The State of AI in 2026: What Actually Changed, and What It Means for Your Business

Every few months the AI narrative resets. A new model launches, the demos go viral, and everyone is told the ground has shifted again. Most of it is noise. But underneath the hype cycle, a handful of genuine shifts have happened over the past year that change what's worth building — and what's now a waste of money. Here's our read from the work we're actually shipping.

Twelve months ago, 'AI agents' mostly meant impressive videos and brittle prototypes that fell over the moment they hit a real workflow. That's changed. The current generation of models is meaningfully better at multi-step tasks — calling tools, checking their own work, recovering from errors, and knowing when to stop and ask a human.

What this means in practice: tasks we wouldn't have quoted on a year ago are now viable. An agent that reads an incoming email, pulls the relevant records from three systems, drafts a response, and flags anything unusual for review is no longer a research project — it's a four-to-eight week build. The reliability is finally high enough that you can put one in front of real customers, provided you wrap it in the right guardrails.

The businesses winning here aren't the ones with the most ambitious agents. They're the ones who scoped a single, well-defined process and automated it end to end.

The cost of running a capable model has fallen off a cliff. Tasks that cost dollars per request two years ago now cost fractions of a cent. This is the most underrated shift in the entire space, because it quietly changes the economics of everything.

Use cases that didn't make financial sense — processing every support ticket, classifying every document, summarising every meeting — now do. If you ran the numbers on an AI project 18 months ago and it didn't pay off, it's worth running them again. The model is cheaper, faster, and better than the one you costed against.

The newer 'reasoning' models — the ones that work through a problem step by step before answering — are a real step up for tasks where correctness matters. Financial calculations, multi-condition policy checks, code generation, structured data extraction with strict rules: these are noticeably more reliable than they were.

They're slower and cost more per call, so they're not the right tool for everything. The pattern we use most often is a tiered one: a fast, cheap model handles the bulk of the volume, and escalates the genuinely hard cases to a reasoning model. You get accuracy where it matters without paying for it on every request.

Open-weight models you can run on your own infrastructure are now good enough for a large share of business use cases. For Australian businesses with data sovereignty requirements — healthcare, financial services, government — this matters enormously. You no longer have to choose between capability and keeping data inside your own walls.

We're now running production workloads on self-hosted open models for clients who couldn't send data to a US-hosted API even if they wanted to. Eighteen months ago that meant accepting a real capability hit. Today the trade-off is small enough that, for many tasks, it isn't a trade-off at all.

Models still hallucinate. They still produce confident, wrong answers. They still break on edge cases you didn't anticipate. None of the progress above removes the need for the boring engineering that makes AI safe to deploy: validation, confidence thresholds, human review for high-stakes outputs, and clear fallbacks when the system isn't sure.

The other thing that hasn't changed: data readiness is still the bottleneck. The most capable model in the world can't help you if your records are scattered across inconsistent formats with no clean source of truth. The companies getting real value from AI mostly did the unglamorous work of getting their data in order first.

Three practical takeaways. First, re-evaluate projects you shelved on cost — the economics have shifted. Second, agents are now worth considering for well-defined, repetitive processes, but resist the urge to automate something sprawling and ambiguous; start narrow. Third, if data sovereignty was the thing blocking you, the self-hosted option is genuinely viable now.

The fundamentals of a good AI project haven't moved: pick a specific problem with measurable value, build the smallest thing that solves it, put proper guardrails around it, and measure whether it worked before expanding. The technology got better. The discipline required to use it well didn't change at all.