Skip to content

AI: What Changed Over the Last 12 Months

What This Is

  • A teaching topic for non-technical professionals.
  • Based on the AI newsletter briefing workstream through 2026-05-18.
  • Intended to help readers understand the direction of travel before learning individual tools or prompts.

Core Lesson

The main change is this:

AI is moving from being a clever chat box to being an operating layer for work.

The early question was: Which model is smartest?

The current question is: Where does AI sit in the workflow, what can it see, what can it do, who reviews it, and how is it governed?

That shift matters because the practical value of AI is no longer just about writing better prompts. It is about designing better loops between people, software, data, and AI agents.

The Short Version

Over the last 12 months, the AI story has moved through six linked shifts.

Shift What changed Why it matters
From models to systems The model still matters, but the surrounding product, tools, memory, permissions, and workflow now matter just as much. Do not judge AI tools only by demo quality or model name.
From prompting to agents AI is increasingly expected to take multi-step action, not just answer a question. The user becomes a reviewer, supervisor, and workflow designer.
From chat to work surfaces AI is moving into browsers, phones, operating systems, email, documents, coding tools, CRMs, and cloud platforms. AI will appear inside the tools people already use.
From experimentation to deployment Companies are asking how to make AI reliable, auditable, secure, and useful inside real processes. Adoption depends on operating change, not just buying licences.
From free-form use to economics Tokens, usage limits, cloud commitments, chips, data centres, and pricing are now part of the AI story. Heavy AI use has real cost and capacity implications.
From abstract risk to controls Security, prompt injection, tool permissions, audit logs, approvals, and data boundaries are now practical deployment issues. Safe use is an everyday operating discipline, not a theoretical debate.

What Happened In Plain English

1. The Model Race Became A Platform Race

The public conversation used to focus heavily on whether one model was better than another. That still matters, but it is no longer enough.

AI companies and large platforms are now competing over distribution and context:

  • Can the AI live in your browser?
  • Can it see your calendar, documents, email, CRM, or codebase?
  • Can it run on your phone?
  • Can it act inside enterprise systems with proper permissions?
  • Can it keep a memory of your preferences and work?
  • Can it be monitored and audited?

For readers, the lesson is simple: the best AI tool is not always the tool with the most impressive demo. It is often the one that fits safely into the work you actually do.

2. Agents Moved From Novelty To Architecture

An agent is not just a chatbot with a new name. In practical terms, an agent is an AI system that can pursue a goal across several steps, often using tools along the way.

That creates new questions:

  • What tools can it use?
  • What data can it access?
  • When does it need approval?
  • What happens if it gets stuck?
  • How do you inspect what it did?
  • How do you stop it doing too much?

Those questions turn a model into something usable at work. The useful parts are the harnesses, review loops, approvals, logs, and operating surfaces around the model.

3. The Hard Part Became Deployment, Not Access

Many organisations can now buy access to strong AI models. That is no longer the main barrier.

The harder questions are:

  • Which workflow should change first?
  • Who owns the process?
  • What data is safe to use?
  • What output quality is acceptable?
  • Who checks the work?
  • How do teams avoid creating unreliable shadow processes?

For a non-technical professional, this means the valuable skill is not just "using AI". It is being able to describe work clearly enough that AI can be inserted responsibly.

4. AI Became Infrastructure

AI now has an infrastructure story behind it:

  • tokens
  • usage limits
  • cloud routing
  • chips
  • data centres
  • pricing models
  • logs and audit trails
  • security controls
  • procurement and vendor risk

That may sound distant from day-to-day office work, but it explains why AI tools can feel inconsistent. A feature may be limited because it is expensive to run, difficult to secure, hard to govern, or tied to a specific cloud or platform strategy.

The practical lesson: if AI becomes part of an important workflow, treat it like operational infrastructure, not like a toy website.

5. Enterprise AI Became More Vertical

The market is moving from general assistants toward packaged workflows for specific domains:

  • finance
  • legal
  • banking
  • insurance
  • sales
  • customer support
  • software development
  • small-business operations

The direction of travel is not just "ask the AI anything". It is "use this governed workflow to complete this kind of work better".

That matters because many people will first meet advanced AI through a specialised workplace tool, not through a blank chat window.

6. Governance Became Practical

AI governance is no longer only about big abstract questions. It is becoming a checklist for normal work:

  • Can this tool use confidential data?
  • Can it take actions or only suggest actions?
  • Are outputs logged?
  • Can prompts leak sensitive context?
  • Can a malicious document manipulate the AI?
  • Who approves high-impact decisions?
  • What happens when the AI is wrong?

The stronger AI gets, the more important these ordinary controls become.

Direction Of Travel

The likely direction is not that everyone becomes a prompt engineer.

The likely direction is that more work will be redesigned around human plus AI loops:

  • AI prepares.
  • Humans decide.
  • AI drafts.
  • Humans review.
  • AI checks.
  • Humans approve.
  • AI follows up.
  • Humans remain accountable.

The most useful people will understand where AI can reduce friction, where it needs supervision, and where it should not be used at all.

What This Means For A Reader

If you are trying to get better at AI, do not start with a giant list of prompts.

Start with these questions:

  • What recurring work do I do every week?
  • Where do I spend time gathering context?
  • Where do I write first drafts?
  • Where do I prepare for meetings or decisions?
  • Where do I chase follow-ups?
  • Where do I compare options?
  • Where do I need a second opinion?
  • Where would mistakes be costly or sensitive?

Then ask whether AI can help as:

  • a thinking partner
  • a drafter
  • a summariser
  • a reviewer
  • a coordinator
  • a research assistant
  • a workflow assistant
  • a controlled agent

Teaching Exercise

Pick one recurring workflow, such as preparing for a weekly meeting.

Write down:

  • inputs: the notes, emails, documents, or systems needed
  • task: what you need to produce
  • judgement: what decisions require human review
  • risk: what information should not be shared
  • output: what a good result looks like
  • follow-up: what should happen afterwards

Then ask AI:

I want to redesign this recurring workflow so AI can help safely. Based on the workflow below, identify which parts AI can draft, summarise, check, or coordinate; which parts need human judgement; what data risks exist; and what a simple first version of the workflow should look like.

Reader Takeaway

The key shift is from prompts to operating models.

Prompts still matter, but the bigger question is how AI fits into real work:

  • What context does it need?
  • What action can it take?
  • What should it never see?
  • What must a human approve?
  • How will the result be checked?

That is the direction of travel.