The State of Applied AI in Mid-2026
We published a literature review on applied AI in mid-2026, surveying ten capability categories, three independent fact-check passes, written for operational leaders and regulated professionals. Here is what it covers and how to use it.
Most of what gets written about AI right now falls into two buckets. There are glossy demos that do not survive contact with real data. And there are confident predictions about how everything is about to change. Neither of them helps an operational leader decide what to do on Monday.
The State of AI in Mid-2026 is our attempt at something more useful for the people who actually have to sign off on risk and budgets. It is a literature review, written in plain language, organised around ten capability categories, and grounded in how these systems behave in production rather than in demonstrations.
Download the full literature review (PDF) →
The review does not try to forecast a distant future. It asks a narrower question: for a typical small or mid-sized organisation in Australia, what can you reliably ship this year, what is still experimental, and where is the marketing ahead of the evidence?
Who it is written for
Three audiences.
- Leaders of small and mid-sized businesses who are being pitched AI transformation and still have payroll, regulators, and service levels to think about.
- Regulated professionals (clinicians, lawyers, accountants, valuers, engineers) who need to understand where AI can safely augment professional judgement and where it cannot.
- Consultants and internal change agents who sit between vendors and operations and need a firmer basis for recommending specific patterns and controls.
The review assumes you are comfortable with workflows, risk registers, and data governance. It does not assume you follow every new model release.
What it actually covers
Ten capability categories that show up repeatedly in real projects. For each one, the paper separates four things:
- What is reliable in production today for typical business and professional environments
- What works in demonstrations but has failure modes that matter on real data
- What is sold as more mature than it is, particularly in regulated or high-stakes domains
- What is quietly further along than most buyers assume, and therefore under-used
The review stays at the level of applied capability rather than model-by-model benchmarking. It is not a leaderboard. It is a map of where you can reasonably place operational bets across the kinds of workflows that small and mid-sized organisations actually run: document automation, knowledge retrieval, drafting, decision support, monitoring, and the slower-moving categories where the marketing is currently running well ahead of the evidence.
Because most of our work is in Australia, the discussion keeps circling back to Australian regulatory expectations and product constraints (AHPRA, ASIC, RACGP, TGA, the Federal Court’s GPN-AI practice note, AUSTRAC Tranche 2, RICS) rather than purely US or EU case law.
Why a literature review, not an opinion piece
The paper draws on more than a hundred cited sources across peer-reviewed work, independent evaluations, regulator publications, and reputable press, with references you can inspect. Claims about capability, safety, and failure modes are anchored in published evidence where possible, and clearly marked as practitioner experience where not. The methodology, the source hierarchy, and the limitations are part of the document, not hidden.
Our view is that AI strategy work should be held to the same standard as any other critical operational decision. You should be able to see the chain from claim back to source. You should be able to disagree with the author using concrete references rather than vibes.
The fact-check standard
The review was drafted using Anthropic’s Claude Fable 5 and then subjected to a three-pass independent fact-check using Claude Opus 4.8.
Across those three passes, 135 individual fact-check findings were logged and resolved. The corrections log is included as an appendix, so readers can see exactly what changed, and why. The pre-fact-check draft is preserved in the archive for transparency.
The unusual thing here is the visibility of the correction trail. Most AI-assisted documents arrive without one. This one arrives with the chain showing.
For a leader, that means the document is not just a snapshot of what the author thought when they wrote it. It is a worked example of how to build AI-assisted documentation with verifiable claims and an auditable correction trail. That standard is increasingly important in regulated contexts.
How to read and use it
The paper is around 13,000 words. It is meant to be returned to, annotated, and reused in strategy and governance work, not read in a single sitting.
Three patterns we see when people actually use it:
- Strategy teams using the ten capability categories as a checklist when reviewing AI project proposals or vendor pitches
- Risk and compliance functions using the “demo versus production” sections to pressure-test vendor claims and shape control design
- Consultants excerpting specific sections (knowledge retrieval, document automation, clinical scribes, legal AI, vertical AI accuracy claims) into client education packs with the original references intact
You do not need to read it linearly. Many readers start with the methodology and the appendices, see what standard is being applied, and then read the capability sections most relevant to their current projects.
How it fits with our other work
The State of AI review sits alongside a separate literature review we published on local PHI masking and de-identification for clinical AI tools, which synthesises two decades of clinical de-identification research into design principles that can be implemented in modern systems, including in ClientJourney.
Together they sit in the Resources section of this site, alongside the technical artefacts behind other things we build.