The role of the Product Owner is one like no other. By combining a variety of skills, this figure is committed to maximising the value of the product and the work of the team behind it, while ensuring that customers and stakeholders’ needs are satisfied. The PO can be seen as the voice of the client inside the company, and as such needs to understand what it takes to adopt a product from the ground up, in order to make the process smooth and conscious for the client. Being this figure so related to the product, there can be significant changes between different types of products, particularly when AI is involved.
To better understand the peculiarities of this role in terms of projects, we got in touch with Alfonso Fraire, Technology Product Owner at the lab of Konica Minolta Digital Services R&D in Rome, and asked him some questions related to his job and the challenges he faces everyday working with Machine Learning technology.
LET’S START WITH THE BASICS, WHAT DOES YOUR JOB AT KONICA MINOLTA CONSIST OF?
Well, typically, a Product Owner in a company that deals with software ensures that the stakeholders’ needs are translated in user stories and inserted in the backlog, and that all the different decisions taken by the team create value for the client. But, in our specific case, each team is treated as a separate entity, like a small new company, which builds a whole new product concept from the ground up, tests its suitability in the product market, and then proceeds to the productisation of the solution. This means that I must also double as a Project Manager together with our clients and to find the perfect fit for the product market. Given the technology behind our products, which are of course AI-based, there are different variables to take in consideration.
WHICH VARIABLES ARE WE TALKING ABOUT?
It really depends on the product, but AI-based solutions are very sensitive, for example, to environmental conditions. Assessing the state of lighting, knowing if the use-case scenario is indoors or outdoors, finding the best spot (if present) to anchor a camera, ensuring that Wi-Fi coverage is good enough and knowing if the solution will be needed in case of adverse weather conditions play a key role in the acquisition of the dataset and on the efficiency of the solution itself.
That’s why you need to manage the expectations of the organisations you work with, and as a PO this is one of the hardest tasks to accomplish. In fact, being Machine Learning’s results mostly nondeterministic, the first important point to make is that the solution will not work on 100% of the cases: realistically, the AI will work 93% to 99% of the times, depending on how good and accurate the Machine Learning model is and on the quality of the available dataset. It is for this reason that we don’t work on mission critical projects that could put people at risk, and also why we carefully qualify the prospects we decide to work with, working as consultants to identify the best use-case for an AI-based solution.
AND ONCE THE QUALIFICATION HAS BEEN DONE (WITH A POSITIVE RESULT), WHAT ARE THE MAIN CHALLENGES THAT YOU FACE IMPLEMENTING AI-BASED TECHNOLOGIES?
Probably the biggest challenge is the need for a huge number of data and information to properly train our AI models. To identify a specific object for instance, we usually need 1,000 to 1,500 pictures of that object to build a reliable dataset, and it is not always easy to get them.
For instance, we are currently working on a model that can recognise children on luggage belts of an airport, but finding thousands of pictures of children on conveyor belts is not as easy as it may seem. In cases like these, we rely on synthetic datasets, which work with artificial data that is generated from the original data and a model trained specifically to reproduce the characteristics and structure of the original.
At this point comes the most tedious part of the process: the labelling. For each photo of the dataset, we have to manually label the area that the AI has to take into consideration. But in order to make this process easier and less time consuming, we are developing AI Studio, a new all-in-one software that enables one to choose the dataset and the algorithm and easily go through the labelling and training phase.
SO, AI IS GOING TO BE MORE ACCESSIBLE SOON, ISN’T IT?
Yes, exactly. But if I may give some advice to an organisation that is looking to adopt AI-based solutions for the first time, I’d tell them to be open to Co-Creation. Reach out to a team of experts, explain to them what your specific needs are and explore the solutions they propose, keeping in mind that those solutions can be progressively adapted.
For instance, there has been a case for our Smart Mirror where the accuracy could only get as far as 85%. Working along with the client, we were able to improve the reliability of the solution, increasing the accuracy by 10%, balancing precision and recall data, enabling, in the specific case, workers to get through the compliance checks even if their hair were slightly exposed.
In fact, it is possible to favour false positives and false negatives to meet clients’ specific needs. Again, what is important is that they are aware of the limits to levels of accuracy: after a while, diminishing returns or no returns at all can make further training unnecessary, but being conscious of it can really make a difference. It encourages an open attitude towards new paths and perspectives and can lead to an effective solution in the perfect spirit of Co-Creation.