AI AGENT PROJECT CASE

BMS unified intelligent office site reconstruction.

An AI Agent and cross-system orchestration concept for closing the loop across meetings, doctors, content operations and IT governance.

01Unified workbench02AgentRun engineering03Controlled execution

✦AGENT
RUN

WECHATONE

WECOM

VMEETING

BI / CRM

This page summarises a project proposal and solution design supplied by the user. It describes the planned approach and target architecture, not a public claim about completed delivery or production performance.

SOURCE MATERIAL

Three visuals.
Three exact claims.

The screenshots below are paired with the architecture, registered tool actions and risk levels visible in the supplied BMS proposal.

BMS five-layer target architecture with AgentRun, MCP and enterprise APIs — **Five-layer target architecture**The diagram separates the Agent experience, AgentRun orchestration, MCP, WeChatOne / WeCom / VMeeting APIs and connected enterprise systems.

BMS registered Agent tool portfolio — **Registered tool portfolio**The table lists concrete actions for meeting status, configuration sync, doctor QR codes, invitations, identity checks, RSVP, attendance, reporting, BI retry and article publishing.

BMS L0 to L3 controlled execution levels — **L0–L3 execution levels**The visual grades capabilities from L0 through L3 so query, write-back, external-impact and restricted actions can receive different controls.

Source note: these are screenshots of the supplied project proposal. They illustrate the proposed solution design and should not be read as a public production-performance claim.

THE PROJECT CHALLENGE

The system boundaries were clear. The employee experience was not.

Employees moved between WeChatOne, WeCom and VMeeting. People passed information by hand, checked status repeatedly and lost visibility when a long task failed. The project direction kept the existing business systems in place and added one governed entry point.

01Scattered entry pointsCross-system tasks depended on manual handoffs.

02Status breaksThe same business state was hard to reconcile.

03No long-task trackingFailures lacked retry, takeover and traceability.

THREE REAL WORKFLOWS

The Agent enters
where work gets stuck.

The project materials describe role-based workflows rather than a generic chatbot.

01 / SALES

From recruitment to RSVP

A sales representative can generate a doctor recruitment QR code, check unregistered doctors, trigger one-time identity verification and follow RSVP and attendance from one journey.

Invite→Verify→RSVP→Attend

02 / OPERATIONS

One request, many systems

Meeting operations can query unsynchronised meetings, update configuration, check valid attendance, generate a report, send data to BI and retry failed records through one orchestrated task.

OTIS→VMeeting→BI

03 / CONTENT

Content, consent and feedback

Content operations can create reminder posts for doctors who have not RSVP’d, manage consent follow-up and send post-attendance surveys across WeChatOne and Tencent Questionnaire.

Draft→Approve→Publish→Learn

TARGET ARCHITECTURE

One experience layer.
Clear system boundaries.

The Agent does not connect directly to databases. It works through registered tools, permissions and auditable service APIs.

01Experience layerUnified workbench · Agent assistant · Task centre · Audit and management

02Intelligent orchestrationAgentRun · Intent recognition · Plan generation · Async tasks

03MCP gatewayTool registry · Permission binding · Risk grading · Rate limiting · Audit

04Business servicesWeChatOne API · WeCom API · VMeeting API

CONTROLLED EXECUTION

More power requires
more control.

The proposed model grades capabilities by risk and adds preview, confirmation, retry, compensation and audit before a tool can affect an external system.

L0Query & summariseNo confirmation

L1Low-risk writeConfirm before execution

L2External impact / batchStrong confirmation and impact scope

L3Restricted operationReject and route to a compliant path

PROPOSED DELIVERY ROADMAP

Read first.
Write with confidence.

The source deck proposes three independently demonstrable phases with security review throughout.

012026 / 09

Read-only foundation

AgentRun environment, tool registry, L0 query tools, unified site, task centre and read-only audit.

022026 / 10 → 11

Partial write-back

VMeeting and WeCom write tools, orchestration, retries, task tracking and audit.

032026 / 11 → 12

Controlled go-live

WeChatOne write tools, higher-risk actions, security review, UAT and phased rollout.

A CASE STUDY IN PRACTICAL AI

Start with the
workflow that matters.

Discuss your Agent case↗