The honeymoon phase of experimentation is over. We are moving from the "magic" of chat to the "muscle" of execution.
The last two years were about the magic of what AI could say. This year is about the muscle of what AI can actually do, and the gap between organisations that can execute and those that can't is compounding.
The State of AI in 2026 marks a definitive shift in the Australian landscape. We have moved past the honeymoon phase of experimentation and entered a period of industrial execution. The biggest evolution this year is the move from AI as an interface (chatbots) to AI as an operator, agentic systems that perceive, plan, and execute.
That shift creates a widening readiness paradox. Adoption metrics look healthy on paper, but the RBA reports that returns remain "variable" and "piecemeal"3. Adoption has been employee-led, tactical use of tools like ChatGPT, rather than strategy-led structural change.
Winners are moving from isolated pilots to a standardised "AI Factory". The 68% / 1% gap2 is an engineering gap, not an ambition gap, though SME-specific data suggests active deployment is lower, at 37–52%, with the spread driven by how broadly "adoption" is defined.
Agents need Agent Experience, machine-readable data, task-based APIs, structured instructions, or they'll hallucinate in production.
Natural-language prototyping is fast but brittle. Production demands deterministic logic, retraining, and automated guardrails.
83% of Australians demand consent before AI use12. Operationalised guardrails, not ethics PDFs, are how you move fast safely.
Major Australian enterprises are moving past simple chat interfaces toward autonomous "agentic squads" that handle complex business logic. This isn't innovation-lab pocket money.
Don't wait for a perfect enterprise-wide blueprint. The leaders are building muscle by deploying targeted agentic squads on high-value problems today. CBA, Lendi, Deputy, and Suncorp have already moved significant agentic workstreams to "closed won".
Source: Mantel field survey, Sydney Data & AI Day (run with Factor).On the market speedometer, the average needle is stuck in the middle.The ghost needle, the AI Natives, is lapping the field.
of Australian businesses have adopted AI in some form5. Adoption is the easy part.
feel fully prepared to support AI at production scale2. The chasm is here.
plan to adopt within the year1.
of Australian AI companies employ <50 staff8.
AI companies mapped across Australia, up from 544 in 20239.
Defining initial business cases. Investment focused on standard assistants and foundational tooling. Starting the journey, not yet scaling it.
Safely testing specific applications and learning from pilot projects, but failing to scale because they treat AI as software, not a system. This is where the capability debt compounds.
Going all in. Rethinking entire operating models, moving to autonomous systems, accepting calculated risk for substantial long-term reward.
Getting a demo to work (Day 1) is easy. Keeping it compliant, fast, and cost-effective at scale (Day 2) is the real work.
Leaders are now facing the hangover of rapid adoption.
If you cannot maintain guardrails dynamically, you cannot maintain the trust required to operate in a regulated landscape.
Latency creeps up as user load increases. Industrialisation demands inference-stack optimisation so the 10,000th user gets the same speed as the first.
Models decay as customer behaviour changes. Without an automated retraining pipeline, your "intelligence" becomes a legacy liability within months.
A chatty agent or inefficient prompt loop can cost thousands overnight. AI FinOps is now a job, not an afterthought.
Moving leadership from strategy to action.
The winners won't be the ones with the most experiments; they'll be the ones who had the foresight to fund the foundations first.
Waiting for a future all-encompassing model to solve complex integration tasks is a trap. The technology available today is sufficient to transform your business, if engineered correctly.
The goalposts for AGI keep shifting. Waiting for a "god model" is an excuse for inaction. The technology available today is sufficient to transform your business if engineered correctly.
Samuel Irvine Casey · Partner, Data & AI, MantelWe are leaving the era of VC-subsidised tokens. As compute costs plateau, the Token Economy becomes the primary constraint. You must treat Intelligence as a metered utility, not an infinite resource.
SLMs, edge / distilled models. Run at the edge, respond in milliseconds, cost fractions of a cent per call. Route ~80% of traffic here.
Mid-tier models, GPT-4o mini, Claude Haiku class. Balanced ROI. Predictable maintenance. The workhorse of day-to-day enterprise automation.
Premium, o1 / Claude Opus class. High cost, high value. Reserve for high-stakes reasoning where the answer justifies the spend. Used when needed, not by default.
Focusing solely on cost reduction is a dead end. If saved time is consumed by maintenance, compliance, or low-value tasks, the enterprise gains nothing.
