What happened
MWS AI, part of MTS Web Services, has launched MWS AI Cowork, which it describes as a universal AI agent for office tasks. According to Vedomosti’s report, the service provides a single interface for finding and analysing information, interacting with files and corporate applications, and creating spreadsheets, presentations, and reports.
The product is built around a chat interface, but its proposed work surface extends into existing company systems. Cowork can connect to email, calendars, messengers, storage systems, CRM, and other corporate applications. It acts on behalf of a user and within that user’s existing access rights. That is a consequential implementation detail: the launch is not presented as a general-purpose chatbot offered alongside business software, but as an agent intended to operate through a company’s existing identity and permission structure.
MWS AI says the service can be deployed in an organization’s closed environment in Russia, framing this as an answer to local requirements around corporate-data security and flexible AI infrastructure. It also says Cowork is not tied to one model. The company lists its own Cotype model, models from Sber and Yandex, and open foreign models as possible options. The report does not specify the routing method, supported model versions, or the operational conditions under which each option is available.
For recurring work, employees can create “skills”: instructions that can be shared with colleagues. MWS AI’s example is a monthly report workflow in which a worker defines data sources, verification rules, and the target document format in advance. Cowork also retains user preferences and task context. The company says retained information can be viewed, corrected, and deleted. MWS AI argues that a company can begin with discrete tasks, such as preparing reports or comparing supplier offers, then add departments and eventually support processes involving several teams.
The company reports pilot results from MTS Web Services units. It says that an employee using the agent for 90% of tasks saved at least one hour a day, which it values at about RUB 263,000 annually per employee. In tests of individual scenarios at a company with 1,000 employees, preparation of a ten-slide presentation took 40–50% less time; a quarterly financial report, 30–50% less time; and meeting minutes, 50–80% less time. MWS AI says those estimates include employee time spent checking and revising the output. An annual 2026 licence costs from RUB 6 million to RUB 20 million, depending on configuration and number of users.
What the archive adds
This is a new event in a related office-agent thread, rather than a new fact about an already covered MWS AI product. Univer Workspace described an open-source office workspace as a shared change surface for people and agents. Cowork brings a comparable direction into a commercial enterprise offer, with closed-environment deployment, application connections, and a stated licence range.
Trend signal
The launch supports Harness architecture outlives model choice. MWS AI’s stated product boundary is the surrounding harness—connectors, identity-bound access, skills, retained context, deployment controls, and reviewable memory—rather than a single named model. The evidence is vendor-reported, but the multiple-model claim makes that architectural framing explicit.
Why now
MWS AI frames Cowork against Russian companies’ requirements for control of corporate data and flexibility in AI infrastructure. Its closed-environment deployment option and model choice are therefore central to the timing and positioning of the launch, not incidental product features.
Connections to other events
The prior Univer Workspace note focused on people and agents collaborating in the office workspace. Cowork addresses a similar operational surface from another direction: it connects an agent to existing enterprise applications and asks it to work under the employee’s permissions. The common question is whether office AI becomes a feature inside individual editors or a governed layer spanning the systems where work already occurs.
Why it matters
For teams building with AI, Cowork illustrates the requirements that turn a document assistant into an operational agent. Useful automation needs access to the systems that hold work context; that access needs to inherit permissions; repeated procedures need reusable instructions; and stored context needs inspection and correction mechanisms. MWS AI also places model choice beneath this layer, which may matter to organizations balancing deployment constraints, internal model preferences, and changing model availability.
The reported licence range also changes how the product should be evaluated. At this price level, a buyer is not primarily buying isolated drafting help. It is buying a governed integration and automation program, and should measure the cost against verified time savings, review workload, integration effort, and the operational risk of granting an agent access to corporate systems.
What happens next
- MWS AI will broaden Cowork from discrete document tasks to multi-team workflows. (6 months, 60%). Watch for: named integrations, workflow templates, or customer deployments beyond the reported pilots.
- Enterprise evaluation will focus on permissioning, memory controls, and human review rather than model choice alone. (12 months, 65%). Watch for: published administrative controls, audit features, or scenario-level outcome metrics.
Limits and uncertainty
The available report describes company claims and pilot results; it does not provide an independent evaluation, a detailed methodology, or sample sizes for the timing figures. It does not establish how reliably Cowork completes tasks, how often human revision is required beyond the assertion that it was included in the measurements, or how permissions are technically enforced across each connector. It also does not specify which open foreign models are supported, the product’s availability date, licence terms, security certifications, or whether the listed price includes implementation and integration work. The claim that employees can delegate up to 90% of repetitive office tasks is a forecast by MWS AI, not a demonstrated general result.
Skill to build now
Choose one recurring office workflow and define it as a testable skill before automating it: list the allowed data sources, the operator permissions required, checks the output must pass, and the expected document format. Measure total completion time including review and rework. This mirrors the product’s stated skill model and prevents a team from mistaking a fast first draft for a completed business task.
Reader takeaway
Cowork is a signal that enterprise office agents are being packaged as controlled work layers, not merely chat interfaces. Its most important claims concern closed deployment, inherited access rights, reusable skills, inspectable memory, and model flexibility. Builders should evaluate those controls alongside any claimed productivity gains.