The “no-code AI” approach to productized business automation

Like anything related to Artificial Intelligence (AI) nowadays, separating hype from reality has become an increasingly difficult task. While the public is mesmerized by low-impact sci-fi AI applications that never make it to market, the efforts to productize and bring value from AI are moving forward from behind the radar – driven by real business requirements. Such valid industry requirements resulted in the recent emergence of the so-called “no code AI” platforms. 

It is already known that a large percentage of AI projects never make it out of the “POC graveyard” (close to 90% according to specialists). The main reason is that AI is still experimenting with problems it can actually solve, but there is also the underlying issue of productizing AI. Data scientists are accustomed to building AI models, but the industry requires stable, secure, production-ready AI products. The models vs products approach to AI is what actually separates failure from successes, and that’s why companies that specialize in producing productized AI, such as RWS, Expert.AI, Cedat85, Bertin IT, Phonexia or Basis Tech, are increasing their market share each year with their Machine Translation (MT), Text Analytics and Speech Recognition products. 

But ready-to-use AI software has its caveats. As you narrow down the use case, the quality and performance of these solutions will decrease. Adaptation, retraining or even creating new AI models based on new production data is required. Up until a couple of years ago, performing a Machine Learning (ML) adaptation task was more suited to the software vendor R&D team. But requirements for data privacy and establishing in-house customer expertise resulted in the “no code AI” approach. 

What is “no code AI”?

Simply put, being able to create AI models without writing software code. Or rather being able to create, adapt and deploy an AI model into production without a cohort of AI researchers, data scientists or software developers. Intended for business users with actual business needs, but no technical experience, “no code AI” is a new paradigm that leads to immediate productized results for very specific AI task that already has a good foundation in terms of state-of-the-art and product maturity.

One example is the RWS Language Weaver machine translation platform. With over 120+ language combinations available out of the box, the solution can cover a good 90% of scenarios – from internal communication, customer support, chat, eLearning, warranty claims, technical support or eDiscovery. But there are always specific needs for AI model adaptation. Whether it is because of a very specific use case (like medical devices, financial mergers or chatbots dealing with specific business lines), or a language adaptation requirements (think of a language dialect or a writing style). To get the most of the Machine Translation engine, it would need to be adapted, using the customer’s data, to such domain or language specificity. This is where no-code AI steps in. 

Traditionally, the ML adaptation task would have been performed by the solutions provider, in this case RWS. But what if the customer cannot share the data required for adaptation, or it wants to build its own competence center in order to provide the adaptation service for various departments or beneficiaries? Having a tool that only requires business expertise, but doesn’t require coding skills is the only viable way to handle such requirements. 

Language Weaver comes with an adaptation component that allows the creation of new language pairs (using existing generic combinations) directly by the customer. It is currently the only such offering in the market. The workflow is specific to other no-code AI platforms: upload data, train model, evaluate and deploy. Everything is done using a simple user interface, with no configuration files, no knowledge about deep neural networks or data science skills. The Adaptation process uses proven “recipes” for each language combination, it automatically cleans and prepares the input data and generates results comparable to those of the RWS data science team. The only thing the customer needs to worry is about acquiring or generating high quality data and establishing a continuous improvement process that relies on the out-of-the box Machine Translation engines in the initial phases, but then constantly improves their performance through adaptation.

There has been an eruption of similar no-code AI platforms recently, mostly for NLP use cases, as described in this Fortune article. The trend is going obviously into the direction of customer controlled AI, faster time to production and getting out of the “AI POC graveyard”. However, it only brings value for already proven AI use-cases, where only adaptation is required on top of foundationally sound AI solutions.