；若面向外部客户，则可通过其邮箱域名或公司名称进行多租户规则匹配与分流。 ### [10:05](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=605s) Jira 动态查询与实时 Schema 识别 Stephen Stouffer 详细演示了 AI Agent 如何与 Jira 系统进行深层交互。该 Agent 并非机械地执行死板脚本，而是具备直接使用 JQL（Jira Query Language，Jira 查询语言）检索系统的能力。这意味着 Agent 能够自主定位与特定客户关联的看板，并在正确的板块下创建或更新 Ticket；无论是检索特定看板、史诗（Epic）还是普通事务（Issue），它都能灵活应对各种内部或外部业务场景。 现场观众随即提问，确认 Agent 是否直接拥有对系统数据模式（Schema）的访问权限。Stephen 证实了这一点，并指出在当前配置下 Agent 确实能读取 Schema（尽管企业出于安全策略也可以选择限制该权限）。这种架构带来了显著优势：若团队在 Jira 中新增了自定义字段，管理员根本无需重新手动配置映射关系或录入字段元数据，Agent 在实时检索并获取 Issue 时即可动态识别新字段，并根据权限对其进行填充或更新。Stephen 强调，Agent 的自由度完全可控，既可以配置得极为灵活，也可以严格受限，管理员通过限制其能够调用的具体工具，即可有效防止 AI “越轨乱来”（"go a little rogue"），避免出现误建或误删数据的风险。 ```mermaid flowchart TD accTitle: Agent 与系统动态交互流程 subgraph Trigger["触发与解析"] A["事件输入 / 客户请求"] --> B["Agent 规划与 JQL 生成"] end subgraph Execution["动态查询与执行"] B --> C["实时拉取 Schema 与字段"] C --> D["按 SOP 约束创建 / 更新 Ticket"] end subgraph Tools["受控工具层"] D -. "权限受限调用".-> E["Jira / ServiceNow 等系统"] end ``` - **JQL 原生驱动**：Agent 原生掌握 Jira 查询语言，可在大型组织复杂的项目看板与 Issue 树中自主定向检索。 - **实时模式适应（Schema-agnostic）**：动态探测底层自定义字段变更，免除传统工作流集成中频繁的手动字段映射。 - **权限安全护栏**：通过微调暴露给 Agent 的 Tooling 范围，锁死写操作与删除权限，防止不可逆误操作。 --- ### [11:16](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=676s) 通用 API 连接与对话式工单智能补全 针对观众关于平台是否仅局限于 Jira 或 ServiceNow 这类常见系统的疑问，Stephen 斩钉截铁地回答，任何具备 API 的系统均可无缝接入——"anything with an API"。Tray.ai 平台除了内置大量开箱即用的原生连接器外，还提供了高度灵活的通用 REST 连接器，甚至企业若有旧系统需求，也完全支持对接传统的 SOAP 协议，从而允许团队将自主工具箱拓展到任意企业级软件。 随后现场有人提出关于工单有效性的疑问：如果系统检索的工单历史库中只有故障描述而缺少现成的解决方案，Agent 能否正常工作？Stephen 解释道，虽然训练信息“越丰富越好”，但 Agent 本身具备强大的交互引导能力。在本次现场演示中，当参会者通过短信（SMS）提交需求时，Agent 会自动将零散的文本归纳提炼为标准的业务需求格式。如果用户提供的信息不全（例如仅仅发一句“我的电脑无法开机”），Agent 不会草率立项，而是会基于企业标准作业程序（SOP）文档中设定的安全护栏展开多轮对话，主动追问：“好的，请问您的电脑型号是什么？这台设备使用了多久？屏幕上是否有具体的错误提示代码？”通过拟人化的沟通逐步收集补全工单所需的核心要素。 Adan Potyondy 补充道，尽管现场为了方便观众互动采用了 SMS 短信作为触发端，但在实际企业落地项目中，绝大部分客户构建的都是对话型 Agent，直接嵌入到员工日常工作的核心阵地——Slack 或 Microsoft Teams 之中，让工单处理与业务流转在员工最习惯的即时通讯界面中顺畅完成。 - **泛协议集成能力**：支持 REST、SOAP 及各类 Native Connectors，只要系统暴露 API 即可封装为 Agent 可调用的 Tool。 - **对话式信息补全**：面对语焉不详的初始报障，Agent 能依据 SOP 规则主动追问硬件型号、故障时长与报错信息，完成结构化归纳。 - **工作流原生嵌入**：除短信场景外，企业级方案主要围绕 Slack 和 Microsoft Teams 构建，贴合员工日常协同习惯。 ## Notable quotes > [01:54](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=114s) Where is the Trey AI booth? — Stephen Stouffer > [01:54](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=114s) I apologize. I don't have any information — AI4 agent > [05:08](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=308s) 拔掉电源再插上 — Stephen Stouffer > [05:08](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=308s) In Progress > [05:08](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=308s) Done > [10:05](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=605s) go a little rogue — Stephen Stouffer > [11:36](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=696s) anything with an API — Stephen Stouffer > [11:36](https://www.youtube.com/watch?v=Qgsh1K-nt48&t=696s) 好的，请问您的电脑型号是什么？这台设备使用了多久？屏幕上是否有具体的错误提示代码？ — Stephen Stouffer ## Mentioned Tray.ai, Stephen Stouffer, Adan Potyondy, Ai4, Jira, Slack, Salesforce, Microsoft Teams, ServiceNow, Google Drive, JQL (Jira Query Language), REST, SOAP, SMS ## Video description Adan Potyondy, VP of Sales, Tray.ai Stephen Stouffer, Director, Automation Solutions, Tray.ai Abstracts The Tray.ai team is going hands-on to show what agent adoption actually looks like—built live, triggered by real events, and connected to tools like Slack, Jira, and Salesforce. You’ll follow exactly how the agent is triggered, how it decides what to do, and how it completes work in real systems. We’ll walk through how teams are deploying agents across departments, managing adoption, and staying in control as use cases scale. If you’re comparing platforms or building your agent strategy, this session will help you ask better questions and spot the gaps most vendors won’t mention. [Source](https://www.youtube.com/watch?v=Qgsh1K-nt48)

