What’s next for AI?

Now: Getting the fundamentals right

LLMs are undoubtedly exciting but require a great deal of groundwork. Instead of building models themselves, many enterprises are partnering with companies like Anthropic or OpenAI or accessing AI models through hyperscalers.⁴ According to Gartner®, AI servers will account for close to 60% of hyperscalers’ total server spending.⁵ Some enterprises have found immediate business value in using LLMs, while others have remained wary about the accuracy and applicability of LLMs trained on external data.⁶ On an enterprise time scale, AI advancements are still in a nascent phase (crawling or walking, as we noted last year). According to recent surveys by Deloitte and Fivetran and Vanson Bourne, in most organizations, fewer than a third of generative AI experiments have moved into production, often because organizations struggle to access or cleanse all the data needed to run AI programs.⁷ To achieve scale, organizations will likely need to further think through data and technology, as well as strategy, process, and talent, as outlined in a recent Deloitte AI Institute report.

According to Deloitte’s Q3 2024 State of generative AI in the enterprise report, 75% of surveyed organizations have increased their investments in data-life-cycle management due to generative AI.⁸ Data is foundational to LLMs, because bad inputs lead to worse outputs (in other words, garbage in, garbage squared). That’s why data-labeling costs can be a big driver of AI investment.⁹ While some AI companies scrape the internet to build the largest models possible, savvy enterprises create the smartest models possible, which requires better domain-specific “education” for their LLMs. For instance, LIFT Impact Partners, a Vancouver-based organization that provides resources to nonprofits, is fine-tuning its AI-enabled virtual assistants on appropriate data to help new Canadian immigrants process paperwork. “When you train it on your organization’s unique persona, data, and culture, it becomes significantly more relevant and effective,” says Bruce Dewar, president and CEO of LIFT Impact Partners. “It brings authenticity and becomes a true extension of your organization.”¹⁰

Data enablement issues are dynamic. Organizations surveyed by Deloitte said new issues could be exposed by the scale-up of AI pilots, unclear regulations around sensitive data, and questions around usage of external data (for example, licensed third-party data). That’s why 55% of organizations surveyed avoided certain AI use cases due to data-related issues, and an equal proportion are working to enhance their data security.¹¹ Organizations could work around these issues by using out-of-the-box models offered by vendors, but differentiated AI impact will likely require differentiated enterprise data.

Thankfully, once the groundwork is laid, the benefits are clear: Two-thirds of organizations surveyed say they’re increasing investments in generative AI because they’ve seen strong value to date.¹² Initial examples of real-world value are also appearing across industries, from insurance claims review to telecom troubleshooting and consumer segmentation tools.¹³ LLMs are also making waves in more specialized use cases, such as space repairs, nuclear modeling, and material design.¹⁴

As underlying data inputs improve and become more sustainable, LLMs and other advanced models (like simulations) may become easier to spin up and scale. But size isn’t everything. Over time, as methods for AI training and implementation proliferate, organizations are likely to pilot smaller models. Many may have data that can be more valuable than previously imagined, and putting it into action through smaller, task-oriented models can reduce time, effort, and hassle. We’re poised to move from large-scale AI projects to AI everywhere, as discussed in this year’s introduction.

New: Different horses for different courses

While LLMs have a vast array of use cases, the library is not infinite (yet). LLMs require massive resources, deal primarily with text, and are meant to augment human intelligence rather than take on and execute discrete tasks. As a result, says Vivek Mohindra, senior vice president of corporate strategy at Dell Technologies, “there is no one-size-fits-all approach to AI. There are going to be models of all sizes and purpose-built options—that’s one of our key beliefs in AI strategy.”¹⁵

Over the next 18 to 24 months, key AI vendors and enterprise users are likely to have a toolkit of models comprising increasingly sophisticated, robust LLMs along with other models more applicable to day-to-day use cases. Indeed, where LLMs are not the optimal choice, three pillars of AI are opening new avenues of value: small language models, multimodal models, and agentic AI (figure 1).

