What is moving in AI right now.
Ten developments worth understanding, in plain terms. For people who want to know what AI is actually doing, beyond the press release.
An AI agent does not just answer a question. It carries out a task across several steps on its own, searching, drafting, filing and triggering actions to reach a goal with little human direction.
It couples a model with tools, memory and a control loop that plans, acts, observes and re-plans. That lets it run multi-step work, and it also lets an early mistake carry through every step that follows.
The bottleneck is the system around the model. Most failures come from authentication, integration and lost context. The model being wrong is rarely the cause. Reliability is a design choice: an agent needs checkpoints, a way to recover from a partial failure, and a clear point to stop and ask. The deployments that hold up take one well-scoped job, give it a named owner and keep a person in the loop.
Get the full brief in your inbox →
Return on investment here means whether the money an organisation spends on AI produces a result it can actually measure, in cost saved or value added.
It is measurable profit-and-loss impact attributable to a deployment. The 2026 evidence shows most spending stalls before it reaches that point, for reasons of integration and workflow fit well ahead of model quality.
The lesson is not to spend less but to spend where the work is. MIT’s NANDA study found around 95 per cent of generative-AI pilots produced no measurable financial return, and that organisations partnering with specialised vendors succeeded about twice as often as those building generic systems in-house. For a community-sector organisation, that usually means starting with one back-office task that has a clear before and after.
Get the full brief in your inbox →
Open models, more precisely open-weight models, are models whose trained parameters are released publicly, so an organisation can download them and run them on infrastructure it controls.
The weights are published for self-hosting and fine-tuning, though the training data usually is not. Leading open-weight models now run within a few per cent of the best closed models on many practical tasks.
Access to capability is no longer the thing holding an organisation back, and neither is the worry about handing data to a third party. The work that remains is choosing the right model for each task and having the capacity to run and govern it. For anyone holding sensitive client information, a model you control is now a serious option, where a year ago it meant accepting clearly weaker results.
Get the full brief in your inbox →
A reasoning model works through a problem step by step before it answers, which improves accuracy on hard questions and takes longer.
It generates intermediate reasoning before its final answer and exposes a setting for how hard to think. Extended reasoning measurably lowers factual error, but it does not remove the model’s tendency to sound just as confident when it is wrong.
When you assess an AI tool, the question is not only how often it is right but what happens when it is wrong and whether you would know. On anything where the facts carry weight, such as a board paper or a funding figure, the safeguard is the checking around the model, because its own confidence tells you very little.
Get the full brief in your inbox →
Model efficiency is the steady fall in the computing power, cost and energy needed to run a capable AI model, which brings serious AI within reach of smaller budgets.
Through distillation, quantisation and mixture-of-experts designs, and through fine-tuning on task-specific data, smaller models now match last year’s frontier on many tasks at a fraction of the cost to run.
Paying for the largest available model on every task is usually waste. A smaller model, chosen well and tuned to the work, often does the job better and far more cheaply. The advantage now goes to the team that can match the model to the task, which is a question of judgement and current knowledge as much as budget.
Get the full brief in your inbox →
This covers the laws and rules that govern how Australian organisations may build, deploy and use AI.
Australia has no standalone AI Act. It regulates AI through existing technology-neutral laws, with a new automated-decision transparency duty under the Privacy Act commencing 10 December 2026 and oversight spread across several sector regulators.
Voluntary at the top does not mean optional on the ground. A board that puts sound AI governance in place now, while the standard is still its own to set, will be ready for the December 2026 duty and defensible in the meantime. The organisations most exposed are those making automated decisions about people without being able to explain them.
Get the full brief in your inbox →
A compound AI system is built from several specialised parts working together, each doing one job well, instead of one model trying to do everything.
It composes multiple models, tools and retrieval steps into a single pipeline whose behaviour exceeds any one model, so the design and the routing matter as much as the underlying model.
When a vendor pitches you AI, ask how the system is designed and where each decision is made, before you ask which model sits inside it. A well-built system of modest models routinely outperforms a single frontier model used alone, and one that cannot account for how it reached an answer is a black box you would be trusting with decisions about people and money.
Get the full brief in your inbox →
A context window is how much text, document and data an AI model can take in and work with at once.
It is measured in tokens. Million-token windows are now standard at the frontier, though accuracy still falls as the key detail sits deeper inside a long input.
A large context lets AI work across a whole body of documents at once, but more input does not produce a better answer on its own. For an organisation that runs on documents, the design question is what you put in front of the model and how you structure it; sheer capacity is the least of it.
Get the full brief in your inbox →
Multimodal AI works across more than one kind of content, text, images, audio and video, instead of text alone.
A single model takes mixed inputs and reasons across them together, so a scanned form, a photo and a paragraph can be read in one pass.
For a sector that runs on scanned forms, photographs of sites and recorded conversations, this widens what AI can take in. The value is in choosing which of those inputs actually improves a decision. Volume for its own sake adds noise.
Get the full brief in your inbox →
AI governance is the set of policies, controls and accountabilities an organisation uses to decide how AI may be used and who answers for it.
It spans the register of where AI is used, the controls around each use, and the supervision that confirms the controls are working, now run as a live operating control day to day.
As AI moves from drafting to acting, the question for a board is no longer whether a policy exists but whether the controls run in practice. Naming where AI is used, who is accountable, what it may and may not do, and how it is supervised is the work, and it is the same map the December 2026 duty will ask you for.
Get the full brief in your inbox →