Sber has introduced Cowork Bench — an open Russian-language benchmark that checks how well AI agents cope with office work. It assesses not the model's answers, but whether the agent can complete a task without human intervention. For example, how AI processes an application, prepares a report, or performs a calculation.
The benchmark includes 475 tasks, 12 business functions, and 26 corporate systems. During testing, the AI agent has to work with spreadsheets, documents, email, calendar, data storage, and a knowledge base.
An example task is budget analysis. For instance, the neural network needs to compare plan and actual figures for four departments, calculate deviations, explain reasons, prepare a report and presentation, revise the forecast, and send out emails. Another task is to organize a business trip, including selecting employees, finding tickets, calculating the budget, and adding meetings to the calendar.
The developers of the GigaCowork platform compared nine software shells for agents — on one open model, with identical tasks and verification. Among them were Goose, Claude Code, Codex CLI, OpenHands, OpenCode, and others. The share of completed tasks ranged from 62% to 80% depending on the shell. GigaCowork coped with 75.2% of tasks.
Previously, GigaPlatform AI Studio introduced the ability to assemble AI agents from ready-made blocks without programming, connecting their own documents and knowledge base to them — these are the scenarios that the new benchmark now tests. In addition, GigaAgent was presented, which breaks down instructions into stages, works with calendars and documents, and prepares reports and presentations.
Similar automation is being developed by other companies: for example, Smart Engines introduced AI agents for creating document templates that eliminate the need for manual setup of non-standard forms.
Read more on the topic:
Never miss our newsAdd this site to your preferred sources to see us more oftenAdd on Google
