Ad hoc analysis designed to answer a one-off question. You have a straightforward question about your data and “ask a data scientist” or “go to the data science team” to answer it.
How many customers are buying product X?
Which customers are spending the most on product Y?
Did last month’s sales go up or down?
Someone on the data science team puts your question on their to-do list, ideally gets an answer to you within a couple of days, and is available to answer any questions you have.
Routine reporting. Dashboards, charts, and queries where the questions are known and settled. You used to have to painstakingly write matplotlib or seaborn code that gave you far more control over the plotting design process than you wanted, make your plots look attractive, and automate the process with a workflow.
Data wrangling. Before about 2020 data scientists regularly complained that 80% of the job was cleaning columns, fixing dates, and converting units, and doing what felt not just unglamorous but uncomfortable in a “this shouldn’t be my job because my job is legitimately something else” way. It took the development of dedicated data engineering teams before the average data scientist could start to work ergonomically.
What remained was that the data scientist could still physically do coding-required tasks that other people couldn’t. If you didn’t know Python or pandas, you had to find someone who did. This both overloaded the data science team and forced others with closer access to the data to rely on them as a bottleneck.
Even with out-of-the-box AI most of these issues are in a position to evaporate. You can ask AI for a snippet of pandas code that will get you the one-off answer you had to wait days for before. You can build automated reporting workflows without writing any code. You can generate reusable scripts and put them into a permanent codebase. If you already know what the results need to be and can check them against what the code produces, you’re someone AI can genuinely remove the bottleneck for.