AI Vision, hosted by Canva 2025 - Cutting Through the Hype to Reveal What’s Next for AI

Table of Contents

• 🪶 Opening with Country and Culture
• 🩺 Early Diagnosis, Real Lives: AI in Healthcare
• 🇦🇺 The 'Fair Go' of Disruption: Australia’s AI Moment 🎯
• ⚖️ Efficiency versus Expertise: The Tightrope of the Decade
• ⚡ The Fourth Industrial Revolution: Infrastructure and Exponential Change
• 🤖 What It Takes to Be AI-First
• 🎨 How AI Is Changing Creativity and Design
• 🌍 Democratization, Voice, and Global Adoption
• 🔧 Practical Playbook: How I Would Start Tomorrow
• 🧭 Governance, Culture, and the Long Game
• 📈 The Business Case: Why Not Adopting Is Risky
• 🔍 Crystallising Quotes and Memes That Matter
• 🧩 Mistakes I’d Avoid (Now That I’ve Heard This)
• 🔮 What’s Next — My Read on the Road Ahead
• 📣 Final Notes I’m Still Thinking About
• ❓ Frequently Asked Questions

🪶 Opening with Country and Culture

I began the evening listening to a powerful cultural framing from Angie Abdilla, a Palawa woman whose meditation on Country did more than open a conference. It set the tone: AI is not merely code and compute; it is a cultural practice that must sit inside ethical systems that predate silicon.

Angie’s piece reminded me that technology sits on land, in story, and in long histories of stewardship. She proposed that if we are serious about building trustworthy AI, we should study systems that have sustained people and place for millennia and try to encode values like reciprocity, relationality, and regeneration into our datasets, models, and guardrails.

"What could AI become if guardrails, standards and rules were governed by a system that has sustained the world's oldest, longest period of peace, well-being and natural resource abundance?"

🩺 Early Diagnosis, Real Lives: AI in Healthcare

Dr Aengus Tran of Harrison.AI shared a story with the kind of human detail that makes the promise of AI tangible. He described Diane, a 69-year-old patient whose chest x-ray was flagged by AI for a subtle sign of stage one lung cancer. What had been missed by human readers was caught the first day Harrison’s system went live, giving Diane "a second chance."

His message was simple: early, accurate diagnosis saves lives, and AI can extend diagnostic capacity where human expertise is scarce. The math he presented was stark. If diagnosis shifts 12 to 18 months earlier, survival numbers improve dramatically.

• Problem: Radiologist shortages and rising error rates as workload increases.
• Approach: Train models on millions of anonymized exams and fine-tune with radiologists in the loop using reinforcement learning.
• Impact: Harrison.AI now processes massive volumes — including a significant percentage of chest x-rays in England and Hong Kong — and has analyzed more than 10 million scans per year.

I took away three operational lessons from Aengus:

1. You do not need a perfect model to start delivering value; you need a model trained for the moments that matter.
2. Clinical validation and clinician oversight are non-negotiable.
3. Once validated, software scales infinitely; that scale is the strategic advantage.

🇦🇺 The 'Fair Go' of Disruption: Australia’s AI Moment 🎯

A frank panel with Sherif Mansour (Atlassian), Pier Luigi Culazzo (Macquarie Group), Marie-Céline Merret Wirström (Made This), and JJ Fiasson (Canva) explored whether Australian traits—pragmatism, scepticism, authenticity—are superpowers or shackles for AI innovation.

I like their framing: scepticism can protect you from hype, but it can also slow curiosity. The sweet spot is shifting from "prove it first" to "experiment and learn fast."

• Authenticity and Trust: Australians value genuineness. That pushes teams to build transparent systems and to prioritize explainability.
• Team Tone and Voice: One size fits none. Businesses need to integrate organisational personality into AI outputs so the tools sound like the teams that use them.
• AI Slop: The phenomenon of generic, unremarkable AI outputs. The cure is craft: skilled practitioners who iterate, add taste, and build pipelines that privilege signal over noise.

