Opinion: Most organisations are still optimising for the last phase of AI. These five shifts explain why yesterday’s playbook is rapidly becoming obsolete and what business leaders should do next, writes Lucio Ribeiro
A few weeks ago, I stood in front of 30 senior business leaders at The Marketing Academy Fellowship and walked them through AiMOT, the AI Moment of Truth: the point in a customer journey where someone meets artificial intelligence before they meet a human, a brand, or a product.
I made the case that AI-driven discovery is the biggest change to customer experience in fifteen years. In the crowd were people running multi-million-dollar businesses, and the idea, and its newness, clearly landed with all of them.
Then it struck me. This concept didn’t exist one year ago. A room of seasoned operators was nodding at something that had gone from non-existent to boardroom-relevant in under twelve months.
That is the real condition of AI right now. The important shifts are either moving too fast to track or sitting too high to see, until they’ve already reshaped the industry around you. By the time a pattern is obvious, it’s already behind you. What follows below is a map. Five structural shifts already in motion in 2026, grounded in what’s happening in the AI industry now.
What I noticed is that each one punishes a behaviour the last three years rewarded. And each one points at the same buried question: whether the way your organisation is built still fits the technology arriving. In twelve months, these will be the conditions everyone is operating inside.
Shift one: the subsidy era is ending
For three years, AI companies priced products to grow, not to profit. OpenAI, Anthropic, and Google absorbed the real cost of compute to acquire users, build habits, and claim market position. That era is closing.
Frontier AI is repricing toward what it actually costs to run. Goldman Sachs projects hyperscaler capital expenditure at over US$1.1 trillion by 2027, with token consumption rising 24 times through 2030 driven by autonomous agent workloads. The median US company still spends US$11.38 per employee per month on AI. The top one percent of fully AI-invested US firms spend around $7,500. That gap will not hold.
The subsidy funded cheap access and a culture of free experimentation. Every team handed a licence to play, every pilot that ran because the compute was effectively free, existed inside a garden where anyone could plant anything and nobody counted the cost of water.
The companies that spent two years teaching their people to experiment freely are about to tell them every query now has a cost. The habit the subsidy era built is the habit the next era can’t afford.
Shift two: frontier models are going enterprise-first
The highest-capability AI systems are no longer being positioned for consumers first. They are going to large enterprises: the organisations with the budget, the compliance frameworks, and the legal appetite to deploy at scale.
For fifteen years, the small fast player could rent the same infrastructure as the giant. Cloud computing made world-class tools available to anyone with a credit card, and that symmetry is what made disruption cheap.
A startup could punch up because it was holding the same weapon. Frontier-first pricing breaks the symmetry. When the most capable AI flows to balance-sheet size first, incumbency becomes an advantage again for the first time since 2010. Frontier AI providers are repricing the disruptor’s toolkit out of the disruptor’s reach.
Shift three: the benchmark has moved
Everyone agrees the public benchmark stopped mattering. Up to 2025, the model got scored on a capability test, but in 2026 the test is more realistic and harder: what does this make possible in our business, and can we prove it?
The benchmark didn’t disappear though; it just moved inside the company. When the test was public, every organisation read the same league table. Now each one has to define its own measure of what AI is worth, and most have no idea how. I see this in client conversations constantly.
The question used to be which model to buy. Now it’s whether the thing we built last quarter moved a number someone in finance cares about, and whether anyone can agree on which number that was.
For most organisations it’s a capability they don’t have and can’t purchase. AI has finished its move from tool to infrastructure, a permanent layer rather than an app you open, and the firms treating it as a project have already fallen behind the ones treating it as plumbing.
You will need to build your own scorecard.
Shift four: from ask to task
The dominant mode of AI interaction is no longer a chat window, as agents increasingly replace prompts. We are moving from ask to task.
Agents are now the operating systems that act across tools, data, and workflows, completing multi-step tasks without a human approving every step.