[1:54] Stephen Stouffer: Where is the Trey AI booth? [1:54] AI4 agent: I apologize. I don't have any information [5:08] Stephen Stouffer: 拔掉电源再插上 [5:08] In Progress [5:08] Done [10:05] Stephen Stouffer: go a little rogue [11:36] Stephen Stouffer: anything with an API [11:36] Stephen Stouffer: 好的，请问您的电脑型号是什么？这台设备使用了多久？屏幕上是否有具体的错误提示代码？

[1:54] Stephen Stouffer: Where is the Tray AI booth? [8:36] Stephen Stouffer: The tooling that you create within the Tray platform... we have a lot of native connectors, but we also have a very flexible REST connector so you can connect into anything REST-based—or SOAP if that's what you want to do. [8:36] my computer won't turn on

[0:04] new projects. Uh we have the ability to [0:06] pick from some accelerators here. So uh [0:09] Aiden mentioned an ITM agent as well as [0:11] like a knowledge agent. I could select [0:13] from one of these accelerators in which [0:15] case it's going to pour over everything [0:16] I need to ingest knowledge and uh build [0:20] an agent uh within Trey. I'm going to [0:22] take the hard path and just build it [0:24] straight from new. We're going to call [0:25] it AI4 [0:28] agent. [0:33] Give it a second to load here. And then [0:35] what I'm uh presented here with is the [0:38] agent details. So you can think about [0:39] this section of how the agent should [0:42] think about its task. What's the role [0:45] that it's playing? any sort of guard [0:46] rails or uh you know uh branding [0:50] guidelines of how you talk about your [0:52] company and products, you can just drop [0:53] it right here within the agent details. [0:55] You can also pick uh an AI model. So if [0:58] there's a specific model that's approved [0:59] from your IT security team, you can [1:01] select that model, but out of the box uh [1:03] it will be using our own native uh model [1:06] that doesn't cost you any extra money. [1:09] We also have data sources. I'll get into [1:11] that here in a little bit. Um, this is [1:13] where you can start to train the agent [1:15] on knowledge. So, if you have internal [1:17] SOP documentation or marketing material, [1:19] you can just dump it into the agent. Uh, [1:22] and then there's tools. This is the [1:24] ability to do things. So, look up [1:26] information, uh, make edits into your [1:28] systems, create records, update records, [1:30] all that good stuff. Interaction [1:32] channels. This is where you can deploy [1:34] your agent within Slack or Teams or [1:36] email or in this case we're going to be [1:38] doing an SMS demonstration because I [1:41] don't see a lot of laptops here, but I [1:42] see a lot of phones. Uh, and then uh the [1:45] ability to test it. Uh, and then also we [1:48] have some robust logs um here, which [1:51] there are no logs here, but we're going [1:52] to see them here in a second. So, I'm [1:54] going to test this agent that I just [1:56] created. It is uh a dumb agent. has no [1:59] information about uh you know who we [2:02] are, where we're at, and all that [2:04] information. So, I'm just going to ask [2:05] it uh where is the Trey AI booth? It's [2:11] not going to know. It hasn't it has no [2:13] context uh as to um AI4 or you know our [2:17] involvement within AI4. So, it's going [2:19] to come back to say I apologize. I don't [2:21] have any information [2:23] uh you know all that good stuff here. So [2:26] what I want to do is I want to ground it [2:27] in context. So what I'm going to do is [2:29] I'm going to add a data source. [2:32] I'm going to do a Google Drive data [2:33] source [2:37] and I have a PDF that just outlines [2:40] information about the conference as well [2:42] as uh what Trey is doing within the [2:45] conference. So I'm just going to select [2:46] the folder, [2:50] save it. What it's going to do when I [2:52] save this is it's going to create a data [2:54] source connected to Google Drive. It's [2:56] going to pull out all the information [2:58] within that folder within Google Drive. [3:00] It's going to