Current models of artificial intelligence can already be trained to distinguish objects in pictures – for example, between cars and motorcycles. With Visual Question Answering, AI wants to go one step further. Machines can identify the color of the previously distinguished vehicles. The question to pose for artificial intelligence would no longer be whether a picture is a car or a motorcycle, but rather which of the two vehicles is red.
To relate this to a business case, a Service Cloud solution could be used where a customer sends a picture of the damages on his car to his insurance company. Salesforce AI would initially identify the damages relative to the undamaged parts in advance through color or shade differentiation, and then pre-classify the damage even before sending it to an insurance agent for further evaluation. This streamlines the damage evaluation process to prevent human errors and to make it more efficient.
NLP is using software to process voice (natural language) for further action. In a scenario where a sales representative finishes a customer call and wants to make changes to their customer records, typing is no longer necessary with NLP. The salesperson must simply dictate the call log into his smartphone or iPad, and Salesforce AI would interpret that to automatically make the changes. This means, the AI does not only transcribe the text, but additionally moves the customer through the pipeline, schedule the next appointment, update opportunities, and more.
Challenges: While AI is continuously adapting to real life scenarios, there are several challenges to consider. First, every company has its own specific language, i.e. acronyms, that simplifies and even optimizes internal communication. There is not always a uniform definition; where in one company, a prospect could be defined as a “lead”, while defined as “opportunity” in another. Each company also has different criteria in classifying a lead as “Hot” and which consecutive to-do’s are attached to such a change in the pipeline. Therefore, to enable Einstein as a cross-company application, it must adapt to the company context and enable configurable models to adjust to the company terminology.
In addition, sales representative’s work does not always take place in a quiet office, but rather close to the customer ideally or en route. This can leads to situations with different audio quality, making speech recognition more difficult. Whether it's a note made in the car, a dictated task in the train station, or a simple voice input at a trade fair, Natural Language Processing moves out of the office into a world of background noise. This requires normalizing audio input and suppressing background noise.
Thirdly, when training artificial intelligence, enough data needs to be collected to be representative of the AI predicted outcome. Problem however lies in that, not all data has potential to be accessed, whether it’s due to data security or simply the nonexistence of data. Therefore, AI providers need to ensure enough data is available, especially considering that an AI powered tool can only be as good as the quality and quantity of trained data records.