Applying ChatGPT on corporate data to improve operational efficiency
In a previous article, we outlined a few ways through which business operations teams can use ChatGPT to enhance their work. In this article, we’ll deep dive in one of those segments (discover hidden insights) and explain how to leverage ChatGPT to identify operational efficiency improvement ideas in your business and how to use it with your data to find answers to your challenges.
The premise of using ChatGPT to improve operational efficiency
ChatGPT (not that it needs any introduction) is an artificial intelligence language model that uses machine learning to generate human-like responses to user inputs. So, if one could give it the right business context (more on that below), it would be able to provide fairly relevant ideas for business operations teams - think of it as a personalized digital consultant.
While one can read about dozens of ChatGPT use cases on the internet, they are mostly focusing on marketing and sales examples. However, we believe there is a large opportunity for business ops teams, too, especially in scanning their data and finding improvement levers with which they could improve their day to day.
We believe that applying ChatGPT on business ops is particularly exciting currently for 3 main reasons:
- Profitability is more important than ever: in a cash rich environment, one can throw money at problems. However, given the current macro context, everybody is looking for ways to increase profitability.
- Some levers are not obvious: Past a certain point, it's difficult to find new levers in business operations. This is primarily due to lack of time (business ops teams are always firefighting) to investigate new avenues.
- Best practices rarely apply directly: Best practices are generally true, but cannot be applied directly without tailoring to one’s specific circumstance. For example, monitoring content quality in classifieds is an evergreen best practice, however, the specifics matter depending on what you are trying to achieve (e.g. improve customer engagement, then aim for complete descriptions and engaging pictures; reduce fraud, then try to detect spammers and fake ads). So, something tailored to one’s specific context and data is much more powerful.
Key requirement for AI: The right business context
The big question is: what data should one use to achieve the best results with ChatGPT in identifying operational efficiency improvement levers? There are limitations to what can be sent through to OpenAIs APIs (due to prompt and response size token limitations) so quality over quantity is key.
In short, based on our experiments, getting relevant recommendations for business ops requires 3 inputs:
- Company sub-vertical: ie. last-mile delivery, vertical real-estate marketplace etc.,
- Business challenge: what is the problem that you need to solve,
- Metadata: namely, what type of data you have available based on which you need to get tailored recommendations.
While the first two are easier to get right, the third can be hard to deploy at scale by the business ops teams. This is because the metadata should be a) available, b) (relatively) self-explanatory and c) relevant to the business challenge
- Available: simply put, business ops teams should have access to their needed metadata without asking a tech / data lead and without typing it in manually into ChatGPT.
- (Relative) Self-explanatory: namely, an operator could understand what the table contains by reading the the metadata (ie. table columns / data types). For example, if you are a ecommerce company and have an “Orders” table, you would expect the columns to be something along the lines of order value, item ordered, status, etc.
- Relevant: if you are trying to reduce order delays or understand potential outliers in your order data, you should have an “Orders” table.
Prioritize and act on operational efficiency improvement levers
We’ve spent some time in the last few weeks thinking about how we can help our clients get relevant operational efficiency levers in a simple and effective way. On top, we wanted to ensure there is a way to act upon the recommendations.
Our (beta) solution to this is AI Scanner (we’re still working on the name 🙂). AI Scanner does 3 things:
- Collects the relevant inputs,
- Scans the data and recommends monitor ideas based on the inputs / metadata,
- Generates the SQL code (& the datasets) for the monitors that best fit the client needs.
The above enables the clients to set up monitoring for the best recommendations and, ultimately, act upon them.
Below is a video showing how this all works in practice:
While AI Scanner is still in beta, it already provided plenty of unique ideas to our clients. Furthermore, and this is probably the most exciting, we made these suggested levers actionable by integrating ChatGPT into our monitor creation flow.
Lastly, a word on security - we do not share any data outside our system, apart from the metadata (which is, as mentioned, column names and data types). In this way, no actual client data leaves our premises.
We’re curious to hear your thoughts on this exciting new avenue. And if you want to try this out, reach out for your free trial here.