In the realm of business support, the utilization of AI presents both remarkable opportunities and unique challenges. While our previous discussions have highlighted the potential of AI in transforming enterprise support and measuring impact, we shift our focus to the complexities of contextualization, especially for Micro, Small, and Medium Enterprises (MSMEs) in Low and Middle-Income Countries (LMICs). While advancements in AI have been highly-publicized, there are still many limitations that exist for billions of people around the world that don't speak English - and therefore miss out on so much potential until this gap is crossed.
The Language Barrier: Bridging the Gap
One of the most formidable challenges in AI-based business coaching is the prevalence of low-resource languages:
Diverse Linguistic Landscape:Â Many MSMEs operate in regions where local languages are underrepresented in AI models, and they simply do not perform well outside of English and especially Latin-based scripts
Cultural Nuances:Â Beyond language, understanding and integrating cultural contexts is crucial for effective AI coaching. Language is only one layer of high-quality business coaching. We need to integrate cultural nuance to achieve truly great results for our users.
Harnessing Prompt Engineering
While the industry begins to make progress with low-resource language LLMs, we have been experimenting with relatively simple prompt engineering as a means to improve output for our users.
Tools like ChatGPT have been tuned for the English language, including difficult-to-translate ideas like idioms, puns, and jargon. While immediately recognizable to native speakers, phrases like "the spice of life" for a food stand business are unlikely to perform well when we try to translate them across languages and cultures.
At ONOW, we've found great success with some simple, but incredibly powerful phrases in our prompt engineering processes:
"Avoid all idioms"
"Avoid all jargon"
"Use simple language"
Additionally, we have combined these techniques with features like few-shot prompting, and multiple agents that review output to assess content and translation quality. By including a suite of tools, we're building a playbook of components to bring high-quality AI output to more languages than expected.
And by automating the process, we can deliver increasingly high-quality and personalized material for people that simply don't have access to powerful tools within ChatGPT.
User Experience (UX) and Channel Delivery Innovations
And while improved content and language is incredibly important to success, the delivery of that coaching is just as important:
Facebook Messenger: Our past experience in Myanmar has shown increased interactions simply by providing our tools within familiar platforms like Facebook Messenger. Ubiquitous in Myanmar, this platform minimizes friction and leads to much greater rates of adoption.
Channel Limitations: While messaging channels have many limitations, we have also found the constraints imposed force a degree of creativity in our innovation and building process. These limitations can be incorporated into the prompt engineering process described above and create a true product that's designed for our users. It isn't simply a clone of ChatGPT. Rather - it's a fully-formed product that uses AI, data analysis, and traditional elements as components of a whole.
Additional Channels: As ONOW is expanding into more countries, we're taking this learning to additional channels like Line, Telegram, SMS, and web applications.
Anticipating Future Advancements
As Language Models (LMs) evolve, we anticipate several improvements. We're doing our best to stay connected to this quickly evolving field that seemingly updates by the day. A couple things we see on the horizon already:
Expansion in Language Coverage:Â Upcoming LMs are expected to include a broader array of languages, reducing the bias towards English and other major languages.
Cultural Sensitivity:Â Future models are being developed with a deeper understanding of cultural nuances, crucial for avoiding biases and ensuring relevance in diverse settings.
Swarm-Like Intelligence: While LLMs are not a panacea for the complex problems of business coaching within developing countries, they can represent a powerful node that exists with multiple other models like machine translation, cultural sensitivity experts, search engines and more to create the illusion of an intelligent being that pulls from multiple tools when forming its advice for business owners.
Conclusion: A Forward-Looking Approach to AI in Business Coaching
The journey towards integrating AI into business coaching for MSMEs in LMICs is filled with challenges, yet it holds immense potential. As we look forward to advancements in LLMs and their applications, a key part of our focus will remain on overcoming the hurdles of contextualization and language diversity. As we seek to provide a coach for every business, we believe these tools will play a key part of the process by bringing high-quality solutions in hundreds of languages to developing businesses worldwide.