Small language models

LLM providers are racing to make AI models as efficient as possible. Instead of enabling new use cases, these efforts aim to rightsize or optimize models for existing use cases. For instance, massive models are not necessary for mundane tasks like summarizing an inspection report—a smaller model trained on similar documents would suffice and be more cost-efficient.

Small language models (SLMs) can be trained by enterprises on smaller, highly curated data sets to solve more specific problems, rather than general queries. For example, a company could train an SLM on its inventory information, enabling employees to quickly retrieve insights instead of manually parsing large data sets, a process that can sometimes take weeks. Insights from such an SLM could then be coupled with a user interface application for easy access.

Naveen Rao, vice president of AI at Databricks, believes more organizations will take this systems approach with AI: “A magic computer that understands everything is a sci-fi fantasy. Rather, in the same way we organize humans in the workplace, we should break apart our problems. Domain-specific and customized models can then address specific tasks, tools can run deterministic calculations, and databases can pull in relevant data. These AI systems deliver the solution better than any one component could do alone.”¹⁶

An added benefit of smaller models is that they can be run on-device and trained by enterprises on smaller, highly curated datasets to solve more specific problems, rather than general queries, as discussed in “Hardware is eating the world.” Companies like Microsoft and Mistral are currently working to distill such SLMs, built on fewer parameters, from their larger AI offerings, and Meta offers multiple options across smaller models and frontier models.¹⁷

Finally, much of the progress happening in SLMs is through open-source models offered by companies like Hugging Face or Arcee.AI.¹⁸ Such models are ripe for enterprise use since they can be customized for any number of needs, as long as IT teams have the internal AI talent to fine-tune them. In fact, a recent Databricks report indicates that over 75% of organizations are choosing smaller open-source models and customizing them for specific use cases.¹⁹ Since open-source models are constantly improving thanks to the contributions of a diverse programming community, the size and efficiency of these models are likely to improve at a rapid clip.

Multimodal models

Humans interact through a variety of mediums: text, body language, voice, videos, among others. Machines are now hoping to catch up.²⁰ Given that business needs are not contained to text, it’s no surprise that companies are looking forward to AI that can take in and produce multiple mediums. In some ways, we’re already accustomed to multimodal AI, such as when we speak to digital assistants and receive text or images in return, or when we ride in cars that use a mix of computer vision and audio cues to provide driver assistance.²¹

Multimodal generative AI, on the other hand, is in its early stages. The first major models, Google's Project Astra and OpenAI’s GPT-4 Omni, were showcased in May 2024, and Amazon Web Services’ Titan offering has similar capabilities.²² Progress in multimodal generative AI may be slow because it requires significantly higher amounts of data, resources, and hardware.²³ In addition, the existing issues of hallucination and bias that plague text-based models may be exacerbated by multimodal generation.

Still, the enterprise use cases are promising. The notion of “train once, run anywhere (or any way)” promises a model that could be trained on text, but deliver answers in pictures, video, or sound, depending on the use case and the user’s preference, which improves digital inclusion. Companies like AMD aim to use the fledgling technology to quickly translate marketing materials from English to other languages or to generate content.²⁴ For supply chain optimization, multimodal generative AI can be trained on sensor data, maintenance logs, and warehouse images to recommend ideal stock quantities.²⁵ This also leads to new opportunities with spatial computing, which we write about in “Spatial computing takes center stage.” As the technology progresses and model architecture becomes more efficient, we can expect to see even more use cases in the next 18 to 24 months.

Agentic AI

The third new pillar of AI may pave the way for changes to our ways of working over the next decade. Large (or small) action models go beyond the question-and-answer capabilities of LLMs and complete discrete tasks in the real world. Examples range from booking a flight based on your travel preferences to providing automated customer support that can access databases and execute needed tasks—likely without the need for highly specialized prompts.²⁶ The proliferation of such action models, working as autonomous digital agents, heralds the beginnings of agentic AI, and enterprise software vendors like Salesforce and ServiceNow are already touting these possibilities.²⁷

