Buying groceries online spiked during the pandemic years with the prolonged lockdowns. Although the trend continued in the post-pandemic years, an IBM survey of 20,000 consumers found that only 14% were satisfied with their online shopping experience.
Clearly, today’s discerning consumers expect a tailor-made shopping experience that comes with the convenience of product choices, detailed information, diverse payment methods, and a seamless integration of in-store and online experiences that cater to their individual preferences.
Preparing for AI integration
The good news is that predictive and generative AI can help meet these expectations. However, questions abound: How can grocers get started with AI and adopt foundational models to support their business? Which is the right AI model? What is needed to tailor these models? And how can we ensure that the models can be trusted and data integrity is maintained?
The reality is that different outcomes require different approaches. The way data sets are prepared and the type of AI models used have a bearing on the outcomes. The risk of using incorrect models can muddy the data and bias the algorithms, which would harm an organisation’s business.
To minimise risks, grocers can start with an evaluation framework that considers the diverse needs and skill sets of multiple decision-makers within the organisation. Specificity is crucial in choosing the right model. An AI model selection framework is a good place to begin with a foundational model.
Here are five tips to keep in mind when choosing the right foundation models.
- Identify your use case clearly: The easiest way to identify the right model for your use case is to start by creating both the prompt and the ideal answer, and then work backwards from there to find the data needed to provide the desired answer.
Getting input from key stakeholders to create actual prompts that are relevant to your industry and business, is a must-do. The more specific the prompts, the higher the accuracy rate. After all, the end game is to have a model that already knows the difference without extra training or input. - Right-size your model: There is a misconception that large AI models are needed to capture complex and nuanced connections within data to generate quality output, especially when it comes to groceries, which can range from dry to wet goods, canned to packed goods, organic, and more. The list is endless.
The recommendation is to right-size the model as it caters to specifics. Once the prompts are yielding the right responses, you can scale down the model and tune it further. Prompt-tuning a smaller model is far more cost-efficient than fine-tuning a massive model that needs loads of data and computing resources. - Test the model: The criteria for testing model performance are accuracy, reliability, and speed, with the weight for each criterion depending on your needs. Accuracy not only provides an objective and consistent measurement, it would also help to create a benchmark for future reference.
Naturally, a model that meets needs is the best option and as such, it is important to assess model performance and the quality of the output using metrics that serve the purpose. Over time, adding datasets to the stack and prompting the models can improve accuracy. - Evaluating risks and governance: Data privacy and security, transparency and traceability of the training data, accuracy and reliability of the output that is free of bias, as well as toxicity and hallucination are crucial in selecting a model.
This means AI governance must be applied throughout the entire selection process—from performance evaluation and optimisation, through prompt engineering, to validation and cost control. The process does not stop after the model is built but continues to enhance the model’s reliability and trustworthiness quotient. - Costing ROI: Cost is the other factor to consider, and the question to ask is: “Would an expensive model provide the ROI to justify its use”? Ultimately, a trusted and cost-effective foundation model that is optimised for cost and performance is preferred. To get there, it is important to continually revisit each AI use case in terms of relevance, model size, and performance.
We look around and see how personalised grocery shopping is already here. What a well-tuned model offers is a highly personalised, almost instinctive experience that caters to consumer needs. When that is available, the number of satisfied and happy consumers will certainly go beyond the current 14%.
