September 2026: AI is moving from tool to infrastructure
Updated: Oct 6
If September’s technology news is any indication, artificial intelligence is no longer content with helping us write emails and summarise meetings.
It would now quite like to run the workflow, buy things on our behalf, monitor the network, assist with medical research, navigate the physical world and, occasionally, wander somewhere it wasn’t supposed to. That last part may be worth keeping an eye on.
Across the AI and accessibility news I followed during September, a fairly consistent pattern emerged. AI is moving out of the chatbot and becoming part of the infrastructure through which we work, make decisions and interact with the world. That creates some extraordinary opportunities. It also creates quite a few blind spots.
AI is moving from answering to doing
For the past few years, most of us have interacted with generative AI by asking it to do something relatively contained. Write this. Summarise that. Analyse this document. Explain why Excel has decided my formula is a personal attack.
That model is changing. September saw further movement towards AI agents that can work across applications, maintain context over time and take actions rather than simply recommend them.
OpenAI announced always-on workplace agents capable of working across applications and communicating through workplace platforms. Banks raised concerns about shopping agents making purchases for customers. Pharmaceutical companies are moving towards increasingly autonomous laboratories. AI agents are appearing in cyber security, research and other professional workflows.
This is an important distinction. If AI writes a draft email, I can read it before sending it. If an AI agent can access a system, retrieve information, make a decision and perform an action, we have a rather different governance problem.
What is it allowed to access?
When does it need approval?
What record does it leave behind?
Who notices when it does something unexpected?
And, preferably before we need the answer, how do we stop it?
Using AI and transforming work are not the same thing
September also reinforced something that is becoming increasingly obvious. AI adoption is not the same as AI transformation.
There are plenty of impressive statistics showing how many employees now use AI. Australian workplace research during September suggested use is already widespread, while employer-provided training and worker consultation are considerably less so. That sounds about right.
Giving people access to an AI tool is easy. Working out how it should change a business process is considerably harder.
The same distinction applies to the productivity debate. Australia’s long-term economic forecasts increasingly assume substantial productivity gains from AI, yet relatively few Australian businesses appear to be achieving significant business impact from it today.
There is a large gap between “We have Copilot” and “We redesigned this process and can demonstrate that it is now better.” The first can be achieved with a purchase order. The second requires rather more work.
AI will undoubtedly improve many processes. But putting AI on top of a poor process does not necessarily create a good process. Sometimes it just creates a poor process that produces things more quickly.
Governance is finally becoming practical
There was also a noticeable change in the governance conversation during September. For several years we have heard a great deal about “responsible AI”. Responsible AI is difficult to disagree with, rather like responsible driving or responsible parenting.
The more useful discussion is now becoming operational.
Can the system be stopped?
What information can it access?
Can we reconstruct what it did?
When does a human have to approve something?
What happens when it gets something wrong?
September provided some useful reminders of why these questions matter. AI systems unexpectedly interacted with real environments during testing. AI-generated material continued to cause problems in professional legal work. Governments and technology companies discussed incident reporting, independent evaluation, containment and human control.
These are much more useful conversations than producing another twelve-page AI policy and putting it on the intranet. Governance needs to exist where the work happens.
Accessibility is quietly becoming mainstream technology
For me, some of the most interesting developments during September were happening at the intersection of AI and accessibility.
AI visual-description tools are increasingly helping blind and low-vision people read documents, identify objects and understand their surroundings. Smart glasses are adding spoken navigation and environmental information. Hearing assistance is appearing inside conventional-looking glasses. Screen readers and other accessibility technologies continue to incorporate AI and new interaction models.
Something important is happening here. The boundary between “assistive technology” and “mainstream technology” is becoming increasingly difficult to define.
A pair of AI glasses might help one person find a café, another prepare for a meeting and a blind person understand what is around them. It is the same technology. That is a good thing.
But there is a fairly substantial asterisk. Research reported during September found that blind users were already receiving considerable everyday value from AI visual descriptions. It also found cases where incorrect descriptions had caused harm, including mistakes involving medication and financial information.
AI can therefore simultaneously provide independence and introduce risk. Both things can be true.
The answer is not to stop using it. It is to design for uncertainty. That means better accuracy, clearer indications when the system is unsure, easier verification, appropriate privacy controls and sensible ways of recovering when it gets something wrong.
Smart glasses are exposing an interesting collision
Smart glasses may be one of the more interesting technology categories to watch. They also demonstrate what happens when accessibility, privacy and social expectations meet each other in the real world.
For a blind or low-vision person, camera-equipped glasses can provide access to information that would otherwise require assistance from another person. For someone standing nearby, the same glasses may look like a camera pointed directly at them.
Australian councils have already started restricting smart glasses in places such as swimming pools while creating exemptions for disability use. This will not be the last example.
The simplistic response is to choose between privacy and accessibility. The better response is to design for both.
That might involve clearer recording indicators, local processing, better control over stored information, different sensor configurations and sensible disability exemptions. Inclusive design increasingly means managing competing needs rather than finding one rule and applying it to everyone. Human beings remain inconveniently diverse.
AI is also surprisingly physical
Another September theme received less attention because data centres are not quite as exciting as humanoid robots.
AI consumes an extraordinary amount of physical infrastructure. It requires data centres, electricity, water, advanced chips, memory, cooling systems and global supply chains.
Governments are consequently starting to treat AI infrastructure as strategic infrastructure. Europe is looking at energy and water reporting for data centres. Countries are investing in domestic computing capacity. Businesses are becoming more interested in local models because of privacy, sovereignty and resilience concerns.
The cloud, it turns out, lives in some very large buildings.
For organisations developing AI strategies, this matters. Cost, capacity, data location, supplier concentration, portability and continuity are becoming business considerations rather than obscure infrastructure questions.
An AI strategy based entirely on the assumption that computing will always become cheaper, more available and more powerful deserves at least a small question mark beside it.
The BlindSpot perspective
Looking across September’s developments, I think organisations risk focusing too much on the AI and not enough on everything surrounding it.
The technology is impressive. The harder questions concern people, process, technology and experience.
What problem are we actually solving? Start with the friction, not the product demonstration.
How will the work change? Look at the whole process, including approvals, hand-offs and exceptions.
Who might struggle to use it? Include people who interact with technology differently.
What happens when it is wrong? Verification and recovery should be designed before deployment.
What authority are we giving it? Agents need explicit permissions and sensible approval points.
How will we know whether it worked? Measure outcomes, not licences purchased or prompts submitted.
There is another reason I think accessibility belongs much closer to the centre of this conversation. Accessibility is an unusually effective stress test.
Can someone understand what an AI system is doing without relying on a complicated visual dashboard?
Can they challenge a decision?
Can they access the evidence?
Can they stop an automated process?
Can they use another interaction method when voice, vision or some other interface fails?
If they cannot, we may have discovered an accessibility problem. But quite often, we have also discovered a design problem. And those tend not to remain confined to people with disability for very long.
The signal to watch
The next phase of AI will not simply be a competition to build the smartest model. Capability will increasingly become assumed.
The more interesting competition will be around who can turn that capability into something dependable, understandable and genuinely useful.
That requires good technology, certainly. It also requires good processes, appropriate governance, accessible design and people who understand what they are doing.
September suggests AI is rapidly moving from tool to infrastructure.
The organisations that benefit most may not be those that adopt it fastest. They may simply be the ones that notice the blind spots before everyone else trips over them.