Chris Bedi, chief customer officer at ServiceNow, believes that domain- or industry-specific agentic AI can change the game for humans and machine interaction in enterprises.²⁸ For instance, in the company’s Xanadu platform, one AI agent can scan incoming customer issues against a history of incidents to come up with a recommendation for next steps. It then communicates to another autonomous agent that’s able to execute on those recommendations, and a human in the loop reviews those agent-to-agent communications to approve the hypotheses. In the same vein, one agent might be adept at managing workloads in the cloud, while another provisions orders for customers. As Bedi says, “Agentic AI cannot completely take the place of a human, but what it can do is work alongside your teams, handling repetitive tasks, seeking out information and resources, doing work in the background 24/7, 365 days a year.”²⁹

Finally, aside from the different categories of AI models noted above, advancements in AI design and execution can also impact enterprise adoption—namely, the advent of liquid neural networks. “Liquid” refers to the flexibility in this new form of training AI through a neural network, a machine learning algorithm that mimics the human brain’s structure. Similar to how quantum computers are freed from the binary nature of classical computing, liquid neural networks can do more with less: A couple dozen nodes in the network might suffice, versus 100,000 nodes in a more traditional network. The cutting-edge technology aims to run on less computing power, with more transparency, opening up possibilities for embedding AI into edge devices, robotics, and safety-critical systems.³⁰ In other words, it’s not just the applications of AI but also its underlying mechanisms that are ripe for improvement and disruption in the coming years.

Next: There’s an agent for that

In the next decade, AI could be wholly focused on execution instead of human augmentation. A future employee could make a plain-language request to an AI agent, for example, “close the books for Q2 and generate a report on EBITDA.” Like in an enterprise hierarchy, the primary agent would then delegate the needed tasks to agents with discrete roles that cascade across different productivity suites to take action. As with humans, teamwork could be the missing ingredient that enables the machines to improve their capabilities.³¹ This leads to a few key considerations for the years to come (figure 2):

• AI-to-AI communication. Agents will likely have a more efficient way of communicating with each other than human language, as we don’t need human-imitating chatbots talking to each other.³² Better AI-to-AI communication can enhance outcomes, as fewer people will need to become experts to benefit from AI. Rather, AI can adapt to each person’s communication style.³³
• Job displacement and creation. Some claim that roles such as prompt engineer could become obsolete.³⁴ However, the AI expertise of those employees will remain pertinent as they focus on managing, training, and collaborating with AI agents as they do with LLMs today. For example, a lean IT team with AI experts might build the agents it needs in a sort of “AI factory” for the enterprise. The significant shift in the remaining workforce’s skills and education may ultimately reward more human skills like creativity and design, as mentioned in previous Tech Trends.
• Privacy and security. The proliferation of agents with system access is likely to raise broad concerns about cybersecurity, which will only become more important as time progresses and more of our data is accessed by AI systems. New paradigms for risk and trust will be required to make the most out of applying AI agents.

• Energy and resources. AI’s energy consumption is a growing concern.³⁵ To mitigate environmental impacts, future AI development will need to balance performance with sustainability. It will need to take advantage of improvements in liquid neural networks or other efficient forms of training AI, not to mention the hardware needed to make all of this work, as we discuss in “Hardware is eating the world.” 
• Leadership for the future. AI has transformative potential, as everyone has heard plenty over the last year, but only insofar as leadership allows. Applying AI as a faster way of doing things the way they’ve always been done will result in, at best, missed potential, and, at worst, amplified biases.³⁶ Imaginative, courageous leaders should dare to take AI from calcified best practices to the creation of “next practices,” where we find new ways of organizing ourselves and our data toward an AI-enabled world.

When it comes to AI, enterprises will likely have the same considerations in the future that they do today: data, data, and data. Until AI systems can reach artificial general intelligence or learn as efficiently as the human brain,³⁷ they will be hungry for more data and inputs to help them be more powerful and accurate. Steps taken today to organize, streamline, and protect enterprise data could pay dividends for years to come, as data debt could one day become the biggest portion of technical debt. Such groundwork should also help enterprises prepare for the litany of regulatory challenges and ethical uncertainties (such as data collection and use limitations, fairness concerns, lack of transparency) that come with shepherding this new, powerful technology into the future.³⁸ The stakes of garbage in, garbage out are only going to grow: It would be much better to opt for genius in, genius squared.³⁹