Engineering Portal

Human review: AI Agents / MCP

Source: deck slides 12–13 · WPP Pilot assessment | Risks #2, #6, #20

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

#6 Process & Approval Bypass#

AI bypasses established controls.

Wrike AI Agents

  • Pilot: WPP's Taskforce reviews the use case before AI Agent activation.
  • Gap: Wrike does not provide a native workflow that routes AI Agent activation to an approver.

Wrike MCP

  • Confirm: WPP LLM client offers enforceable approval capability when human approval is required and which write operations are available.

#20 Autonomous Actions#

High-impact actions lack oversight.

Wrike AI Agents

  • Pilot: Review the agent's configured actions and possible impact upfront.
  • Note: AI Agents do not offer a separate approval step before each action.

Wrike MCP

  • Confirm: High-impact write tools require the intended human approval.

#2 Data Exposure & Scale#

Bulk actions affect unintended records.

Wrike AI Agents

  • Pilot: Keep actions within the agreed scope and exclude operations needing an unavailable preview. Wrike offers agent sandbox for testing agents before running them.

Wrike MCP

  • Confirm: A reviewer can check affected records and the approved time, scope and volume.

Screenshot in the deck (slide 13)#

Callout: "Testing AI Agent in sandbox before activating". Screenshot of the agent builder with the Test tab selected (the Configure | Test switch, "Discard changes" and "Create" buttons are visible). A dialog titled Custom agent shows:

  • Select an action to test: "We'll safely simulate this action so you can check if you're happy with the agent's response. No real changes will be made during this test." A dropdown with "Action 1".
  • Select where to test it: "We'll simulate this action only in the work item you select, not in any of its subitems. For best results, select an item with a name, description, and field values similar to the item you expect the agent to act on." Options Project/Folder (selected) and Task, and a "+ Item" button.

Discussion log#

DateTopicDecision / outcomeOwner