It Started as a Text Box

It is worth remembering how this began, because almost everyone's instinct about what AI is was formed in the first few months and most people never updated it. If your mental picture is still a chat window that writes a passable email, you are working from a photograph taken a long time ago.

November 2022

ChatGPT arrived at the end of 2022 and for a while it was a parlor trick. People typed in requests for poems about their cat in the style of a pirate. They asked it to explain quantum physics to a five year old. They tried to catch it out, and it was easy to catch out.

The serious version of that first year was: a better autocomplete. It helped you start a document you did not want to start. It rephrased a difficult email. It explained an error message. Real value, and every bit of it stopped at the edge of the text box.

Because that is all it was. It could not see a single file you owned. It could not open anything, run anything, check anything, or remember you existed after you closed the tab. You pasted things in, it gave words back, and you did the actual work.

An entire generation of opinions about AI was formed in that period. Including, quite often, the opinions of the people now deciding whether their company should use it.

What actually changed

The interesting story since then is not about models getting smarter, although they did. It is about the thing changing category, three times.

It got eyes

First it could be given your material. Documents, spreadsheets, a photograph of a whiteboard, a recording of a meeting, a codebase. The question shifted from "what do you know" to "what do you make of this." That is a completely different question, and it is the one businesses actually have.

It got hands

Then it could use tools. Search something. Query a database. Call an API. Write a file. Send a request and read what came back. The moment a system can take an action and observe the result, it can correct itself, and self-correction is the whole ballgame. Before that, every mistake was yours to catch.

It got patience

Then it could work for a while. Not one answer to one question, but a goal, held across dozens of steps, over minutes or hours, adjusting as it went. You describe the outcome. You come back to a result.

What that looks like on an ordinary Tuesday

Take software, where the change has gone furthest and fastest.

In 2023, a developer using AI was getting line completions. Useful, marginal, essentially a faster keyboard.

Today a competent engineer describes what they want built, in English, and reviews what comes back. They read every line, they reject plenty, and they own the result completely. But they are directing rather than typing. The teams I know doing this well are not incrementally faster. They ship things they would previously have put on a list marked someday.

You will notice this is not what the research says yet. That is what I would expect. A study takes a year to design, run, and publish, so the good ones out now are measuring the tools of two years ago, in an era when those tools changed shape twice. The people doing the work know before the data does. They always have. That is not a reason to skip measurement, and chapter ten is entirely about measuring your own, but it is a reason not to wait for a paper before looking at your own team.

What is now on the table

Here is the part I find genuinely exciting, and it is not about any of the flashy demos.

A great deal of work in every company exists only because moving information between two systems used to require a person. That was never the job. It was the tax on the job. And for the first time it is possible to take that tax off, not with a six figure integration project, but in an afternoon.

Concretely, these are the things that now work reliably enough to build a business process on:

• Turning mess into structure. An email, a scanned invoice, a voicemail, a photo of a handwritten form. Out comes a record with fields in it.
• Drafting from a template plus lookups. Quotes, reports, summaries, first replies. The shape is known, the content is assembled from what you already store, a person sends it.
• Triage at volume. Which bucket, how urgent, who sees it. Your expert stops spending judgment on the ninety percent that needs none.
• Finding the answer nobody knows where to look for. What did we agree with this customer in 2023. Which contracts have this clause. Where did this number come from.
• Watching things continuously. Every ticket, every review, every change to a vendor's terms. Nobody has time. A system does, daily.
• Explaining the thing to whoever has to use it. Turning a policy into an answer, a schema into a sentence, an error into an instruction.

None of that is science fiction. All of it is available to a company of thirty people, this quarter, at a cost that will surprise you in the good direction.

One of my own

A small one, because it is mine and I can vouch for every part of it.

I produce video from my son's basketball games. The old version took two to three hours of my attention every week: pulling footage, cutting, captioning, exporting, uploading. Not difficult. Just relentless, and it ate an evening every week of the season.

It now takes about fifteen minutes. The work did not get smaller. I do it by checking instead of by doing.

Fifteen minutes is not a number anyone would put in a press release. It is real, it happens every week, and I can show you the machine it runs on. Multiply that shape across the dozen places a company does something relentless, and you have the actual opportunity. Not a moonshot. A tax refund.

The good news has a shape to it

Everything above is why this is worth doing, and I would rather you finish this chapter excited than careful. But look once at what each of those capabilities required.

Turning mess into structure requires access to the mess. Finding an answer nobody can locate requires reach across the places it might be. Watching continuously requires standing access to whatever is watched. Drafting from lookups requires the lookups.

The reach is the value. The reach is also the risk. Not two things to trade off against each other. One thing, seen from two sides.

Which means the decisions that make a system useful and the decisions that make it safe are the same decisions, made at the same moment, by the same person. That is the next chapter, and it is the spine of everything after it.