Making it worth your money: Why businesses are slow to embrace AI at scale

Executives are hesitant to scale generative artificial intelligence beyond proof-of-concepts, despite its promise for enhancing productivity. Here’s why.

Nam Je Cho, director of solutions architecture at AWS Australia and New Zealand, says common concerns include data privacy, output accuracy, unclear return on investment, and potential legal and regulatory implications.

Many AWS customers I’ve spoken to had been hesitant to scale generative artificial intelligence beyond proof-of-concepts, despite its promise for enhancing productivity.

Common concerns include data privacy, output accuracy, unclear return on investment, and potential legal and regulatory implications.

To build confidence and clarity, customers are implementing robust AI governance, policies and standards, clear usage guidelines and deliberate rollouts. However, demonstrating clear ROI with AI to justify project costs, especially at the C-suite and board levels, remains a significant hurdle.

This challenge stems partly from the difficulty in quantifying the productivity gains in knowledge work that generative AI optimises. For example, how do you translate a reduction in resolution time from 10 hours to one hour by a HR chatbot into business value? Without this, calculating ROI to convince boards to invest further becomes challenging.

And without business value, how do you calculate the exact and deliberate ROI to convince the board to invest further? Given these challenges, companies are increasingly exploring a variety of AI solutions to find the right balance of performance, cost and ease of implementation.

Powerful AI models from Anthropic, Mistral, Meta and of course Amazon (we just announced our Amazon Nova family of advanced models at AWS re: Invent) are making generative AI more accessible than ever. These models can produce text, from creative writing to code generation, as well as trend analysis to language translation, video analysis and image/video creation, driving increased productivity and creativity.

Over the past 12 months, as customer generative AI adoption has expanded on Amazon Bedrock, our fully managed service offering for building generative AI application, customers have reinforced that broad and flexible model choices, guardrails for safety, knowledge base and other key features to simplify building AI applications are important to tackling business problems with generative AI. Today, tens of thousands of customers are using Amazon Bedrock to build generative AI applications to solve a wide variety of business problems across every industry vertical.

We’re seeing that getting the best results from models isn’t just about selecting the latest and greatest. Combining fit-for-purpose models with best-practice prompting techniques, often called prompt engineering, can produce significantly better results in terms of accuracy and cost-effectiveness. One impactful technique is named multi-shot prompting. By sharing multiple examples of desired outcomes, users can effectively calibrate the model for each use case, resulting in better accuracy, consistency, cost and performance.

Another approach to level-up customers’ generative AI game is through retrieval-augmented generation. AI models are trained on specific data and their knowledge does not extend beyond the data used for training.

RAG complements a model’s knowledge by providing more up-to-date or context-specific data, or both, and grounds model responses in that data to increase accuracy and relevance, reducing follow-up human intervention. For example, PEXA recently launched an internal generative AI Assistant that leverages real-time company data using RAG and Amazon Bedrock, to ensure every employee chat interaction is secure, accurate and contextually relevant.

While optimal prompting and RAG are powerful, they are not a panacea. Model choice remains paramount and we maintain there is no one model that will rule them all. That’s why we’ve just launched six new Amazon Nova models that deliver industry-leading price-to-performance, to expand the growing selection of the broadest and most capable models in Amazon Bedrock for customers.

While most tasks can be performed by the most sophisticated models, using a model that is too sophisticated for the task will cost more and cause higher latency; the trick is to select the model that is “just right” for a given task, which usually means the smallest, cheapest, fastest model that can perform it.

Human oversight, curation and feedback loops are also essential to ensure the right quality outcome and adhering to responsible AI principles. No generative AI system today is reliable enough to fully automate the end-to-end business process. This human-AI collaboration is key as we work towards more robust, responsible and trustworthy AI systems.

To make it as successful as possible, we need to train the workforce with the right skills. We’ve already trained more than 400,000 people in Australia in cloud skills, and we have a dedicated global Amazon goal to provide free AI skills training to two million individuals by 2025.

The demand is high in Australia, with recent research commissioned by AWS finding 84 per cent of Australian employers and workers expect to use generative AI tools on the job within the next five years, driving 45 per cent more productivity.

We are committed to democratising access and training to generative AI, robust tools for responsible AI development, and initiatives for Australian customers and partners to deliver a more sustainable future.

Responsible use of these technologies is key to fostering continued innovation. One way we’re doing this is providing customers with the tools and guidance needed to build and scale generative AI safely, securely, and responsibly, ensuring Australia is an AI global leader in the digital economy.

Nam Je Cho is director of solutions architecture at AWS Australia and New Zealand.

Originally published as Making it worth your money: Why businesses are slow to embrace AI at scale