Productivity

These productivity tools were judged on whether they preserved real structure, not just whether they sounded smart. Across the tests, the winners stood out when they kept the underlying structure intact: Fellow was strong on speaker attribution and sharing, Unstract reconstructed bank-statement structure, and Zapier AI Agents followed a routing map and built multi-step workflows from plain English.

Structured extraction still needs aggregate checks, because schema-looking output can hide numeric mistakes. Unstract produced good bank-statement structure and correct transaction-type classification, but the summary count was wrong. That means practitioners should validate totals and counts, not just whether the JSON nests correctly.

For no-code agent builders, the real bar is end-to-end task execution against a fixed workflow, not conversational competence. Zapier AI Agents passed all five tasks cleanly, followed the routing map exactly, and auto-structured multi-step workflows from plain English. The test setup makes clear that the useful agent is the one that completes the business process reliably, not the one that merely chats well.

A strong meeting notetaker can still be limited by access and integrations. Fellow was the winner because it handled speaker attribution well and supported open sharing, but the verdict also said the free tier was too restrictive and API/MCP access sat behind paid plans. So the core output can be excellent while deployment features remain gated.