How Businesses Actually Use AI

Dreams and Reality

The classic movie 2001: A Space Odyssey predicted colonisation of space, alien artifacts and a self-aware computer by the dawn of the 21st century. Even experts such as the story’s creator, Arthur C. Clarke, may see the direction of travel but not the speed. Film entertainment is one thing, but are blogs, books and posts describing how AI is changing the world any more accurate?

Optimists champion designer proteins that fight diseases, robots that clean oceans, and brain computer interfaces (BCI) that cure blindness. There are 250 BCI companies with a cumulative $2.3 billion in funding, so there is a lot of optimism around. This is an exciting time to be in the startup economy, but what are established businesses using AI to do?

Respondents to an Andreessen Horowitz survey revealed that saving costs is only a priority 10% of the time. Businesses would rather focus on problem solving to open new markets and create operational scale. The common themes across industries are simulation, prediction, personalisation, report summarisation and customer assistance.

In his book How Innovation Works, Matt Ridley debunks the myth of the creative genius. The same challenges arise across industries at similar times, leading numerous companies to seek solutions. The winners aren’t those who get there first, but those who see the full potential of a product.

As a result, most innovation improves existing practices rather than introducing revolutionary new ones. Of course, new products and services do spring from attempts to solve internal problems. Amazon created Amazon Web Services from a solution for an internal bottleneck, but most companies end up as customers not creators of these breakthroughs.

This is because creating knowledge is hard. Only a handful of companies do it and turn it into successful products. NVIDIA is an exceptional company, at the forefront of knowledge about computing hardware and developing software that allows clients to build their own solutions. Enterprises team up with NVIDIA and its service partners to innovate in the field of machine learning and AI.

Common Data Uses

There is a lot of data – equivalent to 800,000 libraries of Congress by 2027, according to IDC Global Datasphere. Over 80% of it is unstructured and half is either audio or visual. This allows for multimodal AI, in which inputs and outputs are a variety of media. Think text-to-speech, speech-to-audio and audio-to-video. Yet regardless of source, extracting information from data and applying it productively is the core of all AI use cases.

30% of all data generated is in the healthcare sector. As a result, both AI hardware and software providers are focused on delivering services to the industry. These support medical imaging and devices, drug discovery, genomics and digital health. That last category includes customer facing technology, records, and deciphering technical papers.

Other industries at the forefront of AI adoption are automotive, energy, financial services, higher education, manufacturing, retail and telecommunications. While autonomous vehicles grab the headlines in automotive, AI is as prevalent for natural language processing in customer service, demand prediction and process efficiency. Generative AI enables synthetic data and digital twins to accelerate and innovate vehicle design.

As we look across industries we see similar workflows emerging. In energy, AI simulates geological formations to predict the presence of fossil fuels, optimises drilling, and predicts maintenance. Prediction plays an important role in financial services in loss estimation and fraud prevention, and in manufacturing it helps control inventory and monitor machinery. Prediction creates a wealth of cost efficiencies and demand generation opportunities, without threatening a large number of existing jobs.

For all the science fiction outcomes that grab the headlines, a standard list of AI workflows includes virtual assistants and chatbots, knowledgeable copilots, code review and generation, content generation, data analysis and reporting, and language translation. Of those chatbots, copilots and code generation are most common.

Chatbots use pre-trained models to answer general queries. While the Andreesen Horowitz survey revealed that the majority of enterprises are using three or more models, over two-thirds of large language models in production are based on OpenAI.

At the end of 2024, OpenAI revealed that its reasoning models perform at the upper levels of human achievement in coding and PhD exams. Reasoning requires large amounts of computational power and energy, and happens during inference, pushing the cost onto end users. This is a path to profitability for OpenAI, but may limit early access to deep-pocketed enterprises.

Making Money from AI

To make money from AI, businesses need to bring their intellectual property to bear. This means training and running models on proprietary systems. This is where we transition from chatbot to copilot, with the latter being an assistant performing tasks in a specific domain.

GitHub has a successful copilot, which Microsoft is attempting to replicate in Excel. Construction companies train models to provide on-demand health and safety education, while retail stores generate avatars to guide customers with predictive analytics to optimise the buying journey. Copilots squeeze more from existing technologies and the trend is to focus on new opportunities rather than cost cutting.

There is an important caveat, however. Chatbots and copilots are only as effective as the data in their training models. Data is often incomplete or irrelevant and data scientists have been shown to make mistakes such as data leakage, overfitting or coding errors. The most effective way to start using AI is to automate a process that you know works.

