Why Chatbots Don't Finish Work

We've spent the last few years teaching AI to talk. The next challenge is teaching it to work.

A few months ago, I watched someone use an AI assistant to handle a customer refund.

The interaction was impressive. The assistant understood the customer's request immediately, explained the company's refund policy in clear language, apologized for the inconvenience, and drafted a professional response that was ready to send. If you judged success by the quality of the conversation, it was hard to find anything to criticize.

The customer still didn't receive a refund.

Someone had to open the payment system, locate the original order, verify that it qualified under company policy, review the customer's purchase history, decide whether additional approval was required, issue the refund, update the accounting records, and finally notify the customer that everything had been completed.

The AI had answered the question.

The business still had to finish the work.

The more I think about enterprise AI, the more I believe this distinction explains why so many AI projects feel impressive during a demonstration yet struggle to create meaningful operational value.

We have become very good at conversations

The past three years have been extraordinary for conversational AI.

Every new generation of language models has become better at writing, summarizing, translating, coding, and reasoning. Tasks that once required hours of human effort can now be completed in seconds. It is difficult not to be impressed by how quickly these systems have evolved.

Naturally, most organizations have tried to introduce these capabilities into their daily work. Customer support teams deploy AI assistants. Employees ask AI to summarize meetings. Developers generate code. Finance teams draft reports. Legal departments review contracts.

These are all useful applications.

But they also share a common assumption: that producing a better answer is the same as completing the job.

In reality, answering a question is usually only the first step.

Businesses don't run on conversations

Walk through almost any department inside a company and you'll notice something interesting.

People rarely wake up thinking about conversations.

They think about getting work done.

A customer wants a refund. A supplier sends an invoice. A monitoring system reports an outage. A new employee joins the company. A sales representative closes a deal.

Each of these begins with a request. None of them ends there.

The request quickly turns into a series of decisions involving people, systems, policies, approvals, and exceptions. Information moves between applications. Different specialists contribute different pieces of knowledge. Managers become involved only when judgment is required. Eventually, after enough coordination, the work is complete.

What businesses actually manage isn't information.

They manage work moving through an organization.

The real bottleneck is coordination

For years we assumed the problem was a lack of information.

It isn't.

Most companies already know what they should do. Their policies exist. Their customer records exist. Their ERP systems exist. Their knowledge bases exist. Their employees understand the process.

What slows everything down is coordinating all of those pieces.

A simple employee onboarding process might involve HR, IT, security, payroll, procurement, and the hiring manager. None of those individual tasks is especially difficult. The complexity comes from moving the work from one participant to the next while making sure nothing falls through the cracks.

The same pattern appears everywhere. Invoices. Support cases. Insurance claims. Procurement requests. Incident response.

The work itself is often straightforward.

The coordination is what consumes the time.

This is where chatbots reach their limit

Chatbots are remarkably good at one thing: conversation.

They receive an input, produce an output, and wait for the next request.

That model works perfectly for questions.

Business operations work differently.

A customer asking for a refund doesn't need another conversation. They need money returned to their account.

An employee reporting a security incident doesn't need an explanation of company policy. They need someone to investigate, contain the issue, notify the right people, document the findings, and close the incident.

The measure of success isn't whether the AI produced a convincing response.

The measure of success is whether the business objective was achieved.

That requires something more than conversation. It requires coordination.

Perhaps we've been solving the wrong problem

Much of today's AI industry is focused on making conversations feel more human. Larger context windows. Better reasoning. Lower hallucination rates. Faster responses.

These improvements matter, but they all optimize the same interaction: a conversation between one person and one AI.

Businesses, however, are rarely organized that way.

Real work passes through departments. Specialists contribute different expertise. Systems hold different pieces of information. Some decisions can be automated while others require human judgment.

The challenge isn't simply making AI smarter.

It's enabling AI to participate in the same coordinated work that organizations perform every day.

A different way to think about enterprise AI

Perhaps the next generation of enterprise AI won't be defined by better chat interfaces or larger language models.

Perhaps it will be defined by something much simpler.

Whether the work actually gets finished.

That shifts the question from "Can AI answer this request?" to something far more interesting: "Can AI take responsibility for completing this objective?"

Those are fundamentally different problems. One optimizes conversations. The other optimizes operations.

This second question is beginning to have a name. A growing group of operators and researchers have started to describe the shift as the move toward an autonomous enterprise — organizations where AI doesn't just assist with work but is trusted to carry it through to completion. It's a body of thinking worth following, and the clearest ongoing account of it lives at Enterprise Autonomy.

I suspect the next decade of enterprise software will be built around that second question.

In the next article, I'll explore what that might look like — and why I believe businesses will increasingly rely not on individual AI assistants, but on coordinated teams of AI specialists working together toward a shared outcome.