3 Unrealistic Expectations for AI and ML Adoption in Food & Beverage – And A Framework for Managing Them

-->

As a co-founder at Ai Palette, I spend a lot of my time with brand managers and consumer insights teams discussing the many applications of Artificial Intelligence (AI) and Natural Language Processing (NLP) in the Food and Beverage (F&B) Industry.

I always look forward to these engaging conversations that get into the details and discuss the organization-specific use-cases and pain-points.

Truth be told, the outcome of such conversations depends on the willingness to manage the perceived expectations from AI technology.

It also requires a preparedness to embrace the change in processes and outlook that AI implementation will demand.

Through this blog, I am attempting to reach out to CPG product development and consumer insights stakeholders.

And I am proposing Expectation Management as a prerequisite for Change Management.

I have developed this belief that the onus of Expectation Management is with both the AI Technology Partner (like AI Palette) and the product and consumer insights teams.

In this blog, I will throw some light on three critical expectations and share a framework that may help you with expectation management during the initial days of your food-tech conversations.

An expectation management framework for AI and NLP applications in F&B

This framework is my three-step and honest approach to expectation management during AI in F&B conversations:

  • Step 1: Discuss the reason why AI is falling short of meeting the perceived expectation levels.
  • Step 2: Compare the extent to which AI can live up to the expectation (solve the pain-point), vis-à-vis traditional methods.
  • Step 3: Discuss if adoption of the AI solution, even below the perceived expectation levels, can put the organization in the driving seat.

Let’s put this framework to test as we evaluate three expectations from the AI technology that I have encountered the most during my conversations.

Expectation #1: Outright Accuracy

Artificial Intelligence, along with its Natural Language Processing capabilities, and Machine Learning, are cutting-edge data-driven technologies. There is a preconceived notion that the output of all AI-based applications ought to be outright accurate!

But this preconception requires a lot of reconsideration.

And this reconsideration is specifically needed in the case of AI and NLP applications, trying to solve consumer insights related pain-points for the F&B, CPG, or FMCG businesses.

Step 1 of our framework, let’s discuss the reason behind why AI-based F&B applications fall short of this perceived expectation:

  • Such AI applications process Big Data (being generated in volume, variety, and velocity) procured from new-age data-sources like social media platforms, e-commerce stores, other F&B-related online marketplaces, retail sales data, and more.
  • We are talking about highly unstructured and extremely valuable user-generated and system-generated data.
  • This data becomes available due to activities ranging from light-hearted banter between friends on social media to POS transactions at retail outlets.
  • This data was never generated and stored to decode the hidden consumer insights.
  • Just to put things in perspective. If we compare this with the use-case like predictive maintenance of equipment in factories and shop-floors – the available data is already structured and generated with the very intent of analysis and monitoring.
  • In technical terms, ML models for F&B use-cases need to work with a lot of unlabelled data. And labeling publicly available data is an uphill task.
  • For example, a comment on social media – ‘enjoying hot and spicy noodles for my lunch’ – can be easily decoded by the ML models as preferred physical attributes of that food item.
  • But due to lack of data labeling, another comment – ‘enjoying smoothie, a perfect retreat on a hot day’ – can misinterpret an attribute for the outside environment as an attribute for the food item.
  • All this impacts the accuracy of the consumer trend predicted by the AI application. However, this impact can be mitigated, as ML models continue to unlearn and learn based on the feedback.
  • In the future, we may have the tools and technology to automate the data-labeling process.

Step 2 of our framework nudges us to compare the capabilities of traditional methods to uncover consumer insights from similar or other sources of data:

  • Legacy tools, frameworks, processes available to the consumer insights teams lack the required processing capabilities to deal with Big Data.
  • And the insights derived by our legacy systems from the consumer data collected via surveys, interviews, group discussions, and more are equally prone to errors due to various kinds of human biases.
  • The not so encouraging – 90% failure rates of the new product ideas – can also be attributed to legacy tools and systems to a certain extent.