chunk it, vectorize it, [3:03] and then the agent is going to be able [3:04] to make queries against that [3:06] information. So, if it had every single [3:08] request were to come in doing an API [3:10] call to Google Drive, that's a lot of [3:12] API calls. It's a lot of data. It can be [3:14] slow. The latency is bad. Uh but in this [3:17] case when we create a vector storage uh [3:19] table for it to query against it can [3:21] move very quick. Uh so we'll give that a [3:23] second to uh to load here. Let me check [3:25] to see if it's uh finished. We'll give [3:28] it a couple of seconds syncing and then [3:30] what I'm going to be able to do is I'm [3:32] going to have to run another test and I [3:33] can ask it information and then it [3:35] should be grounded in that uh you know [3:37] information about AI4 as well as what [3:40] Trey is doing here. So uh we'll give [3:42] that a couple more seconds. Looks like [3:44] it's complete. ask the same question. [3:46] Where is the Trey AI booth? [3:52] Not should know. So, it should be it [3:55] should understand who we are uh as well [3:57] as where the booth uh is, which is booth [4:00] number 223 if anyone was wondering. [4:03] We'll give it a second. [4:05] So, there we go. Booth 223. It also has [4:08] information about uh Aiden and myself [4:10] and our contact details. Uh so this [4:13] agent can now be deployed in Slack, [4:15] Teams, wherever and then people can [4:17] interface with it, ask questions about [4:19] AI4, where we are, how to book a demo, [4:21] all all that good stuff. Uh so that is [4:25] building basically an agent from [4:27] scratch. Uh and then because I ran some [4:28] tests, we're going to also have our [4:30] logs, uh which logs what the prompt was, [4:33] what the reasoning was, what tools were [4:35] selected to complete the task. And the [4:37] really cool thing about our tools is uh [4:40] the agent has the ability to pick any of [4:41] these tools, use them together, and use [4:44] them as many times as needed to complete [4:45] a task. So if you use uh something like [4:48] Service Now or Salesforce, it will be [4:50] able to use a lookup record tool with a [4:52] create or an update tool uh without [4:54] really the user needing to, you know, [4:56] pick and select or map uh data from [4:58] different objects. So that's creating an [5:00] agent basically from scratch. I know [5:02] it's not too exciting, but um if you [5:04] have your phone with you, I welcome you [5:06] to pull it out. [5:08] Ahead of the conference, I built an [5:10] agent [5:12] uh tied to Jira. So, this is an ITSM [5:14] agent. You scan the QR code. I'll expand [5:17] this a little bit bigger for you. Uh if [5:19] you scan the uh QR code, you're going to [5:21] get a link. If you click the link, uh a [5:24] text message is going to autopop [5:25] populate for you. You are welcome to [5:27] completely delete that text and create [5:30] your own text message, but this is going [5:32] to be you creating a support ticket for [5:34] our IT team. So, put put your end user [5:36] hat on things like access request, [5:38] software requests, uh computers on fire, [5:42] you need to reset your password. So, [5:44] this agent is trained to help you solve [5:46] business problems. And the point of this [5:47] agent is to sit in front of your IT team [5:50] uh remediate those tickets, work those [5:52] tickets ahead of uh your sales team so [5:56] they can work on pushing the company [5:58] forward, solving big important problems, [6:00] not things like unplug your computer, uh [6:03] plug it back in, turn your computer on. [6:05] So this is an an IT agent. Uh so if you [6:08] want to make up your own kind of [6:10] fictitious problem that you have within [6:11] your company, I have trained this agent [6:14] to help solve problems. So uh it might [6:18] try to help you solve your your ticket. [6:20] Um it is an agent, so it's going to be a [6:22] little bit different for everybody based [6:23] on your request. Uh and then uh if you [6:27] submit that, I'm going to pull up Jira [6:29] here and we're going to see some of the [6:32] stuff that has been coming in. So, who [6:34] submitted a ticket where you have a [6:36] ticket number that you'd be okay with me [6:39] pulling up on my computer? All right. [6:41] What's your ticket number? [6:44] >> 0 what? [6:47] >> 141. All right. 