Generative AI is here to stay, and is set to revolutionise digital products in a fundamental way. One key principle remains standing: you need to think from the end-user experience backwards to the technology. That's exactly how we approached a recent brainstorming session for Valcori, a procurement platform for medium-sized businesses to create tenders and streamline the supplier selection process.
On average, 80% of the tendering process is repetitive manual work that is prone to impactful errors. We asked ourselves how Large Language Models (i.e. ChatGPT) could assist our users to save time and improve overall tender quality.
The plan: replacing a sizeable part of the heavy lifting for our users when creating a new tender through the power of AI, with minimal manual data input
The result: an intuitive user flow to instruct the LLM to return an elaborated set of questions and parameters that fit neatly in Valcori's tender structure
The timeline: just 3 weeks to go from plan to implementation in production of the MVP, saving users 55% on the usual time spent on tender creations
The proof of the pudding is in the eating, let's demonstrate a concrete example:
Need a maintenance service provider for your elevators over the next three years? Valcori will generate all the crucial questions to ask your potential suppliers.
Link to demo video: https://youtu.be/UhpO3-XNY1w
Overcoming ChatGPT's creative tendencies
During the implementation, we discovered that ChatGPT occasionally gets a bit too creative and returns invalid responses, even when the prompt is consistent. To address this, we implemented Triple Modular Redundancy for our production environments. We send three parallel requests to ChatGPT, and return the first valid response to the client (product backend). While this increases costs, the current pricing structure of the OpenAI API makes it feasible. As AI models improve and with chatGPT-4 being rolled out, we expect such redundancy measures will become obsolete.
A dedicated microservice
The process described above is handled by our serverless microservice that converts JSON input from the backend into a ChatGPT prompt using product-specific templates. The ChatGPT response is then converted into JSON and sent back to the backend. This technical setup was an easy fit into our existing infrastructure with Google Cloud Platform.
Broadening use cases
We see the use case described here above only as the starting point. The possible use cases in procurement workflows are infinite, both for buyers and suppliers. In the end it's in everyone's interest to find the most optimal path towards beneficial procurement outcomes for all parties involved!
Power up your digital product with generative AI
Are you interested to discover how genAI can unlock new opportunities in your digital product? Reach out to our team (email@example.com)