Two practical takeaways from the panel:

1. Make experimentation routine. Create small, safe environments where teams can tinker and learn.
2. Design for team-level customization. People want tools that reflect their voice and domain knowledge.

⚖️ Efficiency versus Expertise: The Tightrope of the Decade

Dr Sandra Peter framed the coming years as a decade of disorientation. That sounds dramatic, but she made a useful point: leaders will constantly balance efficiency gains with the urgent need to preserve and grow human expertise.

Her data points were crisp: AI speeds tasks up, but when we offload thinking without intent, cognitive engagement drops. She called the by-product "work slop" and warned that junior talent, in particular, risks atrophy if we replace training opportunities with autopilot workflows.

Her key advice for leaders:

• Create learning cycles that use AI as a tutor, not simply as a replace-and-forget tool.
• Accept you will be wrong often and design organisational rhythms that let you experiment, fail fast, and pivot publicly.
• Build AI fluency into leadership and board conversations—this is now a strategic competency.

⚡ The Fourth Industrial Revolution: Infrastructure and Exponential Change

Craig Scroggie of NEXTDC gave an infrastructure-first view. His core thesis was straightforward: AI is not just an algorithmic revolution. It is an infrastructure revolution.

Craig reintroduced me to two useful metaphors:

• Wang’s law: GPUs and accelerated compute have rewritten the economics of compute in the AI era, much like Moore’s law shaped the prior decades.
• AI factories: Large-scale, GPU-driven facilities where models are trained and served at scale, analogous to the cloud a decade ago.

He underlined the physicality of AI: submarine cables, data centers, edge infrastructure, 5G/6G and satellite constellations (Starlink). These are the levers that determine whether a country can use AI to improve productivity, deliver services, and close digital divides.

Two operational reminders:

1. Policy and regulation matter. Infrastructure investments require stable frameworks and public-private partnerships.
2. Compute is a bottleneck for many use cases. The countries and companies that secure reliable, low-latency compute will unlock enterprise-grade AI faster.

🤖 What It Takes to Be AI-First

I sat in on a pragmatic panel of founders and investors—Antoinette Lattouf, Amina Rosenberg, Jackie Koh, Steve Hind, and Tom Humphrey—where the discussion moved from the conceptual to the executable.

Here are the ideas that stuck with me.

Two-by-two adoption framework

Jackie Koh’s simple matrix is one of the best heuristics I’ve heard recently. She classifies initiatives along two axes: co-pilot vs autopilot and current vs aspirational tasks.

• Co-pilot: AI augments humans in the loop and is quick to deploy.
• Autopilot: AI performs tasks end-to-end with high reliability, which takes more investment.
• Current tasks: Automate what teams already do to free time and reduce friction.
• Aspirational tasks: Let AI do work you could not previously afford to do—this is where new value is created.

The highest leverage projects sit at the intersection of autopilot and aspirational tasks. They take effort, but they unlock step-changes in capability.

Start with domain experts

Many founders described a common failure: building agents without embedding domain expertise. Models trained on generic web data can mimic competence, but they will not deliver production quality without the guidance of in-house specialists.

Jackie was blunt: deploy AI with the people who understand the work. That means getting domain experts to write prompts, to act as human-in-the-loop verifiers, and to curate training data so the agent represents real organisational knowledge.

Raise ambition first

Steve Hind told a story I liked. His team had been tinkering with small feature improvements when a hackathon revealed a bolder path: feed the system everything and let it reason across code, tickets, docs and customer threads. The step change came from ambition, not incremental optimisation.

That resonated with Tom Humphrey’s "clean slate" advice: if you were designing the workflow today, unconstrained by legacy systems, what would it look like? Start there.

Common missteps

• Pretending large language models are out-of-the-box experts.
• Under-investing in data quality, particularly entity and network data.
• Focusing on marginal efficiency wins instead of unlocking new, aspirational work.