Step 3 compels us to evaluate whether the insights delivered by the AI solutions, at their current accuracy level, will enable the consumer insights and product development teams to take a step forward in the innovation and technology adoption curve?  

  • The answer is a resounding, Yes! AI and NLP applications enable F&B companies to derive insights from non-legacy sources. And these insights are deprived of inherent human-biases.
  • The accuracy of such insights will be higher than those derived by the legacy systems and sources.
  • And with the maturity of input data and ML models, it is expected that the accuracy of trend spotting will improve.

Expectation #2: Replacement for Human Intelligence

Human Intelligence is governed by Bounded Rationality. Herbert Alexander Simon (an early pioneer of Artificial Intelligence, co-creator of the Logic Theory Machine and the General Problem Solver, and a Nobel Prize-winning economist, introduced this theory)

As per this theory, individuals take rational decisions within the limits of information that can be processed by them, their personal experiences, knowledge, and specific brain anatomy.

The computational powers of the machines continue to increase exponentially. And they have also become cost-effective.

As a result, Artificial Intelligence (AI) has emerged as an ideal companion to assist humans with rational decisions based on Big Data, which is otherwise beyond the scope of our bounded rationality.

But many of us are unaware that even Artificial Intelligence and Machine Learning are governed by the same theory of Bounded Rationality (step#1 of our expectation management framework activated!)

  • Dr. Peter Angeline very aptly explained The Bounded Rationality in AI and ML during his keynote speech at the Transforming Food Insights Summit in November 2020 (you can watch the recording here – AI’s Inevitable Disruption of CPG).
  • According to Dr. Peter Angeline, a Strategic Futurist at The Hershey Digital Innovation Lab, the rationality of AI tools and applications is bounded by the Data Bubble around them.
  • This data bubble is formed based on the amount and kind of data to which these tools have access.

  • Hence, it is essential that we re-align our expectations by accounting for the limitations in AI, which in this case, depends on the amount and kind of data and data sources that can be made available by our systems and processes.
  • Having realized this, we can now rejoice that both Human Intelligence and Artificial Intelligence have bounded rationalities in complementary ways.
  • And this can enable us to leverage Augmented Intelligence, a powerful overlap, for co-creation by humans and machines!

On the contrary, traditional methods may continue to work in the realms of the Bounded Rationality of Human Intelligence (step#2 of the expectation management framework activated).So, can Artificial Intelligence bounded by the Data Bubble deliver a value-add for F&B, CPG, and FMCG businesses (step#3 of the expectation management framework activated).

  • Let’s say that within the limitations of the Data Bubble at your organization, your AI application can analyze 300-400 probable consumer trends.
  • And as your ML Model continues to unlearn and learn, the application can pinpoint to 20-30 trends (out of the initial list of 400), with a certain degree of confidence, for further ratification by your team’s Human Intelligence. A perfect interplay of Augmented Intelligence!
  • Hence, an AI tool can help you filter out the noise by extracting insights from the data beyond your bounded rationality. With this, your team can have a better-informed shot at success.

Expectation #3: Speedy Learning Curve

Adopters think AI and ML models, by design, start interpreting like humans right from the word go.

The input data and output, in this case, are generally labeled. Akin to brain development in human beings, from infancy to the adult stage, Machine Learning (ML) Models also need data and time to achieve their optimum decision-making capabilities. Again following step 1 of the expectation management framework, let’s evaluate the learning curve of any ML Model.

Machine Learning Models either follow a Supervised or an Unsupervised learning curve.

In the Supervised learning method, between the input data and the associated output, there is always a mathematical equation or a well-defined correlation. The ML model observes and learns from the data made available to it.

My decade long experience in the Food and Beverage industry makes me believe that the “experience” and “intuition” of your teams will continue to play a critical role even after switching to the “data-driven models” of AI.

© 2020—2022, AI Palette