An AI agent is not a chatbot—it is a digital coworker inside the workflow

Challenge: a chatbot cannot carry a workflow

The organization in this case study is an AI startup that began in Taiwan and is expanding quickly across multiple markets—fast-moving, geographically distributed, with all day-to-day collaboration running on Lark. As it scaled, three pain points became increasingly clear:

• Meeting notes depended on people. Decisions and action items were scattered across chat logs, and important items had no tracking mechanism.
• Repeated questions. The same questions—product specs, legal terms, and pricing—were answered again every week.
• Slow content production. A single external deck took half a day to produce, and the output was hard to keep consistent.

Off-the-shelf AI chatbots could not solve these problems. They can answer questions, but they cannot carry a workflow: they do not pull meeting transcripts on their own, they do not write action items into the task system, and they do not deliver finished work back into the conversation.

Solution: an AI agent that enters the workflow

This project is built on an agent architecture, integrated deeply with the Lark open platform (messaging, tasks, documents, multi-dimensional tables, and interactive cards), with every model call routed and governed through ATP Petrichor.

1. Meeting automation. The agent polls for meeting transcripts on its own; when it detects a new meeting, it produces a structured summary document (summary, positions, decisions, and risks). Action items are written directly into Lark's native task system, with owner, due date, and reminder set in one step, and decisions are stored for traceability. Distribution follows privacy routing: meeting information reaches only the relevant members, and each member's space stays isolated.

2. Task governance. A single command in the conversation creates an action item, which lands directly as a native task and reminder and is tracked automatically before it comes due.

3. Enterprise knowledge base. Product Q&A, legal templates, and sales tools are stored centrally and synced automatically; the agent answers questions in the group chat, and every answer cites its source.

4. Content production line. Describe a need in the conversation, and the agent produces decks, product one-pagers, and social assets that follow the company's brand identity—vector-text PDFs delivered straight into the conversation.

5. Confirmation before execution. Before running a high-cost task, the agent reports the intended direction and estimated time on an interactive card, and executes only after confirmation; if a task fails, it reports back on its own, keeping progress visible throughout.

ATP's role in this case

The team also uses ATP's end-to-end Token API service—a single point of integration that switches between models as the task requires, with every agent call executing on the platform:

Governance area	How it works
Multi-model routing	Routine tasks use cost-effective models and deep tasks upgrade automatically; if a primary model fails, traffic fails over to a backup
Cost control	Scheduled and conversational tasks are billed by tier, with quotas set in advance—rather than discovering an overrun at month-end reconciliation
End-to-end audit trail	Every call records its model, usage, and source, so a usage anomaly can be located and fixed the same day

Results

• After a meeting ends, the summary document and action items are generated automatically, with no manual write-up.
• Repeated questions are handled by the knowledge base, with consistent, sourced answers.
• Decks and marketing assets went from "half a day" to "a single command in the conversation," with fully consistent brand styling.
• Model costs are visible and controllable: an audit trail once helped locate the same day an abnormal consumption on a scheduled task, which was fixed immediately after the model tier was adjusted.