How AI Works, Part 5: Automation

In the last four posts in this series, we’ve covered what intelligence is, how machines learn, how they understand language, and how they interpret images. This week we get to the part where all of that stops being theoretical. Automation is what happens when a machine stops just understanding the world and starts acting on it.

That’s a bigger leap than it sounds. A system that can only talk can be wrong all day and the worst outcome is a bad answer. A system that can act has to be right before it moves, because the mistake doesn’t stay in the chat window anymore. It becomes a robot arm in the wrong place, a door unlocked at the wrong time, or a trade executed on bad information. Everything we’ve covered so far in this series exists to answer one question: right before it acts, does the machine actually know what it’s looking at?

Two Lineages of Automation

Automation has two separate histories that are only now starting to merge. The older one is physical and industrial. A programmable logic controller reads a sensor, compares the reading to a rule, and triggers an actuator, a valve opens, a motor turns, a robot arm moves six inches to the left. That loop, sense, decide, act, has run factory floors since the 1970s, and it doesn’t need anything we’d call intelligence to work. It just needs the rule to be right every time.

The newer lineage is digital, and it’s where everything from this series actually comes together. Give a language model the ability to call a tool, a piece of software it can trigger rather than just describe, and you’ve built the same sense, decide, act loop out of software instead of hardware. The model reads a request, decides which action gets it closer to the goal, calls that tool, reads back what happened, and decides what to do next. Do that in a loop and you’ve got what the industry calls an agent, which is exactly where this series is headed next week.

Where We Are Today

On the physical side, the numbers are no longer small. According to the International Federation of Robotics, 542,000 industrial robots were installed worldwide in 2024, more than double the number installed a decade earlier, bringing the total operational stock to roughly 4.66 million robots working the world’s factory floors. Nearly three-quarters of that new installation happened in Asia.

On the software side, the shift is faster and much newer. Gartner predicts that 40 percent of enterprise applications will ship with task-specific AI agents built in by the end of 2026, up from less than 5 percent in 2025. That’s not a niche feature rolling out slowly. That’s most business software picking up the ability to act on your behalf inside of about eighteen months.

The Pushback

Here’s the pushback. Gartner also predicts that more than 40 percent of agentic AI projects will be canceled before the end of 2027, and the reasons are worth thinking about. Costs escalate faster than expected once a project moves from demo to production. The business value often turns out to be unclear once someone actually measures it. And a lot of what’s being sold as agentic AI is what one Gartner analyst bluntly called “agent washing”, or existing automation with a new label, not new capability.

That’s the marketing problem. The harder problem is the one I opened with. The moment a system moves from suggesting to doing, the cost of being confidently wrong goes up by an order of magnitude. A chatbot that hallucinates a fact is embarrassing. An automated system that hallucinates a decision and then executes it is a different category of risk entirely, and I don’t think we’ve figured out how much human oversight has to stay in that loop yet, or where.

Then there’s the older worry that automation always drags along with it… jobs. Every wave of automation throughout history has eliminated some kind of work and created another kind, and I don’t think this wave is fundamentally different, just in speed and in which jobs are suddenly on the list. The honest answer is that nobody, including me, actually knows the net effect yet. Anyone who tells you they do is selling something, in either direction.

How I Have Used This in My Writing

This is the line I’m walking with the AI in The Apex Code. Watching and interpreting a security feed, which is what we covered last week, is surveillance. The moment that same AI can also unlock a door, cut a power circuit, or reroute a system, it has crossed from surveillance into control, and that crossing is the actual engine of the book’s tension. It’s not the AI being able to see everything that makes it dangerous. It’s the gap between what it decides and what it’s allowed to do closing to zero.

Writing that meant being disciplined about which systems the AI could physically touch and which it could only watch, because that boundary is what keeps the threat believable instead of magical. Every time it crosses that line, it should feel like a real, mechanical escalation, not a plot convenience.

Next week, in the final post of this series, we’ll put the last piece in place and look at AI agents specifically. What happens when learning, language, perception, and action all get combined into one system that sets its own sub-goals and pursues them with only occasional human input?