What actually changed in AI between 2025 and 2026 — the boring stuff that's eating real workflows

The loudest AI stories of the past twelve months have been about things that don't matter much to a plumbing business in Geelong or a physio practice in Parramatta. New frontier models. Billion-dollar valuations. Benchmark records. Lawsuits. The usual. Meanwhile, underneath all of that noise, a set of quieter, more structural changes have been steadily making AI genuinely useful for everyday Australian business operations — and most owners haven't heard about them, because they don't make good headlines.

This post is about those structural changes. Not the theatre — the infrastructure. The shifts in what AI costs, where it runs, and what it can reliably do without a machine learning engineer in the room. Some of it is genuinely interesting. Most of it is, frankly, boring. That's the point.

Boring, in this context, means it actually works.

What did AI inference costs actually do between 2025 and 2026?

They fell substantially — and this is the single most consequential change for small business operators. Running a capable language model used to require either expensive API calls that added up fast at any real volume, or enough technical infrastructure that only well-resourced teams could justify it. That constraint has effectively dissolved.

The mechanism is straightforward: as chip manufacturing scaled and model architectures became more efficient, the cost per token processed dropped significantly across the major providers. What that means in practice is that workflows involving thousands of documents per month — supplier invoices, job reports, intake forms, customer emails — are now economically viable to run through AI without building a business case around them. You just do it.

For a trades business processing fifty supplier invoices a week, or a retail wholesaler sorting through product specification sheets from a dozen different suppliers, the per-document cost is now measured in fractions of a cent. The decision about whether to automate a document workflow is no longer primarily a cost question. It's an integration question — which is a more solvable problem.

Has AI capability become commoditised, or are the top models still meaningfully different?

At the top end, the frontier models are still differentiated — but the gap that matters for most business workflows is no longer between the best and second-best model. It's between models capable of doing a task reliably and everything else. And the capable tier now includes a lot of options.

For the practical work that shows up in Australian SME operations — drafting standard correspondence, extracting structured data from messy inputs, summarising meeting notes, classifying incoming requests by type — multiple models now perform well enough that the choice between them is a matter of price, data residency, and integration fit rather than raw capability. Picking the "best" model the way you'd agonise over picking the best tradesperson is the wrong frame. Most of them can hang the door.

The commoditisation is most visible in the extracting-and-classifying category. Pulling fields from a PDF, routing an email to the right department, summarising a phone call transcript — these tasks were genuinely hard to do cheaply and reliably two years ago. They're table stakes now. The interesting question isn't whether AI can do them; it's which business processes are still being done by hand because nobody's connected the pieces yet.

Are there AI models that run in Australia, on Australian infrastructure?

Yes — and this has changed materially. Twelve months ago, "run it locally or send it offshore" was a real binary for operators with data residency obligations. That binary has largely collapsed, though it still takes some deliberate setup to get right.

AWS Sydney — the ap-southeast-2 region — hosts capable model deployments that keep data on Australian soil. Smaller open-weight models can run on modest hardware, including on-premise servers that never touch a public cloud endpoint. For businesses in healthcare, legal, government supply chains, or anywhere the Australian Privacy Principles create a meaningful compliance obligation, these aren't theoretical options anymore — they're practical ones that teams have been deploying quietly throughout 2025.

The catch is that AU-hosted options still require more configuration than sending data to a US API endpoint. You're not just signing up for a SaaS product — you're making infrastructure decisions. But the configuration overhead is now a one-time cost rather than a permanent specialised-engineering dependency. A properly set up AU-hosted workflow runs without ongoing intervention. That's a meaningful shift from where things were.

What does "multimodal" actually mean in practice for a small business workflow?

Multimodal means the model can handle more than text — images, audio, documents with mixed layout, handwritten notes, photographs of physical objects. In 2025 this moved from "impressive demo" to "reliable enough to build workflows around" for a specific set of use cases.

The practical applications that have matured are narrower than the demos suggest. A model reading a photo of a handwritten site measurement and extracting the numbers into a spreadsheet — that works reliably now. A model looking at a product image and generating a structured description for a wholesale catalogue — also works. A model transcribing a voicemail and routing the request to the right person — works. These aren't futuristic capabilities; they're available and cheap enough to run at small-business volumes today.

What multimodal doesn't mean is that you can point a camera at anything and get perfect, structured, legally reliable output. The failure modes are real — poor image quality, ambiguous handwriting, unusual document layouts — and any workflow that relies on multimodal extraction for consequential decisions still needs a human review step. Knowing where the seams are is more valuable than being impressed by what the technology can do when conditions are ideal.

What changed about voice and transcription specifically?

Transcription accuracy improved to the point where it's genuinely useful as a default step in several business workflows, not just a novelty. Meeting notes, voicemail summaries, job site walkthroughs recorded on a phone — the transcription quality for clear Australian-accented English is now reliable enough to act on without heavy editing.

The workflow unlock this creates is underrated. When every phone call or site visit produces an automatic, searchable, structured record, the administrative overhead of running a service business drops meaningfully. A sparkie doing three assessments a day doesn't need to type up notes that evening — a short recorded voice memo, transcribed and summarised automatically, lands in whatever system the business uses before the next job starts. That's not a minor efficiency; for owner-operators who are currently doing that typing at 9pm, it's the kind of change that actually changes how the evening goes.

The caveat is accent variability and industry jargon. Transcription still struggles with heavy regional accents, rapid speech over noisy backgrounds, and highly specialised terminology — the sort of thing a cardiologist dictates, or a manufacturing floor supervisor calls out over equipment noise. For standard business communication in reasonable conditions, though, the current generation is good enough to be the first step in a real workflow rather than a curiosity.

So what's the honest picture for an Australian SME looking at this now?

The honest picture is that the structural prerequisites for useful AI adoption — affordable inference, AU data residency options, reliable extraction and transcription, commoditised capability across common tasks — are now largely in place. That's genuinely new. A year ago, several of these were still conditional or required significant technical lift. Today they're table stakes for anyone willing to do the configuration work.

What hasn't changed is the integration gap. The technology being capable doesn't automatically mean it's wired into your quoting tool, your practice management system, or your inventory software. That gap is where most of the real work sits — not debating which model is best, but deciding which specific workflow to connect first, and building something that runs without babysitting. The AI automation work that produces lasting value looks like that: one workflow, properly connected, running quietly in the background while the business gets on with the actual job.

The businesses that will look back in two years and feel like they got ahead of this aren't the ones that spent time reading every model release announcement. They're the ones that picked one slow, manual, repeatable process — the one that eats an hour a day — and replaced it with something that runs on its own. That's it. That's the whole strategy. The technology is boring enough now that the interesting problem is entirely on the process and integration side, and that's a problem that rewards patience and specificity over excitement. If you want a sense of what that looks like across different industries, the industries overview shows where these patterns tend to surface first.