Next big thing in AI applications

AI coding assistants are undoubtedly one of the most successful and widely adopted applications of GenAI to date.

These tools help software developers write code. Or, more precisely, to outsource code writing. Some developers even claim they haven't written a single line of code in months.

Every major LLM vendor has entered the race: Anthropic with Claude Code, OpenAI with Codex, Google with Gemini CLI, and so on.

The competition is still ongoing, but the market is starting to look crowded. And there are some natural limits to how big it can become.

First, there are only so many software developers in the world. Even if every single one of them starts using a coding assistant, the potential user base still has a ceiling.

Second, not every developer has a directly measurable impact on company revenue or profit. Of course, software engineering creates enormous business value, but linking an individual developer's productivity to financial outcomes isn't always straightforward.

And that brings us to the third limitation: token budgets. Companies will pay for productivity, but how much are they willing to spend on inference per developer? Especially when the financial return isn't always easy to demonstrate.

AI labs seem to have recognized these limitations and are now expanding into a much larger market: knowledge workers.

Why?

Even with a conservative estimate, roughly 15–20 knowledge workers exist for every professional software developer worldwide. That's a substantially larger potential user base.

And here's another interesting difference: some knowledge workers are directly involved in revenue generation, purchasing, investments, budgets and other financial decisions. If an AI agent can help them make better decisions, close more deals or manage larger budgets more effectively, the potential financial return could be substantial.

This means companies might be willing to spend considerably more on inference for certain knowledge-work tasks than they do on coding assistants.

The challenge, however, is that knowledge work is much less clearly defined than software development. And building an agent for such open-ended tasks is considerably harder. Nevertheless, over the last few months, we've seen an impressive number of products attempting to do exactly that.

How do AI coworkers actually work?

The idea is simple: instead of asking an AI assistant to generate content that you then use to complete a task, you give it the task itself.

For example, rather than asking: "Help me prepare a quarterly business review."

You could say:

"Review the quarterly results, compare them with our targets, identify the biggest deviations, and prepare a management presentation."

To do this, the agent needs much more than a powerful LLMs. Looking across today's available products, I would distinguish six important capabilities.

1. Access to information

An AI coworker needs access to the information you normally use to do your job.

This could include local files on your computer, documents stored in SharePoint or Google Drive, email conversations, calendar events, Slack messages, Jira tickets, databases and other systems.

Some agents can directly access files on your computer, subject to permissions. Others work primarily through cloud connectors and enterprise search.

There's a fundamental difference between an AI assistant that reads a document you've uploaded and an AI agent that finds the right document itself.

2. An execution environment

A conventional chatbot generates responses. An AI coworker needs somewhere to execute its work.

Different vendors approach this differently.

Some agents operate on your actual computer, using its applications, browser and file system. Others are given their own virtual computer in the cloud, where they can open a browser, create files, run scripts and use applications independently of your laptop.

The second approach is particularly interesting because the agent can continue working even when your computer is switched off.

3. Access to tools and applications

Understanding a task is one thing. Actually doing it is another. AI coworkers need ways to interact with software.

They can use application APIs, prebuilt connectors and plugins, or interact with user interfaces much like a human: clicking buttons, navigating websites, entering information and downloading files.

API integrations are generally more structured and reliable. Computer use can reach applications that don't offer suitable APIs, but it also introduces additional opportunities for errors.

Overall, the more systems an agent can operate, the broader the range of tasks it can complete.

4. Memory and persistent context

Imagine hiring an assistant who forgets everything about your work after every meeting. That would be rather inconvenient. The same applies to AI agents.

More advanced products maintain persistent context: your preferences, previous tasks, project information, working instructions and sometimes progress towards longer-term goals.

This is what could let an AI coworker move beyond isolated assignments.

5. Identity and permissions

This is where things get particularly interesting from an enterprise architecture perspective.

When an AI agent reads a document, accesses a database or sends an email, whose identity is it using?

There are two fundamentally different approaches.

In the first, the agent operates on behalf of a human user. It inherits access through the user's authenticated applications and connected accounts.

In the second, the agent receives its own identity and permissions. In principle, you can grant it access only to the systems it needs, and later you can attribute its actions to that specific agent.

I believe this distinction will become increasingly important as companies move from personal AI assistants to agents that execute business processes.

6. Autonomy and supervision

Finally, there's the question of how independently an agent can work.

Some assistants execute a task while you watch and ask for approval before taking important actions. Others support long-running tasks, scheduled execution or triggers. An agent might automatically prepare a report every Monday, check for changes in a project or react when a relevant email arrives.

However, autonomy isn't binary. Being able to run for several hours without supervision doesn't mean an agent should independently approve payments, send sensitive information or make consequential business decisions.

The interesting question is how much responsibility we're prepared to delegate, and under what conditions.

Meet your AI coworker

It would be too ambitious to name every AI Coworker released to the market, but it's worth naming the most famous ones.

OpenAI: ChatGPT Work and dots

OpenAI has two related approaches.

ChatGPT Work extends the familiar assistant experience into multi-step execution. Instead of simply generating a response, it can research information, work across files and connected applications, and produce finished documents, spreadsheets, presentations, and other deliverables.

