色控传媒

Like many powerful new technologies in their early stages, generative artificial intelligence (GenAI) is just beginning to show its potential in the workplace with early applications like chatbots, agents, medical research breakthroughs, and software coding assistants. But the real value is yet to come—and humans will play a critical, irreplaceable role.?

I’ll make a bet with you: Over the next two years, we’ll see a significant uptick in real AI production deployments for high-value work across many industries and businesses.

If the history of technology is any guide, the first move by big businesses will be to use AI to improve office productivity, business-related insights, and workflows. This will help orchestrate human teams, optimizing the value-added human activity that computers can’t do. Compared to chatbots or document summarization, this is a far bigger deal in terms of financial impact, efficiency, growth, and management.

Here’s why (and how) I think that will happen.

It may seem odd that GenAI is still looking for its best use case in business. It’s drawn so much attention in the past two years that you would think it's obvious by now what it is for. On one side are ballyhoos and near-apocalyptic predictions of robot overlords, while on the other, pragmatists sneer that business spending on AI doesn’t come close to matching the investments hyperscalers have made in AI infrastructure.

My advice: Ignore this and focus instead on the ways big enterprise technologies typically come into their own.?

• Client-server systems were around for years before most offices had personal computers and email became popular. A new management concept, business process reengineering, which relied on faster information flows, put client-server over the top.
• Smartphones were around for several years before a combination of the iPhone and the development of the App Economy made “bring your own device” a corporate standard for the past 20 years.?
• Zoom and other affordable videoconferencing services were around for years but gained traction when COVID forced innovations to make the home office just as capable as the work office. There are entire businesses built as fully remote offices which would not have been possible without these new scalable technologies.

There is a pattern here. New computing technologies emerge, costs come down, customers grow used to them as they fit into existing workflows, and enabling events accelerate them into workplace essentials.

If you look at GenAI this way, you get a better sense of where it’s headed.

GenAI comes out of machine learning and deep learning, which for several years, and adds several innovations, primarily around data—in particular, large language models. Ideas around attention and brute force computing are leading to breakthroughs.

Now, innovators—from startups to established companies, including 色控传媒—are making it even easier to deploy and easier to use. You can see the pattern of market acceptance followed by broad adoption coming around AI, just as it did with client-server all those years ago.

DeepSeek is a particularly interesting development that came out of China in open source form, free to use by all, with none of the usual secret bits held back. , which is both relatively easy and affordable to train, allows enterprises to deploy their own dedicated fine-tuned models based on their own data, at an affordable price. The echoes of the DeepSeek coming-out party are still reverberating as other models improve, with many used for fine-tuning and optimizing business-specific knowledge and questions.

The data hoarding of large language models (LLMs) is yielding to a new discipline: the management of data. Data management is the art of tracking and organizing data into refined and indexed, filtered, and represented form. The data flywheel, as the practitioners call it, allows data to be scaled up in different contexts to facilitate collaborative systems where many models work in concert.?

The business benefits will be insights, efficiencies, and new discoveries and innovations in even the most entrenched areas of business. Some vendors will approach this as agents, others as domain experts. As usual in our industry, a good approach to technology will take on many names as marketeers try to corner a technology domain and make it unique to themselves.

There is also talk about using more powerful AIs to filter data for use by less complex AIs (e.g., a lower parameter model). This process of distillation shows a lot of promise and may become a strong factor in fine-tuning AI for targeted enterprise deployments. Open source will be disproportionately favored for distillation as the big AI service providers enforce licencing terms that prohibit users from using their system to train more specialized AI.

What do these new data-centric workflows look like? Think about the processes by which drugs are discovered and tested or how corporate financial systems are put through their quarterly and annual reconciliation chores. Quality data is obtained and used in bias-defeating systems like double-blind trials for drug companies or audited records for financial analysis. Analysis follows well-defined workflows, judging both results and specific risks with what-if hypotheses, prior to actions like releasing a drug or making a strategic investment.

Those are the big outlines. GenAI may work within corporate workflows like team collaboration by mimicking several of these processes, including better data identifiers and refining, to support the creation of more dependable results. The background processes, the math in creating data reliability, won’t change much from one team to another, but the metadata or labels will be specific to teams, tasks, and industries. This may not be super obvious but consider that how people collaborate can differ by industry and each organisation.

A new discipline, the data engineer, is already on the rise. That person or team will organize, prepare, and transform data to support this focus on .

This framework for using AI inside processes brings a number of virtues to collaboration. It can strengthen team investigation and growth and create . It can be organized to execute on opportunities and also work out risk profiles of various actions inherent in these opportunities.?

Additionally, we’re seeing AIs increasingly able to by which they arrive at answers. Building an awareness and commentary on its own actions has many benefits, including better-defined data sets, rules on what data can change and how, a means to independent verification (frequently through another model or a reference process), and a process of validation, including whether results spun up as expected. All of these improve system integrity, ensure compliance, and de-risk repeated processes.

Lastly, technologists are now increasing the capability of AI by the development of world models, or simulations of complex interactions within a given context. Think of these as akin to the mental models humans have that enable them to cross the street: They know how sidewalks function, how fast traffic is going, where crosswalks are, what a crossing light is, etc.?

Fei-Fei Li, one of the key players in the development of modern AI, is now building out world models. Think of world models as a sequence of states and the rules that must be adhered to as the AI traverses across each chain of these states. Instead of mimicking what it has seen in the data, AI forms a more coherent understanding of how the states interact with each other and the rules that govern that interaction, whether those rules are codified in physics or in rules.?

This will enable better decision-making, faster error correction, and the utility of AI in more diverse and complex tasks. We have yet to see this in the very large models, but huge bodies of research work are underway to develop these systems of knowledge.

As we build up these remarkable capabilities, what is left for humans to do? So much. Naming the problems to solve is both hard and requires a high degree of imagination (unless you are dealing with the blatantly obvious, in which case you should already have automated the well-known problem).?

The focus of the human will be on the imaginative, future-looking work that needs prioritization. So far, AI shows no capability there equal to what a skilled group of people can do in everyday work. Here’s why.

Collaborative teams are specialists in subdomains that are attuned and managed in concert with other teams and changing market needs. These teams can name and articulate problems—problems of risk, corporate culture, and judgment—and identify solutions consistent with the enterprise’s mission, history, goals, and constraints. They choose the data sources and determine how the AI will work to uncover patterns and evidence to successfully (and compliantly) solve the big-picture problems. These aren’t the sorts of things that machines are good at.?

The reason is the signals are nonstandard. The data is of varying quality. The risk/reward balances are often affected by quasi-emotional, but nonetheless critical, dimensions like reputation, brand value and reliability, and leadership charisma. Trust is analytical, but trust is also sometimes very emotional. So are culture and creativity, the heart of customer and partner relations, and ultimately of enterprise success. These are, for the foreseeable future, sovereign human zones.

Identify workflows in your organisation and interject AI within that workflow to get alternate answers that can be found within data. Seek a parallel path to find answers within workflows that are in the early stages of maturity. I think that, coupled with a good data set management strategy, will enable you to find that AI rapidly will enhance your decision-making.?

The hardest part is getting started—but, like AI, the more you do, the better at it you’ll get.