What are cross-functional AI projects and why do they get messy?
If you have spent any time in the boardrooms of Barangaroo or the innovation hubs of Cremorne lately, https://www.techguide.com.au/news/computers-news/why-australian-tech-professionals-are-going-back-to-study-ai-in-2026/ you’ve heard it: the call to "do more with AI." It sounds simple. You take a product manager, a data engineer, a legal counsel, and a business analyst, shove them into a room, and tell them to ship a feature powered by a Large Language Model (LLM). Six months later, you’re looking at a bloated budget, a stalled rollout, and a "pilot" that nobody actually uses.
Cross-functional AI projects are notorious for being messy because they force a collision between two very different worlds: the rapid, experimental nature of machine learning and the rigid, risk-averse requirements of enterprise governance. When these teams don’t speak the same language, the project hits the wall before a single line of production code is deployed.
The definitions: Familiarity versus expertise
Before we dive into the wreckage, let’s clear up a distinction that haunts almost every project I’ve audited in the last two years. We need to stop conflating AI familiarity with AI expertise.
- AI familiarity: This is the ability to use an AI assistant to draft a Jira ticket or summarise a meeting transcript. It is a productivity multiplier, but it requires zero knowledge of how the model functions.
- AI expertise: This is the ability to understand model bias, hallucination rates, data provenance, and the ethical implications of deployment. It is about understanding the how, not just the output.
Too many projects fail because the stakeholders have high familiarity but zero expertise, leading to unrealistic expectations about what an LLM can actually do in a regulated environment.
Why these projects turn into a ‘mess’
The "mess" usually starts with a failure in stakeholder alignment. In a standard software project, you are building features that follow deterministic logic. In an AI project, the outcome is probabilistic. When the business side expects a "correct" answer and the data team provides a "likely" one, friction is inevitable.
The Governance Gap
You cannot bolt on AI governance workflows at the end of a project. Yet, I see companies try this every week. If you haven’t mapped out how data flows, who owns the model's output, and how you will handle a request for information (RFI) or a privacy breach, you are building on sand. A report from PwC on Australia’s AI readiness highlights this clearly: organisations that treat governance as an afterthought face significantly higher remediation costs when the model eventually trips over a compliance hurdle.
Product Engineering Ethics
You know what's funny? then there is the issue of product engineering ethics. It isn't just about whether the model works; it’s about whether it should be doing the work. If your cross-functional team lacks a shared framework for ethics, you end up with biased outputs that damage the brand. You need more than a list of "thou shalt nots"; you need a technical workflow that detects bias at the inference layer.
The Australian Skills Gap: A Mid-Career Pivot
The Tech Council of Australia has been vocal about our national skills gap. We have a massive cohort of professionals with 5 to 15 years of experience—the mid-career veterans who understand our local market, regulatory landscape, and customer pain points. These people are the ones we need to upskill.
The trend I’m seeing is a shift away from the "hire a new PhD" strategy. Companies are realising that a fresh graduate with deep ML theory but no business acumen is just as dangerous as a business analyst who thinks AI is magic. The solution is the mid-career pivot.
Education: Online vs. Campus
For years, there was a stigma around online postgraduate study. That era is dead. Institutions like The University of Melbourne have bridged this gap effectively. One client recently told me was shocked by the final bill.. Their online postgraduate offerings in data and AI are now viewed by hiring managers as effectively equivalent to their on-campus counterparts. Why? Because the assessments in these programmes are increasingly project-based, forcing students to apply theory to messy, real-world datasets rather than theoretical classroom problems.
Role Required Shift Focus Area Business Analyst From requirements to probabilistic mapping AI Governance Workflows Product Manager From feature lists to data-product lifecycles Product Engineering Ethics Software Engineer From CRUD apps to RAG/LLM orchestration AI Assistant Integration
How to fix the messy middle
If you are leading a cross-functional AI project right now and it feels like it’s going off the rails, you need to pivot your approach. Stop obsessing over the "AI" part of the name and start focusing on the "alignment" part.
- Standardise your definitions: Ensure everyone in the room has a baseline understanding of what a model can and cannot do. If your legal counsel doesn't know the difference between a fine-tuned model and a RAG (Retrieval-Augmented Generation) pipeline, educate them.
- Implement "Governance-as-Code": Don’t rely on spreadsheets to track compliance. Build governance workflows into your CI/CD pipeline. If the model output doesn't pass the toxicity test or the data provenance check, it shouldn’t reach the staging environment.
- Value industry experience: Stop hunting for the "AI unicorn." Look for the senior BA who has been running successful digital transformations for a decade. Pair them with a junior data scientist. The domain knowledge they bring will save you more money than any cutting-edge model configuration.
Final thoughts
The "mess" of cross-functional AI isn't a bug; it’s a feature of the transition from legacy IT to AI-augmented workflows. We are moving away from the era where we could treat technology as a black box that just "works."
Australia has the talent base to lead here. Our workforce is resilient, our regulatory environment is maturing, and our universities are finally producing graduates—and upskilling veterans—who understand the blend of technical precision and business reality required for the job. Stop calling it "AI engineering" when you’re just prompting a chatbot, and start focusing on the rigorous, messy, and rewarding work of building sustainable AI systems. That is how you actually move the needle.