Then, on September 29, OpenAI introduced dots.

Dots take the concept further towards persistent AI workers.

Each dot has its own cloud computer, can connect to thousands of applications through plugins, learns from feedback, and is designed to continue working towards goals over time.

The distinction is interesting: ChatGPT Work focuses on delegating a substantial task, while dots are designed for a longer-term working relationship with an agent.

Anthropic: Claude Cowork

Anthropic has been particularly active in extending its coding-agent experience to general knowledge work.

Claude Cowork brings agentic capabilities to documents, spreadsheets, research and everyday desktop tasks. With appropriate access, Claude can work with files on your computer, interact with applications and websites, and complete multi-step assignments.

For example, you could give it a folder containing interview transcripts, spreadsheets and background materials, then ask it to analyze the information and prepare a report.

The interesting aspect is its ability to operate within an existing working environment, rather than requiring you to copy every relevant piece of information into a chat.

Microsoft: Copilot Cowork

Microsoft has an obvious advantage: millions of knowledge workers already spend much of their day in Outlook, Teams, Word, Excel, PowerPoint and SharePoint.

Copilot Cowork is built to execute multi-step tasks across this environment. It can prepare documents, analyze information, manage calendars, draft and send emails, and carry out recurring workflows, with approval controls for important actions.

It can also keep working in the cloud after the user closes their laptop.

Microsoft's approach is particularly interesting for large organizations because the agent operates within the existing Microsoft 365 ecosystem, using corporate information and established access controls.

There is a catch, however: the pricing. Copilot Cowork requires a Microsoft 365 Copilot licence, but that's not the whole story. Agent execution is charged separately based on consumption through Copilot Credits. In other words, companies pay for both: access to the assistant and the work it performs.

This makes Copilot Cowork an interesting example of the emerging economics of AI agents: moving from fixed per-user subscriptions toward usage-based pricing, where the cost of delegated work must be justified by its business value.

xAI: Grok Bot

Grok Bot, introduced in August'26, is another interesting example of the persistent AI worker approach.

The concept is to create AI bots that have access to a cloud computer, applications and ongoing context, and can carry out tasks on your behalf.

The idea of having several specialized AI assistants is appealing: one for research, another for administrative work, perhaps a third for monitoring particular topics.

But there's an important technical nuance. Giving a bot a name and its own workspace doesn't necessarily provide it with an independent security identity or fully isolated access to systems. That difference matters much more in enterprise environments than in personal productivity scenarios.

Glean: Agents built on enterprise knowledge

Glean approaches the problem differently.

Rather than starting with a general-purpose chatbot, Glean built its business around enterprise search, helping employees find information scattered across SharePoint, Slack, Jira, Confluence and other corporate systems.

But its most interesting capability is the Enterprise Graph.

Glean automatically builds a network of relationships between people, teams, projects, documents and activities. Using metadata, interactions, and AI, it tries to reconstruct how an organization actually works, including who owns specific projects, who has expertise in certain areas, and how different activities connect.

Imagine asking an AI agent:

"Why is Project X delayed, who is responsible for resolving the blockers, and what decisions are still outstanding?"

Instead of simply retrieving documents mentioning Project X, Glean can use its Enterprise Graph to connect relevant information, discussions, and people, helping the agent understand the broader organizational context.

Its agents can then use this context to perform tasks. Glean also supports independent agents with their own identities and permissions.

I find this approach particularly promising because one of the biggest challenges in enterprise AI isn't accessing information, but understanding how it all fits together.

Even the most powerful LLM cannot figure that out from a handful of documents.

So, are we about to get digital employees?

Technically, we're getting considerably closer. But I wouldn't rush to call such AI Agents digital employees just yet.

There are still plenty of open questions around reliability, security, governance and economics. How do we evaluate agents, control their access to sensitive information, assign accountability, and make sure the value they create justifies their cost?

But perhaps the biggest challenge is whether people will delegate meaningful work to AI in the first place.

Unlike software development, where results can often be tested objectively, knowledge work is much harder to evaluate. Checking an agent's output may require almost as much expertise and effort as doing the work yourself. And that could become one of the biggest barriers to adoption.

The Next Competition in AI

For the past few years, AI labs have been competing on model capabilities, benchmark scores and increasingly powerful coding assistants.

Now the competition is shifting towards a much broader challenge: building AI agents that can reliably perform real work.

Success here will depend on much more than the underlying model's intelligence. Agents need access to the right information, an understanding of organizational context, the ability to interact with business systems, and sufficient reliability to operate with limited human supervision.

And, of course, the economics need to make sense. An agent that costs more to operate and supervise than the value it creates is hardly a productivity breakthrough.

I suspect the winners in this market won't necessarily be those with the most powerful LLMs. The real competitive advantage may lie in how well vendors combine AI capabilities with enterprise context, execution, governance and a genuinely convenient way to delegate work.

After all, the ultimate measure of an AI coworker isn't how impressive its answers are, but how much useful work we can confidently hand over to it.

I'm curious how far we've already come in practice.