Chuangli AI Workspace · Enterprise AI Employee Platform

The company had already accumulated role SOPs, compensation methods, recruiting playbooks, and operating knowledge, alongside a Q&A MVP. But finding an answer is not the same as completing work: employees still need to organize inputs, break work into steps, confirm standards, and turn responses into usable artifacts.

I reframed the product from more Q&A entry points to AI employees that can complete business tasks. Every capability must define its use case, required inputs, execution steps, approval points, and final deliverable instead of merely swapping prompts.

An employee can begin with a one-sentence goal or an Agent card. The Agent Router recommends a primary and alternative Agent from capabilities the user is authorized to access. Deep tasks first create a structured brief that confirms goals, materials, constraints, and expected outputs before execution.

Execution is not a black box. Steps, progress, sources, tools, and waiting-for-approval states are visible; failures recover at the step level, while critical actions retain human confirmation. Completed work becomes a plan, report, table, checklist, script, or response library with version history and task provenance.

Employees should not need to understand models, prompts, embedding dimensions, or workflow orchestration. The user experience exposes only Knowledge Q&A, Quick Generation, and Deep Execution, expressed through business tasks such as recruiting plans, competitor analysis, livestream reviews, and service responses.

The product uses two interface layers: employees see an accessible, page-based AI front end, while administrators see Agent versions, knowledge scope, failed tasks, index health, and knowledge gaps. One system must be both easy to use and governable.

Enterprise conclusions must retain citations. When authoritative material is missing, the system labels an AI inference or asks follow-up questions rather than presenting model advice as company policy. Personal, project, and enterprise memory have different approval and visibility rules, and only reviewed content can become enterprise knowledge.

Tasks, steps, tool calls, and deliverables preserve versions and audit relationships. Long tasks support approval waits, step-boundary pauses, retries, and idempotent deduplication. These may look like implementation details, but they determine whether a company can safely entrust real work to AI.

I led product planning and experience: starting from existing enterprise knowledge assets to define the multi-Agent proposition, first-release scenarios, information architecture, Agent Router, task briefs, state machine, deliverables, and the knowledge-learning loop.

I also structured permission, memory, file, recovery, human-approval, and safety boundaries, translating product decisions into a PRD, high-fidelity prototype, architecture decisions, and executable development tasks. AI supported research comparison, prototype engineering, documentation, and consistency checks; I owned business trade-offs, product definitions, and acceptance boundaries.