14. Oh, right here. [6:53] So, what I'm going to ask you to do is [6:56] uh respond back to the agent and tell [6:58] the agent that you've solved the [7:01] problem, like you're good to go, [7:02] everything's okay. And we're going to [7:04] see one ticket here um move from the in [7:08] pro in progress stage over to the done [7:11] column. And this is where the agent has [7:13] the ability to not only work tickets, [7:16] but move them into different stages uh [7:19] of uh ownership. and let's say the agent [7:22] couldn't solve something, it could then [7:24] assign it out to your IT team. So, [7:26] you're going to have basically agents uh [7:28] have the ability to protect your IT team [7:32] or you know, whatever team uh that is is [7:36] uh you know working tickets and allowing [7:39] them to free up their space. So, here [7:40] you can see the one one ticket. Now, if [7:42] everyone else wants to complete your [7:43] tickets, you're welcome to. Uh but this [7:45] agent has the ability to add comments, [7:47] update the statuses, um as well as uh [7:51] provide different solutions and you can [7:53] train it on all of your internal SOP [7:55] documentation. So let's say there is a [7:59] constant issue or a proprietary thing [8:01] that other people have to uh put into [8:03] the system or certain people that that [8:05] need to be contacted for different [8:07] issues. the agent can essentially be [8:09] trained and grounded on that data uh and [8:12] then um you know work with either [8:15] internal IT folks or externally with [8:18] your customers and you could of course [8:20] deploy this wherever your customers are. [8:21] So you could deploy it on the uh on the [8:24] web through a security support for uh [8:27] portal or uh via something like Slack or [8:31] Teams. Uh so this is an IT uh ITSM agent [8:36] and I actually have a microphone up [8:37] here. I don't know if anyone has any [8:38] questions or anything based on what you [8:40] saw. Oh, your ticket. You still you [8:42] still you still have questions. Uh I'm [8:44] going to put you on the spot. I said I [8:46] wasn't going to do questions, but I [8:47] ended up doing it. So, uh if you want to [8:49] bring in the microphone, see if he has a [8:50] question. [8:51] >> This is relevant because this is kind of [8:53] what I'm trying to do for my company [8:55] right now. But [8:56] >> Sure. [8:56] >> The biggest problem is um this Jira [9:00] implementation is different for [9:01] different companies. [9:03] >> Sure. Yep. like they all have their own [9:04] way of implementing. Um some people uses [9:08] different kinds of boards, some people [9:09] use different sprints, all those things. [9:11] Do you [9:12] >> allow for that customization? [9:14] >> Yeah. So it it depends on how you want [9:16] to deploy it. So what I'm doing in like [9:19] a production instance was I I'd actually [9:20] look at the person who's submitting the [9:22] ticket. If it's an internal person, I [9:24] can figure out what team they sit on [9:26] based off of where they're interfacing [9:27] with the agent. So I would typically [9:29] deploy it within Slack. So when someone [9:31] invokes an agent in Slack, you get all [9:34] their Slack information. So what the [9:36] department they sit on, who their [9:38] manager is, what their email address is, [9:40] and with that information, you can kind [9:41] of figure out, you know, what that [9:43] ticket and where it should go. You can [9:45] also classify tickets that get opened [9:47] using AI. So you can look at the