Contrary to public fears of mass automation-led job loss, leadership is prioritising efficiency and customer experience. AI is being used to clear backlogs and improve service quality, but most organisations aren't measuring whether staff actually like the tools they're handed.
If your people hate the tool, your "efficiency gains" are imaginary. Track user advocacy as rigorously as technical uptime.
Source: Mantel field survey, Sydney Data & AI Day (run with Factor).Engineering deterministic outcomes today, not waiting for "god models", while foundation models continue to accelerate.
AI is a probabilistic tool today, and businesses need deterministic outcomes. You need to have repeatable, explainable outcomes.
2025–26 marks the end of "AI as an interface" and the beginning of "AI as an operator". Agentic AI doesn't just predict the next word; it executes the next step in a business process by perceiving, planning, and using enterprise tools autonomously.
If generative AI gave computers the ability to create, agentic AI gives them the ability to decide and act.
Samuel Irvine Casey · Partner, Data & AI, MantelInvisible helpers woven into the UI of existing tools (SaaS, ERP, CRM). They don't switch apps; they inhabit them.
An agent inside your CRM that automatically prioritises leads and drafts personalised follow-ups based on real-time LinkedIn signals.21
Conversational partners where the user provides intent and the agent manages execution. These require deterministic planning, the agent must show its work to bridge the trust gap.
Lendi's agentic loan assistant that doesn't just "talk" about loans, it gathers docs, performs credit checks, and prepares the application.21
Fully autonomous "digital employees" that monitor event triggers and act without human intervention. The peak of systemic intelligence, and the most expensive to maintain.
A Deputy-powered workforce agent that monitors weather alerts and proactively suggests (or executes) shift changes to prevent safety violations or understaffing.21
I would much rather use deterministic software than look to solve every problem with an agent-sized hammer.
Emma Bromet · Partner, Data & AI, MantelBecause agents lack human intuition, they require structured data, task-based APIs, and explicit instructions, the machine-readable equivalent of UX. We've spent decades designing for human eyes. Now we must design for machine reasoning.
Agents are literal; they lack intuition. If an agent has to spend 40% of its reasoning power deciphering unstructured data, it will hallucinate.
Traditional APIs are designed for developers who understand the schema. Agents struggle to chain generic CRUD updates without breaking business logic.
Vague instructions lead to hallucinations and process failure. "Use good judgment" means nothing to an agent.
Think of it like a Subway sandwich shop: you need your ingredients prepped, sliced, and organised in the display before the lunch rush.
Thomas Maas · Head of Client Solutions, Strategy & Consulting, MantelMoving from pilot purgatory requires a platform-engineering mindset, building reusable patterns and infrastructure that serve every future agent.
Define use cases & route requests
Curate data products with contracts
Compose agents on shared patterns
Eval gates, guardrails, drift checks
Deploy to API layer with observability
The "Invisible" Foundations Matter Most. Standardised APIs lack a "shiny interface" for Board demos, but they are the accelerant for every future project.
"Live" is a technical milestone; "embedded" is a cultural one. Stop counting launches and start counting workflows that have been fundamentally redesigned.
Source: Mantel field survey, Sydney Data & AI Day (run with Factor).From AI as a chatbot to AI as an operator.
Workflows in businesses are actually shaped by the SaaS software that they have, but AI doesn't need any of those screens. This is not a technology that you put in human organisations. You need to create something different.
We can push code to production in months. Embedding AI into the daily habits of a workforce takes a different kind of muscle. In 2026, 74% of Australian workers worry about accidentally misusing AI42, and that's not solved with a handbook.
Technology is rarely the ultimate point of failure in an AI transformation. Success in the Agentic Era requires shifting focus from the software itself to the mindset of the people using it.
Cath Jordan · Principal Data & AI Consultant, MantelClosing the AI literacy gap means meeting each layer of the organisation where it actually sits, executives, managers, and practitioners need different muscle.
Leaders must view AI as critical infrastructure, not a plugin.
48.5% of AI ambitions are now CEO-led43. AI has transitioned from technical project to strategic mandate. The C-suite provides the mandate (the why).
Middle managers are the bottleneck of fear.
They need technical awareness to supervise an agentic workforce, knowing when to trust the model and when to hit the kill switch. Not experts; stewards.
The threat isn't AI taking your job. It's AI taking your brain.
If staff lose agency, they stop giving feedback. Models decay. Culture must reward the humans who catch hallucinations, not punish them.
Success lies in the overlap of three mandates: the C-suite provides the why, the tech teams provide the what, and the enablement team provides the how.
