Writing.

The model is rarely the bottleneck now. Context is. Why every person and company is about to need a knowledge base a machine can read.

Autonomy is not the goal. Useful systems know when to hand the step back to a human.

The first AI transformation is not a chatbot. It is making the work legible.

AI did not come for the work. It changed the unit of work. On building factories instead of features, and what stays human when producing gets cheap.

The first concrete factory for a growing store. Your support queue is leaking money to the same five questions. Build the system, not the chatbot.

You cannot query a company whose knowledge is fragmented and dirty. Before the brain, someone has to drag every scattered fact into the open and clean it with a model.

The people who use AI all day and the people who say it cannot work sound like they live in different worlds. They do. It mostly comes down to which model you pay for and whether you gave it any context.

A quote that should take minutes takes two days, because the knowledge to make it is scattered across a rate sheet, an inbox and one broker’s head. The first in a series: take one industry chokepoint and sketch the system that clears it.

MCP and LLM operating systems could replace fragmented application switching with a dynamic visual map of a company and its work.

For thirty years the interface was the product. The operator is changing from a person clicking a screen to an agent calling an API, and most software is still built for the wrong one.

Most AI images come out generic because people ask the model to invent too much. The fix is almost dull: find a reference first, then build the prompt around what to keep, change and avoid.

Most ideas used to die before an honest test because discovering whether they were good was too expensive. AI makes the first experiment cheaper without making production easy.

Automation can remove a task without removing the need for people. Often it lowers the cost of work, increases demand and moves the constraint somewhere new.

AI is a capability, not a business strategy. Start with repeated, measurable operational pain, then use the simplest mechanism that can solve it safely.

Documentation rots because a person has to write it. Drop your raw material into one folder and let a model compile and maintain the wiki. You keep the pile of truth current; the machine keeps the wiki in sync.

Once a model has compiled your knowledge into a wiki, you can ask it questions a person used to need an afternoon for, file the answers back in, and point it at the whole thing to find where it contradicts itself.

A pile of what you have read sits in your head, leaks, and cannot be queried. Keep one folder of everything you read, let a model compile it into a wiki you can browse and question, and your curiosity finally adds up.

A forwarder's real asset is everything it has learned moving shipments and never wrote down. How I would gather that scattered history into one pile and let a model compile it into a living wiki of lanes, carriers and customers.

A field-service team keeps energy assets running on knowledge written down almost nowhere: asset histories, site quirks, fault patterns, the procedure that keeps a person safe. How I would compile it, built read-only first.

Construction has a knowledge problem with a name: every project relearns what the last one knew, because the lessons die at handover. How I would compile drawings, RFIs and records across projects into memory a firm owns.

Making one task faster does not improve the whole workflow if the constraint moves into data, review, permissions or the next human handoff.

Coding agents are improving faster than almost any other AI tool. The durable skill is not loyalty to one of them. It is a workflow that absorbs better tools without losing context, tests or source control.

Frontier AI makes old software weaknesses easier to find. Important systems need security review against agent-speed analysis.

The AI systems I run live inside Slack and the team talks to them by name. A plain guide to the stack, the Slack setup, the guardrails and the database behind a team agent you can build yourself.

AI lets one capable person cover more ground across functions. The worker who gains the most is the high-context generalist: technical enough to build, commercial enough to choose the right problem, and careful enough to know where a specialist is still required.

WhatsApp is adding usernames. It made me notice that ChatGPT, Claude and Grok are single-player silos, with no second person in the room. The next round may go to whoever already has the social graph and learns to put the model between people.

Coding agents made typing cheap. They did not make knowing what to build cheap. The winning skill is knowing which building blocks a system needs and choosing the right ones.

Bottom-up pilots are how a business finds its real friction. They are not how it decides what to fix. A point solution stays small unless someone with scope across the business chooses to redesign the workflow around it.

Most companies do not have a lack-of-data problem. They have a permission model built for a person browsing a folder, applied without change to an agent that was never in the room.

Australia can build the physical layer of the AI economy, but hosting foreign data centres is not the same as building national AI capability.

Anthropic has not shown that Claude feels anything. It may have shown something stranger and more useful: a machine can develop a private workspace for thoughts it can report, reason with and act from.

AI vendor choices will change. Keep the workflow, knowledge, evals and evidence in a form the company can take elsewhere.

External AI specialists can accelerate delivery, but the company must retain the understanding required to run and improve the system.

As AI lowers the cost of execution, the valuable work shifts towards defining useful jobs, supplying context and judging the result.

Easy access to capable AI can hide the real learning curve. Fluency compounds through repeated use on real work.

If AGI can learn almost any computer-based job, retraining is not a sufficient answer. The problem becomes ownership and distribution.

Agents will not remove browsers, websites or software. They will change who operates them and move the interface towards outcomes.

A capable model can explain the industry and still have no idea which procedure, exception or approval rule applies inside your company.

If AI makes computer-based labour cheap sooner than expected, how much of today's work should become something you own?

Portable intelligence will spread across specialised machines until robots become an ordinary part of the landscape.

Musk says money will not matter in 2036. The argument underneath the headline is about physical patience getting cheap, and what stays expensive after it does.

The first output is a starting point. Serious AI work still requires inspection, testing, judgement and correction.