The enterprise AI harness
that lives inside WeCom
AIPaddle doesn't ask anyone to open yet another system. WeCom is the entry point, WeDrive holds documents and trackers, the ERP keeps the books; AIPaddle sits on top to create, retain, coordinate and govern AI capabilities.
From one message in a group to one row in a tracker
- WeCom project groups · @PMO to assign, report, approve
- WeCom workbench (mobile) · chat, view, start/stop
- Personal assistant (web) · build AI in a sentence, voice, drag files
- Project pages / admin · teams, templates, permissions, approvals
- PMO Agent · checks, reminders, escalation, acceptance, reports, KPIs
- Personal assistant · understand → build → schedule → report
- AI workers · take tasks in a role, accepted by a human owner
- Nine AI object types: agent, workflow, digital employee, digital team, knowledge base, skill, MCP, API, extension
- Scheduled jobs · WeCom push node
- WeDrive project space: 11 standard folders for originals and deliverables
- Online sheets are the source of truth: plan, risk and issue logs, two-way sync every 5 minutes
- Knowledge base: records, team and policies indexed automatically, answers cite sources
- Roles drive assignment · permissions by role · department data scopes
- Approval center · tiered autonomy (reversible acts go ahead, irreversible ones need approval)
- Multi-tenant isolation · full audit trail
Five layers, each talking only to its neighbours. Select a layer to see what it holds.
Eight criteria a real enterprise AI harness must meet
Use them as a selection checklist; they are also how AIPaddle's modules are organized.
AI shows up where people already work (WeCom groups, the workbench), with no new system to open.
C2Capabilities as assetsEvery AI capability is stored as a standard object that can be edited, versioned and reused, and stays when people leave.
C3Enterprise contextKnowledge, files, org structure and roles become AI context, filtered by the caller's permissions.
C4Tools and systemsMCP, business system APIs and WeCom interfaces are connected and governed in one place, with credentials isolated per company.
C5Orchestration and teamworkWorkflows, digital employees and teams; AI workers share roles with people, and a person accepts every deliverable.
C6Built into processesPolicies become executable rules inside projects, trackers and reports, not just a chat window.
C7Governance and securityRole, permission and department data scopes; publishing approval, security scans on save, tiered autonomy and full audit.
C8Measurable and under your controlUsage, cost and KPIs are measurable; models are replaceable, SaaS or on-premises is your choice, and the data stays yours.
Four application objects do the work, five foundation objects power them
Every AI capability on the platform is one of nine object types. References between them are enforced in code, so whatever you compose stays traceable and governable. Select an object to learn more.
Digital team
A group of digital employees or agents working toward one business goal, orchestrated by workflows.
- References
- Members are digital employees or agents; collaboration is drawn in a team workflow
- Assignment, parallel work, hand-offs, review and summaries are all orchestrated
- Pre-publish check flags members missing from the flow and blocks non-members
- Call it in WeCom
数字团队#名称 内容- Where to edit
- Digital employees › Teams
Teams cannot contain teams; collaboration across employees must go through a team workflow.
Digital employee
An AI role with defined duties, made of one or more agents and their workflows; it can be assigned to a project role as an AI worker.
- References
- An AI role with a name, position, department, duties and KPIs
- Holds 1 to N agents, routed by keyword or bound to a workflow
- Can be @-mentioned in groups and speaks up within its duties, rate-limited
- Call it in WeCom
数字员工#名称 内容- Where to edit
- Digital employee workbench
A digital employee cannot contain another digital employee.
Workflow
A traceable flow that orchestrates agents, tools, skills and knowledge bases, and can push results to WeCom.
- References
- Can orchestrate agents, digital employees, retrieval, tools, HTTP requests and WeCom pushes
- Generate in one sentence or build on the canvas, with version history
- Can be the target of a scheduled job
- Call it in WeCom
工作流#名称 内容- Where to edit
- Workflow canvas
Multiple agents can only collaborate through a workflow; a workflow cannot be registered as a tool.
Agent
The smallest unit of execution: a model plus instructions plus bound resources.