Source: Mantel field survey, Sydney Data & AI Day (run with Factor).Agents need nurturing: the cross-functional squad.
You wouldn't hire a staff member for three months and then never speak to them again. If you launch-and-leave an agent, you are creating a legacy liability. Product models require permanent teams.
As Australian regulation evolves, this role updates the agent's deterministic guardrails to ensure continued compliance.
Ensures time saved by the agent is actually reinvested into high-value work, not leaking into administrative waste.
Monitors model health and triggers retraining loops the moment accuracy begins to dip below enterprise threshold.
Ensures APIs remain functional even as legacy systems of record are updated or swapped out.
The next 12 months will see a pivot toward AI culture, not just AI code. One-off webinars are not enough; literacy must be treated as a core professional skill.
Source: Mantel field survey, Sydney Data & AI Day (run with Factor).Organisational antibodies, fear, inertia, legacy pride, will instinctively attack and reject AI initiatives. This isn't bad behaviour; it's a survival mechanism from a workforce that has seen innovation theatre come and go.
If staff fear for their jobs, they stop giving the feedback required to fix data drift. The model decays.
Beat it by making compliance the architects of the guardrails, not the victims of them.
Practitioners who saw vibe-coding fail will justify total rejection from a single glitch. Counter with rigorous engineering and shared wins.
Culture, workforce evolution, and overcoming organisational antibodies.
Humans will move to higher-order activities: taste, timing, task knowledge, all these things that are going to have to be imparted into these systems.
In Australia, "she'll be right" is not a security strategy. Build the safety cage, not because you expect to crash, but because it's the only way you can legally and ethically go 300km/h.
If an agent tries to exceed its authority, approving a $1M invoice on a $10k limit, the security layer is a deterministic circuit breaker that blocks the action instantly.
Let users see exactly why an agent made a decision. You close the trust gap and accelerate enterprise adoption.
Data drift can turn a secure agent into a liability overnight. Industrialised security demands continuous monitoring of reasoning paths to detect logic drift.
When you give autonomous agents "write" access to your core systems, you are effectively granting them "Power of Attorney."
Trust is no longer a feeling; it is your Technical Architecture.
Trust as a strategic asset.
In the agentic era, trust becomes a strategic asset, not just a compliance requirement.
40-page Responsible AI PDFs sitting on SharePoint don't stop biased models. Hard-code your values using automated guardrails that check for bias and privacy leaks before an agent acts.
Against a backdrop of global tension, the concern isn't AI failure: it's unpredictable success. Clients are prioritising guardrail engineering to prevent rogue AI in volatile times. Fear of falling behind is overcoming the technology fear, shifting the Australian market from caution to calculated risk tolerance. Australia's regulatory posture has shifted significantly: the National AI Plan (December 2025), a new Australian AI Safety Institute, mandatory high-risk AI guardrails in active consultation, and the ACCC pushing for standing oversight powers mean that "we'll deal with compliance later" is no longer viable. The organisations building operationalised governance now will be ahead when mandatory obligations land, not scrambling to retrofit them.
From checklist compliance to operationalised responsibility.
We have guardrails that we have to stay within, and I think most days that's a good thing, because we have to be buttoned up. We have to do the right things in the right way.
Do not run 20 pilots. Build one thing that works, scales, and pays for the next one. Here's the 100-day plan.
Score Data, Cloud, and People on a 0–5 scale. Output: a ranked list of the handbrakes that have been quietly killing every previous pilot.
Pick one high-value, high-feasibility project and ship the agent, data products, guardrails, and monitoring. Output: a complete reference implementation the next ten use cases inherit from.
Commit budget to API layers, data governance, observability, the AI factory itself. Output: shared infrastructure every future agent will run on.
Moving from strategy to execution: the leadership mandate.
I would treat that piece of paper, an AI strategy, like ice cream in the sun. It's gonna melt fast. I'd be asking: is this like SaaS in disguise? I'd be looking at everything through a sceptical lens.
Unlike previous years that relied primarily on survey data, the 2025–26 edition is grounded in Mantel's direct experience bringing agentic systems to production, validated by the market at the Sydney Data & AI Day.
Direct insights from deploying agentic solutions across banking, media, retail, and government.
Input from Mantel practitioners across AI, Cloud, Digital, and Cyber Security.
Analysis of Australian government, industry, and research data: CSIRO, RBA, ACCC, Department of Industry.
Proprietary survey of leaders at the Sydney Data & AI Day, a raw look at the transformation gap between executive ambition and execution. Run in partnership with Factor.