🎨 How AI Is Changing Creativity and Design

Stef Corazza from Canva laid out a clear product vision: put AI in the middle of the creative operating system so it amplifies human imagination rather than replaces it.

I found three product lessons particularly actionable.

Make AI part of the workflow, not a separate toy

Tools that sit outside the editor invite friction. Canva’s new design model shifts from "generate and discard" to "generate and edit." An AI-produced image becomes an editable design component that fits the user's existing canvas. That makes a difference: users finish projects instead of endlessly iterating on a generator.

Invest in an evaluation layer

Canva runs continuous benchmarking and A/B testing to discover which models are actually useful. This internal evaluation pipeline is a hidden product muscle that I think every product team should build. It shortens time-to-ship for high-value capabilities.

Shared context, persistent memory, brand awareness

Future AI agents in creative workflows need persistent memory, awareness of brand rules, and cross-document context. That is how you move from a one-off image generation to a consistent system that supports teams and preserves brand integrity at scale.

"We are building AI around human creativity. The humans are at the center, the ideation comes from humans, and AI enables and augments." — Stef Corazza

🌍 Democratization, Voice, and Global Adoption

Oliver Jay from OpenAI brought a global perspective. His point resonated with Craig’s infrastructure message: democratising AI requires both software design and physical infrastructure.

One neat detail he shared was a concrete example from India where study mode was developed after observing how students use reasoning traces to learn. That product tweak—designing the model to teach reasoning instead of handing over answers—was directly inspired by local user research.

Oliver also emphasised voice as a fundamental unlocking modality. Voice combined with vision opens access for billions who are currently under-served by keyboard-first interfaces. Voice plus local language models is the next major inclusion vector.

He added that OpenAI is partnering with governments through an "OpenAI for a country's" program to accelerate local adoption, policy alignment, and upskilling. For national ambition, that program and local compute investments matter equally.

🔧 Practical Playbook: How I Would Start Tomorrow

If I were advising a leader eager to make AI real in their organisation, this is the checklist I'd hand them.

Step 1: Pick a problem, not a tool

Start with an outcome. Is the goal to speed hiring, improve customer onboarding, or detect disease earlier? Define success metrics first.

Step 2: Map tasks to co-pilot/autopilot and current/aspirational

Use Jackie Koh's matrix. Prioritise projects that combine aspiration with automation when you can, and deliver co-pilot implementations for quick wins.

Step 3: Ship and learn quickly

Give teams the tools, raise output expectations, and create space for iteration. The organisations that treat exploration as routine will learn faster.

Step 4: Protect and grow expertise

Build guardrails that keep cognitive work where it matters. Put mentoring and copiloting practices in place so junior people still build judgement while leveraging AI.

Step 5: Invest in data hygiene

Data quality is the silent variable. Spend more time on data enrichment, entity resolution, and domain-specific corpora than on the latest model headline.

Step 6: Make ethics operational

Adopt cultural and technical standards. Be explicit about transparency, provenance, and how humans remain accountable for decisions.

🧭 Governance, Culture, and the Long Game

Long-term success is rarely a product problem alone. It is organisational.

Three governance ideas stood out:

• Experimentation with guardrails: allow teams to tinker but log changes, track hallucination rates, and set escalation patterns.
• Human accountability: AI should augment decisions but not obviate responsibility. Keep human sign-offs for high-risk outcomes.
• Value-aligned objectives: bake cultural values—reciprocity, shared benefit—into procurement choices and how you measure ROI.

📈 The Business Case: Why Not Adopting Is Risky

Several speakers returned to the same economic logic: there is a growing gap between early adopters and laggards. That gap compounds. Falling behind is not a linear cost. It’s multiplicative.

Practical consequences include fewer customers, worse product-market fit, and the risk that competitors embed AI to create new product experiences you can no longer match. The financial argument is compelling, but so are cultural and talent risks: people want to work on modern stacks and in learning organisations.

