HackerNews中文版

大多数企业工作效率低下的原因，并非数据本身质量差，而是因为访问数据的界面过于分散。一个简单的提问，比如“哪些交易停滞不前？”，需要涉及仪表盘、电子表格、客户关系管理系统（CRM）、商业智能（BI）工具、内部脚本，以及几个Slack线程。而根据答案采取行动，又需要在不同的系统之间切换。这种摩擦存在于中间环节。 Worqlo 正在尝试通过将对话作为界面层，将确定性工作流作为执行层来消除这种摩擦。其理念很简单：自然语言输入 → 验证后的工作流输出。大型语言模型（LLM）负责处理意图。结构化的工作流引擎负责执行：CRM查询、字段更新、通知、权限管理和审计日志。模型本身从不直接执行操作。以下是其工作原理。为什么选择对话？人们用问题思考。系统用模式思考。仪表盘介于两者之间。由于每个系统都暴露了自己的用户界面（UI），因此界面会成倍增加。工程师最终不得不构建内部工具、过滤器、查询、分析页面和一次性自动化。这就是UI税。对话减少了界面。工作流增加了安全性和确定性。架构（简化版）用户 → LLM（意图） → 路由器 → 工作流引擎 → 连接器 → 系统 LLM 提取意图和参数。没有执行权限。意图路由器将意图映射到已知的工作流模板。工作流引擎按顺序执行步骤：模式验证权限检查 CRM查询 API更新通知审计日志连接器针对CRM、ERP、内部API和消息传递系统的严格适配器。如果出现以下情况，工作流引擎将拒绝运行：字段不存在数据类型不匹配权限失败工作流模板与用户意图不匹配这可以防止常见的LLM失败情况：虚构的字段、不正确的API调用、不安全的操作等。示例查询用户： “给我看看本周DACH地区的销售线索” 内部流程： intent = llm.parse("pipeline query") validate(intent) fetch(data) aggregate(stats) return(summary) 后续： “将汉莎航空的交易重新分配给Julia，并提醒Alex跟进” 工作流：按名称查找交易验证所有权变更编写CRM更新发送Slack通知编写审计日志所有操作都通过确定性步骤运行。为什么从销售开始？销售CRM是结构化且可预测的。工作流重复（重新分配、提醒、跟进）。延迟很重要。输出是可衡量的。这使得该领域成为会话式工作流的良好测试环境。长远来看，这并非仅限于销售。同样的模式也适用于运营、财务、市场营销和人力资源。为什么不直接使用“ChatGPT + API”？因为那样容易出问题。 LLM不是可靠的执行引擎。它们会虚构字段名称、ID、端点和逻辑。企业系统需要安全、可审计的操作。 Worqlo将LLM视为解析器，而不是执行者。执行在受控环境中进行，具有：工作流模板模式合约基于角色的访问控制（RBAC）日志可重复的结果这既保持了自然语言的便利性，又保持了经典自动化引擎的可靠性。我们正在测试什么？我们想看看：对话是否可以取代UI，用于狭窄、结构化的任务确定性执行是否可以与自然语言意图共存多轮工作流是否真的可以减少运营负荷连接器模型是否可以扩展，而不会造成另一个集成混乱工程师是否更喜欢通过工作流而不是UI层来暴露功能现在还处于早期阶段。但该模型对于高容量、低级别的运营工作似乎很有前景。

查看原文

Most enterprise work isn’t slow because of bad data. It’s slow because the interface to that data is scattered.A single question like “Which deals are stalled?” touches dashboards, spreadsheets, a CRM, BI tools, internal scripts, and a few Slack threads. Acting on the answer requires switching between systems again. The friction is in the middle.Worqlo is an experiment in removing that friction by using conversation as the interface layer and deterministic workflows as the execution layer.The idea is simple: natural language in → validated workflow out.The LLM handles intent. A structured workflow engine handles execution: CRM queries, field updates, notifications, permissions, and audit logging. The model never executes actions directly.Below is how it works.Why Conversation?People think in questions. Systems think in schemas. Dashboards sit between them.Interfaces multiply because every system exposes its own UI. Engineers end up building internal tools, filters, queries, analytics pages, and one-off automations. That’s the UI tax.Conversation removes the surface area. Workflows add safety and determinism.Architecture (simplified) User → LLM (intent) → Router → Workflow Engine → Connectors → SystemsLLMExtracts intent and parameters. No execution privileges.Intent RouterMaps intent to a known workflow template.Workflow EngineExecutes steps in order:schema validationpermission checksCRM queriesAPI updatesnotificationsaudit logsConnectorsStrict adapters for CRMs, ERPs, internal APIs, and messaging systems.The workflow engine will refuse to run if:fields don’t existdata types mismatchpermissions failworkflow template doesn’t match user intentThis prevents the usual LLM failure cases: hallucinated fields, incorrect API calls, unsafe actions, etc.Example QueryUser:"Show me this week's pipeline for DACH"Internal flow:intent = llm.parse("pipeline query") validate(intent) fetch(data) aggregate(stats) return(summary)Follow-up:"Reassign the Lufthansa deal to Julia and remind Alex to follow up"Workflow:find deal by name validate ownership change write CRM update send Slack notification write audit logEverything runs through deterministic steps.Why Start With SalesSales CRMs are structured and predictable. Workflows repeat (reassign, nudge, follow-up). Latency matters. Output is measurable. It makes the domain a good test environment for conversational workflows.The long-term idea is not sales-specific. The same pattern applies to operations, finance, marketing, and HR.Why Not Just Use “ChatGPT + API”?Because that breaks fast.LLMs are not reliable execution engines. They hallucinate field names, IDs, endpoints, and logic. Enterprise systems require safe, auditable actions.Worqlo treats the LLM as a parser, not a worker.Execution lives in a controlled environment with:workflow templatesschema contractsRBAClogsrepeatable resultsThis keeps the convenience of natural language and the reliability of a classic automation engine.What We’re TestingWe want to see whether:conversation can replace UI for narrow, structured tasksdeterministic execution can coexist with natural language intentmulti-turn workflows can actually reduce operational loada connector model can scale without creating another integration messengineers prefer exposing functionality through workflows instead of UI layersIt’s still early. But the model seems promising for high-volume, low-level operational work.

Show HN: Worqlo – 企业工作流的对话层