Unclear data science tasks delay research projects
A central data lab receives 200 data science tasks a week. Unclear descriptions cost more than 160 hours and delay analyses.
- Science
- Data analytics
Key metric
Reduced clarification effort per week
16070h
Weekly clarification effort in the data lab
At a glance
Starting point
A non-university research institution with 3,400 employees runs a central data lab that supports researchers from labs and projects. Around 200 data science tasks come in every week, from simple evaluations to complex modeling. Requests arrive by email, ticket system and chat, often without a clear question, data basis or priority. Three data science teams with twelve full-time staff in total spend about 160 hours a week on follow-up questions, sorting and tracking tasks before they can even start the actual analysis.
- 200
- Tasks per week
- 160 h
- Clarification effort per week
The challenge
The data science teams are expected to deliver short-term analyses for ongoing experiments while also developing long-term models for strategic projects. In practice, however, there is no structured intake channel for tasks. Every week, the 200 incoming requests generate dozens of follow-up questions about data sources, target variables and deadlines. Many tasks are logged twice, for example as an email and as a ticket, or lose context when staff change. Research leads have no reliable overview of which tasks are already running, which are blocked and which have been deferred. As a result, evaluations slip, experiments have to be repeated and funding deadlines come under pressure.
Typical symptoms in the company
Why this remained unsolved
The problem had been known for a long time but was considered an unavoidable side effect of dynamic research operations. Previous improvement attempts relied on new ticket templates and training for requesters. These measures helped only to a limited extent, because many researchers continued to prefer email or chat messages and ask spontaneous questions. A classic workflow system seemed too rigid to keep up with the variety of tasks, data sources and project setups. At the same time, the effort of a large IT project was hard to justify as long as the teams could somehow absorb the extra work.
Our solution
The institution introduced an agentic workflow that bundles and structures all data science tasks through a shared intake channel. Incoming emails, tickets and chat messages are automatically assigned to a task, turned into clear short descriptions and tagged with metadata such as project, data source, urgency and expected deliverable. A virtual coordination assistant guides requesters through targeted follow-up questions, suggests suitable categories and documents decisions directly in the project tool. Data science leads get a central board that sorts tasks by priority, effort and dependencies and supports handovers between teams. The result is a continuous, traceable process from request to completed analysis.
Where generative AI comes in
Generative AI analyzes free text from emails, tickets and chats, condenses it into clear task descriptions, suggests categories and priorities and phrases follow-up questions in plain language. It can handle incomplete, multilingual and vague descriptions and adapts its suggestions to the data lab’s style.
- Central intake for emails, tickets and chat messages that automatically bundles, deduplicates and consistently assigns requests.
- Virtual coordination assistant that guides requesters with follow-up questions, fills in missing details and documents them in the project tool.
- Prioritization board for data science leads that ranks tasks by project, urgency, effort and dependencies.
- Automated logs that document the history of every task from request to analysis in a traceable way and are ready for audits.
Introducing the agentic workflow
Current-state analysis
Survey of existing intake channels, types of data science tasks and tools in use, including time spent on clarification and prioritization.
Workflow design
Defining a target process with a uniform task format, roles for coordination and prioritization, and rules for handovers between teams.
Integration
Connecting email, ticket system and chat to the new intake channel, including write access to the data lab’s project tool.
Assistant logic
Configuring the virtual coordination assistant with typical questions, categories and examples from the institution’s everyday research.
Pilot and rollout
Testing with selected projects, followed by adjustments and a gradual rollout to all data science teams and research areas.
Measured: after six months in production
Period: after six months in production.
| Metric | Value | Basis |
|---|---|---|
| Reduced clarification effort per weekKey metric | Clarification effort down from 160 to 70 hours per week | Time tracking before/aftermeasured |
| Lead time until analysis starts | Median reduced from 3 days to 1 day | System measurementmeasured |
| Assistant suggestions with manual correctionQuality measure | Adjustments needed in 15 percent of tasks | Sampleestimated |
Cost-effectiveness
The benefit depends heavily on researchers accepting the shared intake channel and providing basic details. With many new project types, the need for manual control increases, because the assistant initially makes less fitting suggestions.
Integration and operations
Transferability
The agentic workflow for data science tasks transfers easily to other knowledge-intensive areas that receive many unstructured requests and where specialized teams are scarce. What matters is that tasks arrive through several channels, clarification takes a lot of time and priorities change often. Wherever free text has to be translated into structured work packages, a similar approach can help relieve capacity and create transparency.
A similar case in your organization? Let’s talk.
Briefly describe your process. We’ll get back to you with an initial assessment of whether the pattern can be transferred, and what it realistically takes.