🔍 Crystallising Quotes and Memes That Matter

Some lines stuck in my head. They are useful memory anchors:

• "Pessimists sound smart. Optimists make money." —a rallying cry for leadership that chooses action.
• "AI is a cultural practice." — a reminder that technical design without cultural embedding fails at scale.
• "Raise ambition first, then give people tools." — an operational nudge for change agents.

🧩 Mistakes I’d Avoid (Now That I’ve Heard This)

Hearing founders confess their missteps made the room more useful than a lecture. Here’s a short list of practical errors I would not repeat.

• Rushing to deploy general LLM outputs as authoritative without domain supervision.
• Treating model selection as a binary choice; instead, build an evaluation loop to test models quickly.
• Optimising narrow tasks instead of rethinking entire processes from a clean slate.
• Underinvesting in data quality and believing a model will "figure it out".

🔮 What’s Next — My Read on the Road Ahead

Expect the next five years to deliver three converging trends:

1. Specialised, trustworthy models

We’ll see more vertical, fine-tuned models that embed domain expertise and provenance. The era of one-size-fits-all generalists will give way to ensembles of specialists.

2. Invisible AI inside workflows

As AI becomes integrated into daily tools (think editable designs, document-aware agents, voice-native interfaces), the novelty wears off and productivity gains show up in consistent outputs.

3. Infrastructure and policy maturity

Countries and companies will invest in compute capacity, edge deployments, and public-private partnerships. Policy will mature to reconcile safety with innovation.

📣 Final Notes I’m Still Thinking About

There’s no single manifesto coming out of this conversation—only a practical ethic. It’s about three things held together: real ambition, human stewardship, and patient infrastructure.

If you ask me what to do tomorrow, I’ll say:

1. Choose an aspirational project and commit to it.
2. Hire (or designate) domain experts to work with engineers and prompt designers.
3. Invest in data and evaluation pipelines before you invest in models.
4. Build a simple governance loop to protect expertise while enabling exploration.

❓ Frequently Asked Questions

What does "AI-first" actually mean for a company?

AI-first means designing processes and products with AI at the centre of the workflow rather than grafting AI onto legacy work. It implies rethinking how teams produce outputs, how data is organised, and how human roles shift toward oversight and high-value creativity.

How do I choose between co-pilot and autopilot approaches?

Start with the task’s risk and repeatability. Use co-pilot for high-risk, creative or judgement-heavy work and when you need fast adoption. Reserve autopilot for stable, high-volume tasks where reliability can be measured and monitored.

What’s the quickest way to show ROI from AI?

Automate high-frequency, low-value tasks first—reporting, summarization, routine coding and ticket triage. Those wins free time and create internal champions. Pair those quick wins with a roadmap to aspirational projects.

How do we prevent "work slop" and avoid degrading expertise?

Mandate apprenticeships where juniors work alongside AI and experienced practitioners. Use AI modes that surface reasoning (study mode, chain-of-thought) to preserve learning, and design performance metrics that reward judgement development, not just outputs.

What investments should a country prioritise for AI readiness?

Focus on compute infrastructure (data centers and GPUs), low-latency networks (fiber, 5G/6G, satellite), and large-scale upskilling programs. Public-private partnerships for industry-specific pilots accelerate adoption while informing policy.

How do creative teams maintain authenticity when using generative AI?

Embed human taste by making human designers the final arbiter: use AI to draft and expand ideas, but route output through branded templates, editorial review and human refinement. Invest in tools that let designers control style, voice and nuance.

Are specialised models better than general-purpose LLMs?

Specialised models outperform general models when high precision, domain knowledge, and regulatory compliance are required. Use general models for exploration and prototyping, then fine-tune or build specialist models for production use-cases.

What’s the single most important organisational habit to build?

Regular, structured experimentation. Make learning a routine with time-boxed pilots, shared learnings, and transparent criteria for scaling or killing projects. That habit keeps you adaptive in a fast-moving landscape.

---