Governance: review and action oversight

Source: deck slides 5–7 · WPP Pilot assessment | Risks #6, #20, #21

Legend: Today = Wrike capability | Confirm = setup/evidence | Pilot = proposed control. See the overview for all labels.

#6 Process & Approval Bypass#

Established controls are bypassed.

Wrike AI Agents

• Today: Agents cannot change access permissions or approval rules.
• Pilot: WPP's Taskforce reviews the proposed use case before the agent is activated.

Wrike MCP

• Confirm: Check which write actions the assistant asks a person to approve, and which actions WPP's middleware blocks or allows.

#20 Autonomous Actions#

High-impact actions lack oversight.

Wrike AI Agents

• Today: AI Agents work similar to automations within their assigned location, with no delete action.
• Note: AI Agents do not ask for approval before each action.

Wrike MCP

• Confirm: Test a high-impact write in the chosen assistant to confirm LLM client controls

#21 External AI Platform#

Data leaves approved boundaries.

Wrike AI Agents

This risk concerns an external AI platform connecting to Wrike through MCP. Native-agent ownership and creation rights are covered on the earlier slide.

Wrike MCP

• Pilot: WPP approves the external platform, connection design and data scope before use, then records each approved connection and its owner.

Screenshots in the deck (slides 6–7)#

AI Agent configuration interface and available actions (slide 6)#

Callouts: "AI Agent configuration interface" and "AI Agent available actions".

Configuration interface. A "Back to AI agents" link, a Configure | Test switch and a "Discard changes" button. The left panel is a conversational builder:

• Greeting: "Hi, Client Service! I can help you get started with AI agents. What would you like to do today?"
• Suggested starts: "Describe a problem you want to solve", "Share an idea for an agent", "Learn how agents work in Wrike (like setting one up or testing it)".
• Notes under the chat: "As we chat, I'll keep the settings in sync.", a "Copy chat" button, a feedback prompt ("How do you like this conversational agent builder?") and "AI-generated content may be incorrect".

The right panel has these sections:

• Agent scope: "Select which tasks, projects, or folders will define this agent's scope. The agent can watch for triggers and act only in these items and their subitems." A "+ Scope" button and an info box: "Scope can be defined now or anytime later from Space settings > AI agents."
• General instructions: "Define the agent's main purpose and how it should behave in any situation… perform the specific actions below." An example prompt starts "You should act as a project assistant and monitor assigned wo…" (cut off).
• Trigger: "Select what triggers the agent and where to watch for it." A "Select trigger" dropdown.

Available actions (the "Select action" dropdown, outlined in red):

• Change field values (submenu)
• Change location (submenu)
• Create task, project, or folder
• Create item from blueprint
• Manage approval
• Post a comment
• Send email (submenu)
• Send Microsoft Teams message…
• Send Slack message…

Nearby text, partly covered: "…above trigger happens. Drag actions to change their … action to complete its next actions." Each configured action has a delete (bin) icon.

AI Agent reasoning logs (slide 7)#

Callout: "AI Agent reasoning logs". Screenshot of Agent action details for one run:

Summary shown: "The project is significantly overdue by approximately 161 days with all 6 active tasks past due, indicating critical schedule risk. No progress has been logged against planned effort, raising concerns about resource allocation and execution. Immediate attention is required to address delays and realign resources to mitigate further impact."

Highlights shown: the work item above, "Overdue by 161 days with 6 active overdue tasks; no logged hours or effort used; assigned to Resource Manager" (name omitted), then a link to the item on wrike.com. A "Recommendations" heading follows, but the content is cut off.

Discussion log#