The chat window succeeded because it asked nothing of your operating model. You bolted it on, a person typed a question, a tool answered, and nothing underneath had to change. Adoption was easy precisely because it was structurally weightless.
Agents remove that comfort.
A system that acts across a workflow forces the decision the chat window let you defer: who is accountable for what the machine does; who owns the output when no human wrote it, and what gets escalated to a person versus automated past one.
These are operating model questions, and most organisations have not answered them because the previous phase rewarded them for never having to ask. The easy adoption that looked like progress demanded no restructuring, which is exactly why the next phase stalls for the companies that mistook the bolt-on for the change.
The real work is unglamorous: redesigning how decisions get made, and what a human is actually for in a process the machine can now largely run.
And then there is the question most boards haven’t reached yet: what happens when an agent gets it wrong, and nobody can prove what it did. In regulated industries like financial services, healthcare, and legal, this is no longer theoretical. Audit trails for agent decisions are becoming a compliance requirement, the same way data privacy did after the General Data Protection Regulation (GDPR) arrived and organisations scrambled to retrofit what they should have built from the start.
The commercial response is already forming. TAigentsphere, an Australian startup backed by Main Sequence and chaired by former Optus CEO Kelly Bayer Rosmarin, raised $4 million in April 2026 to build exactly this infrastructure: a governance layer that gives enterprises visibility into what their agents are doing, what they cost, and where the compliance risks sit.
Shift five: the rocket isn’t always the right vehicle
As frontier tokens get expensive and compute becomes a constrained resource, the most important decision in AI is one most organisations are not yet making: matching the tool to the task. Frontier models are the rockets. Open source and simpler models, well-configured and task-specific, are the bicycles. Not every problem needs a frontier model. Some need a rocket, but most need a bicycle.
In my work building AI applications with clients, the question I return to is whether we’re reaching for frontier capability because the task demands it, because it seems obvious, or because it feels more serious.
A state-of-the-art model answering an internal FAQ is a rocket delivering the mail. A smaller open model, well configured and task-specific, does that job at a fraction of the cost, with faster iteration and lower risk.
The deeper shift is what counts as skill now. For three years the scarce thing was the ability to build with AI at all, but that ability is becoming common. The new scarce thing is restraint: knowing what not to build with the expensive model. And it runs against every incentive in the room.
Engineers want the best tool, vendors want to sell the frontier tier, and the CEO likes saying the company runs on the most advanced AI available.
Every arrow points at the rocket, but restraint is the competence nobody is hiring for yet.
The shift underneath all the AI shifts
These five movements are not independent. They converge on one thing.
Look at what each shift actually punishes. Free experimentation. Renting the same tools as the giants. Chasing the most capable system. Bolting AI on without changing the structure underneath. All of it was sound practice eighteen months ago, and all of it is now a liability. Running the 2023 playbook today works against progress.
What the shifts have in common is that they all land on the operating model and architecture. The way an organisation is built, where money and judgement sit, what a human is actually for. That was designed for an earlier version of this technology, one that was cheap, consumer-facing and self-contained. The version arriving now is expensive at the frontier, enterprise-first and embedded in the work itself.
The organisations moving fastest rebuilt the operating model before the rebuild became obvious. That is a leadership decision, after all, not a technology one.
These patterns are being set now, while they’re still cheap to act on. In twelve months they’ll be the ground everyone stands on, and the only question left is who moved early and who is paying to catch up.
Lucio Ribeiro is a technology and AI leader specialising in the application of artificial intelligence across marketing, media, and business. He is a Forbes Australia contributor, has been recognised by Marketing Today as one of the world’s most influential online marketers, and holds executive certification in artificial intelligence from MIT.
Want to see more Forbes articles on your feed? Tap here to make Forbes Australia a preferred source on Google.
Look back on the week that was with hand-picked articles from Australia and around the world. Sign up to the Forbes Australia newsletter here or become a member here.
