The Next Wave: How Domain-Specific AI is Set to Redefine the BPM Industry

In Brief:

• Domain-specific AI – honed to meet specialised industry needs – is fast becoming the next big thing in the tech space.
• In this article, we explore how domain-specific AI can be leveraged by the BPM industry.
• We also touch upon the exciting developments happening in the Indian LLM space.

Introduction

By now most of us are well acquainted with AI and its many avatars.

At one level sit speech recognition systems (think Alexa, Siri), image recognition systems (hello Google Lens) and recommendation systems (all those Netflix prompts).

Then there are the heavyweights – machine learning, deep learning, natural language processing – powering the interfaces and applications we now take for granted?

All of them, in tandem, have reshaped the way we do business across domains, functions and geographies.

In the business processing management (BPM) domain itself, it is amazing to see how AI is being leveraged every day to:

• Improve decision-making through data-driven insights
• Enhance end-user satisfaction through personalized experiences
• Reduce operational costs through automation and streamlined processes
• Increase efficiency and productivity through process automation

And now, with the advent of Generative AI - GenAI of the ChatGPT and Gemini fame - new, increasingly sophisticated solutions are coming to market every day.

However, the usage of GenAI is still limited in scope and depth.

If we are to truly transform operations over time, we need a form of AI that is hyper-specific, precise and sharp – able to dig deep into the demands of our particular business case.

Which leads us to the next big thing in the AI world – domain-specific AI: GenAI contextualized for specific industries, businesses and tasks.

Essentially, GenAI transformed from a butter knife to a surgeon’s scalpel.

The Case for Domain-Specific AI

By 2027, over half of the GenAI models used by enterprises will be domain-specific - Gartner.

We briefly touched upon ML/DL earlier, which forms the basis of a large number of models and applications.

One of these applications is the Large Language Models or LLMs – an algorithm that can process and understand human languages or text using self-supervised learning techniques

Now, for specific business needs, general-purpose applications based on LLMs might not be sufficient.

So, the next step would be to take a generic model and train or modify it for a particular field or area of knowledge. This is domain-specific AI.

Domain-specific models would have deep understanding of context, industry data, corporate policies, and industry terminologies. They would:

1. Provide more accurate results as they are trained on field- specific data and are therefore more conversant with the domain’s intricacies, risks, and context.
2. Consume less time and fewer steps than general models to obtain the required information.
3. Factor in the context and language of the domain to provide responses and suggestions that contain specialized terminology and vocabulary.
4. Reduce costs as they would make workflows faster and more efficient, without compromising quality and context.

Domain-Specificity and BPM

When I think about domain-specific AI in the? BPM context? – be it payroll, compliance, AR/AP, HR or other backend solutions – I find the framework presented by Vidgof, Bachhofner and Mendling in their paper on the use of LLMs for Business Process Management helpful.?

The paper categorizes BPM activities in 3 phases: identification, discovery, and analysis.

Identification

This is the unstructured part of the process, where a domain-specific model can be put to use extracting relevant information from heterogeneous internal documentation and logs.

The goal is for the model to glean what kind of processes the firm is undertaking, and rank these in order of strategic importance.?

Discovery

At this stage one or more process discovery methods is selected to produce process models.

When automating process discovery, process mining is the way to go – a technique of extracting process models and other relevant data out of event logs left by information systems supporting the execution of a process.

However, now domain-specific AI excels at sifting through large volumes of information and making sense of them – so process discovery can unfold via document analysis.

Further, it can extract patterns from communication logs: emails and chats between internal and/or external stakeholders. It can then convert this to human-readable narratives that explain trends, anomalies, and insights.

Finally, going beyond evidence-based discovery, sometimes firms conduct interviews with domain experts before producing a process model based on several interviews. Domain-specific AI can solve parts of this problem by providing a chatbot interface for domain experts.

In this way, the domain experts answer questions in the chat and responses can be collated and analysed in a quicker way.

Analysis

Now, the discovered processes are analyzed to find problems and bottlenecks. While this is a cognitively loaded task, domain-specific AI can assist human analysts.

If an issue exists in a process, it's likely that somebody has already complained about it. AI can be put to work on scraped data from social media, support service or internal communication tools to find patterns and gaps.

After an issue in the process is found, the next step is to spot the part of the process that creates this issue. The task of the AI is to analyze task names and descriptions to make suggestions on tasks that may be responsible for the issue.

In advanced cases, domain-specific AI might be even capable of suggesting some fixes – such as new business strategies based on current market trends, or competitor analysis.

The Way Forward

The next steps are clear – adoption of a domain-specific AI is an investment in your growth and future.

There are three possible (but not equally probable) methods to achieve this goal.

1. Start From Scratch – This was the older method of building a unique model from the ground up, investing huge amounts of time, effort and money just to bring the model up to par with general-purpose models. On the grounds of efficiency, this might not be the best bet. By 2028, more than 50% of enterprises that have built their own large language models (LLMs) from scratch will abandon their efforts due to costs, complexity and technical debt - Gartner
2. Fine-tuning An Existing Model – This entails picking an existing foundation model of your choice (such as Meta’s LLama 3.1) and building on top of it. The aim is to optimize the model using domain-specific knowledge ( fine-tuning, RAG, etc.), but most businesses will still find this a relatively arduous, expensive and inefficient undertaking.
3. Enriching Prompts – The simplest option would be to leverage a plug-and-play domain-specific model that can immediately meet your requirements. You can also enrich the model with your private data, making it unique to your needs and use cases.

An important footnote, if you are seeking to expand AI adoption in India, you may want to explore indigenous LLMs being developed within the country.

The surrounding environment plays an important role in the development and adoption of LLMs. And, as the CEO of an Indian AI startup pointed out recently, global LLMs are trained on data which at times may not capture the nuances of the subcontinent.

Indigenous LLMs, still in pilot phases, may offer a more India-focused foundational model.

Also, of late, I have been reading reports about Indian players looking beyond generic LLMs and training products on domain-specific datasets in fields such as healthcare, financial services and law. This targeted approach could yield quick dividends.

Whatever route you choose, the potential competitive advantage domain-specific AI represents is huge. It is an opportunity you cannot afford to ignore.