- References
- Can use: models, knowledge bases, skills, MCP / API tools
- Can use a workflow as its "brain", running it each turn
- Cannot reference other agents (multi-agent work goes through workflows)
- Call it in WeCom
agent#名称 内容- Where to edit
- Agent orchestration page with a live test chat
An agent cannot contain another agent.
Knowledge base
Takes files or data sources, chunks, indexes and retrieves them; answers cite sources, and when nothing is found the code forbids making things up.
- References
- PDF, Word, Excel, PPT, TXT and Markdown, up to 50 MB per file
- Retrieval respects the caller's own permissions; inaccessible libraries are skipped
- Project and WeDrive folders can be indexed directly
- Call it in WeCom
Used by agents; ask it directly through the assistant- Where to edit
- Knowledge base detail: documents, chunking, testing, settings
Chunks stay under 512 tokens to keep retrieval precise.
Skill
A reusable method that guides an agent through a type of work: SOP, prompt and dependencies.
- References
- SOP, prompts, rules, examples and dependencies, shared by many agents
- Can depend on MCP / API tools and knowledge bases
- Generated through a guided chat and testable in the marketplace
- Call it in WeCom
skill#名称 内容- Where to edit
- My Skills editor
A skill cannot depend on a workflow.
MCP
Standard tool services exposed through the Model Context Protocol, with tools discovered dynamically.
- References
- Auth: Bearer, Basic, custom headers, OAuth (authorization code + PKCE)
- Platform-level servers are registered once; each company keeps its own credentials
- Tool lists are discovered at runtime
- Call it in WeCom
Called by higher-level objects- Where to edit
- MCP page
Business objects reference governed tools, never a whole server.
API
Interfaces to external business systems, such as ERP master data and document queries.
- References
- Import OpenAPI and expose only the operations you pick
- Parameter constraints, field masking, domain allow-lists
- External databases are connected through a read-only HTTPS layer
- Call it in WeCom
Called by higher-level objects- Where to edit
- Plugins › API
Credentials are encrypted per company, and permissions can only narrow, never widen.
Extension
The only way in from outside: WeCom bots, external system APIs and webhooks.
- References
- Targets: published agents, digital employees and teams
- Forms: WeCom bot (long connection), external API, webhook
- API keys stored as hashes, shown once, revocable; rate-limited by key and IP
- Call it in WeCom
@-mention the bot in a group- Where to edit
- Extension settings
Once the target object is retired, calls are refused immediately; each extension runs as a machine user under the same data isolation.
Built by the assistant or by hand, every object follows one lifecycle
One sentence produces a build plan; once confirmed, objects are created in dependency order.
Every save is scanned: hard-coded secrets are blocked and injection risks are fixed automatically.
Run it once with pushes previewed, not sent; schedules start only after you confirm it works.
The creator can use it right away; nobody else can see it yet.
Change prompts, knowledge bases or flows in the editor, or just say "that's not right" in chat.
Choose the company space, specific people or departments, and submit for approval.
The org admin reviews first and can pass it to the department AIBP for final approval; it takes effect for the grantees immediately.
Published objects are retired before deletion, and anything that depends on them is flagged.
Authorized, audited, measured
About 80 permission actions granted by role and department data scope; hidden menus are not security, every entry point checks.
Reversible actions run directly and can be undone; irreversible ones go through the approval matrix.
Every row carries a company ID and database row-level security backs it up, so even an application bug cannot leak another company's data.
Every AI call records tokens, model and source; people and AI workers share one KPI set.
Four steps to keep individual AI capability in the organization
- 01
Connect
Open SaaS or deploy on-premises, sync your WeCom org structure and set up WeCom management.
- 02
Retain
Each department turns frequent work into agents, workflows and knowledge bases with the assistant, then shares them after approval.
- 03
Collaborate
Put the PMO Agent on one key project, turn policies into rules and assign AI workers to roles.
- 04
Measure
Review usage, acceptance rates and KPIs to choose the next capabilities to build.
Want to see it running in your business?
A 30-minute online demo built around your industry, including how we would connect to your WeCom and ERP.