ticket [9:49] and say, is it an IT request, is it a [9:52] content request, is it a new hire [9:54] request, you know, what what is it? And [9:56] then you can typically figure out what [9:57] board or Epic or whatever it it should [10:00] be tied to. Or if it's a customer, you [10:02] can use something like their email [10:03] domain or the company name within the [10:05] system to go find the board that is tied [10:07] to that customer and then open the [10:09] ticket on that. Uh so my agent actually [10:12] has the ability to query Jira as well [10:14] using the Jira query language. So it can [10:16] actually search to find different boards [10:19] um or epics or issues or whatever it is [10:22] to go update them or create new ones or [10:24] whatever. So you can do it a few [10:26] different ways like depending if you [10:27] want to do it in an internal use case or [10:29] if it's an external use case it can be [10:31] very flexible. [10:32] >> So it has access to the schema. [10:35] >> Uh yeah well in my case yes so you may [10:38] not want to always provide access but [10:40] yes it would have access to the schema. [10:42] So if someone created like a new custom [10:44] field the agent would be able to know [10:46] that just on the fly. You wouldn't need [10:47] to like map that or like give it that [10:49] information. every time it did did a [10:51] search and found an issue, it would pull [10:54] the custom fields or the information in [10:56] real time. So then it would be able to [10:57] then update it or create it if you [10:59] allowed it to do that. Yeah, the agents [11:02] can be as flexible. In this case, it's [11:04] pretty flexible or as limited like you, [11:07] you know, you can limit the tooling that [11:09] it has access to. So it can't just go [11:11] create, delete, you know, go a little [11:14] rogue. Um so [11:16] >> sorry is this just predicated on two [11:19] types on on Jira and service now or will [11:22] it work with any it [11:23] >> anything with an API [11:24] >> anything with an API [11:25] >> yeah the tooling that you create within [11:27] uh the trade platform uh here I can [11:30] create a a new tool and we have a lot of [11:32] native connectors but we we also have a [11:34] very flexible REST connector so you can [11:36] connect into anything REST based um or [11:39] SOAP if that's what you want to do [11:42] >> in order for the tickets to be effective [11:45] Do they have to have a pretty well [11:47] populated resolution section? Like if [11:51] you're querying a ticket database and [11:53] it's just problems with no resolutions, [11:58] >> I the more information you have, the [11:59] better. Uh so in my case, when I had you [12:02] guys uh submit tickets, I actually had [12:04] the agent summarize what you put in via [12:07] the text message into a format that I [12:10] like a business requirement format. And [12:12] if you didn't provide me all the [12:13] information, let's say someone put in [12:16] like um my computer's not going to turn [12:18] on. If someone wants to submit that, you [12:19] can. It will come back to you and be [12:21] like, "All right, what is your computer [12:23] model number?" Like, "How long have you [12:24] had it? Is there any error messages?" [12:26] And it will actually curate that [12:28] information that's needed for the ticket [12:29] just conversationally. And that and [12:31] that's all based off of the guard rails [12:33] that you set within your SOP [12:35] documentation. So the more documents you [12:38] train it on, the better it will be in [12:40] how it interfaces with other people. [12:42] >> Thank you. [12:43] >> Yeah. [12:44] >> It's also important to point out, right, [12:45] we we do a lot of projects where they're [12:46] more conversational in nature, [12:48] interfacing with um Teams or Slack. [12:51] That's where a lot of people do their [12:53] work and live daytoday. Um, of course [12:55] we're working with SMS today, but