The aim of chatbots, copilots and code generation is to increase the quantity and speed of workloads. This will generate demand with superior and personal service, and design new products through simulation and testing. The intention is that staff without technical training will be able to use AI models. Imagine testing and creating products, predicting and generating demand, and serving and maintaining customers, all without writing code.

In most use cases, AI accelerates what we already do today. This has the highest probability of a positive return on investment. For companies with the ambition to go further, let’s look at where it is realistic to close the gap between the AI optimists and reality.

Real-time Decision Making

At CES in January 2025, NVIDIA CEO Jenson Huang name checked Meta for transforming enterprise workflows. Meta’s Llama open source models are foundational for the AI reinventing corporations. NVIDIA hopes to do the same for robotics.

The simplest, successful business idea is meeting a customer’s needs. Scaling that idea requires interpreting feedback. To do this, managers want to make data-driven decisions.

Right now, data-driven means examining historic behaviours to project the future. AI changes that across a range of industries. Data will be real-time and predicted, allowing companies to make informed decisions about customer demand, cash flow generation and product performance. Processes will be recreated from ground level, using spontaneous scenario analysis.

Synthetic data holds the key to this transformation. Tech leaders talk about running out of data, but what they mean is there is not enough useful data to train models. AI generates relevant new data by making minor adjustments to real world text, audio and video. What will companies do with this data?

Scaling for Simulation

Huang talked about two new ways of scaling intelligence beyond training data. The first uses AI to critique AI in a process known as reinforcement learning and the second is reasoning. This means scenario testing solutions at the speed of machine thought, which requires the massive compute NVIDA’s hardware is designed to handle.

As an example, Huang says robots need three models – one to train, one to deploy and one to simulate. That third one is where the transformation occurs.

Training equips robots with core skills, while the decision about what to do next uses real-time scenario testing. At a junction, a car can simulate outcomes and take the best course of action. This on-the-job training will be what allows driverless cars to function in crowded cities.

Expensive heavy trucks are being retrofitted as autonomous vehicles, monitored through digital twins and remotely controlled. Humanoid robots working production lines are a few years away, but they can already haul rubbish into skips and load vehicles in the factory yard. What’s missing is enough video of humans doing this. To solve the problem, people record themselves performing a task in virtual reality and AI generates multiple versions with slight variations to provide comprehensive datasets.

This idea is not limited to robots. Synthetic data is used to model protein behaviour for drug discovery, control traffic flow on roads and in stadiums, and test customer service agents for bias. Simulation allows real-time adjustments to behaviour and, for example, may be how automated medical assistants develop their bedside manner.

A New Internet

It’s a common misconception that Microsoft missed mobile. On the contrary, it attempted to shoehorn Windows onto mobile, but this required too much memory and was abandoned in 2017. 99% of cellular phones use either Apple iOS or open source Android, which were purpose built for mobile.

Generative AI promises a new operating system for the internet. This will choose and perform actions at a moment in time, obliterating the idea of traditional applications. It may also restore the internet to its original conception as a decentralised communication network, before Google, cloud service providers, and the large telcos took control of traffic.

A related use case is a decentralised energy grid. Renewables and storage are turning energy into a technology industry, where prices fall through time. An AI operated grid will balance load and provide energy independence to small communities.

Wearables may be the entry point for a GenAI operating system. Meta has shown with Llama that it has no intention of being trapped by a competitor, as it is with iOS. Meta Quest has its own operating system, built on Android and optimised for virtual reality.

At the moment, GenAI is built into operating systems, for example with Windows Copilot. NVIDIA customises Linux into DGX OS for AI workloads on its hardware. The need for low latency and energy efficiency to run AI at device level, may be the trigger for the next generation internet.

From Small Automations to Reinventing the Workplace

Most business innovation involves small improvements to the way we work each day. This may be automated filing of emails, data entry, and searching the web for missing information. These are mundane tasks that workers do every day and which have been automated thousands of times. The benefits of automation are lower costs, fewer errors and happier humans.

There is a lot more that can and will be done using AI. The first and most important steps are to have enough of the right data and to know how to use it. AI can help in generating synthetic data and in correcting the common errors in data science. Agentic AI is the preferred path to integrating these systems today. We remain a long way from developing a self aware computer like HAL in 2001, but are capable of reinventing the workplace in ways that were unimaginable when Arthur C. Clarke was writing.