Most AI Projects Fail Before They Start
The pattern is familiar. A board mandates an AI strategy. A CTO spins up a proof of concept. Six months later, the project is quietly shelved because the data was messier than expected, the integration was harder than scoped, or the business case never quite held together.
Australian organisations are spending real money on AI initiatives that stall at the pilot stage. According to CSIRO research, while adoption intent is high, execution capability consistently lags behind ambition. The gap is not usually about the technology itself - it is about organisational readiness.
An ai readiness checklist is not a bureaucratic exercise. It is a structured way to surface the gaps that will kill your project before you commit significant resources. The 15 questions below are organised into five domains: data, infrastructure, governance, talent, and strategy. Work through them honestly.
Data Foundation: Can You Actually Feed the Machine?
AI systems are only as good as the data that trains and operates them. This sounds obvious, but the specifics catch most organisations off guard.
Question 1: Do you know where your critical data lives?
Not in a general sense - specifically. Which systems hold it, what format it is in, who owns it, and how it is accessed. A retail client we worked with assumed their customer purchase history was clean and centralised. It turned out to be split across three legacy ERPs, two of which were running on-premises with no API access.
Question 2: What is the actual quality of that data?
Run a basic profiling exercise before you commit to anything. Look at completeness rates, duplicate records, inconsistent formatting, and missing values. A dataset that is 40% complete is not a foundation - it is a liability.
Question 3: Do you have the legal right to use this data for AI purposes?
This is where Australian organisations need to be particularly careful. The Privacy Act 1988 and the Australian Privacy Principles govern how personal information can be used. If your intended AI use case involves customer data, check whether your existing consent frameworks actually cover model training and inference. Many do not.
Question 4: Is your data labelled, structured, or annotated in a way that supports your intended use case?
Raw transaction logs are not the same as labelled training data. If you are building a supervised learning model, someone has to do the labelling work. That cost and time needs to be in your project plan.
Infrastructure: Will Your Systems Support What You Are Building?
A technically sound AI model sitting on inadequate infrastructure is a performance problem waiting to happen.
Question 5: Do you have sufficient compute resources, and where will they run?
Cloud, on-premises, or hybrid - each has implications for cost, latency, and data sovereignty. For organisations handling sensitive data, particularly in government, health, or financial services, you may need to consider whether your cloud provider's Australian data centres meet your compliance requirements. AWS, Azure, and Google Cloud all have local regions, but their specific service availability in those regions varies.
Question 6: Can your existing systems integrate with an AI layer?
The most common integration failure mode is an AI model that produces good outputs but cannot push them into the workflow where they are actually needed. If your CRM is a 15-year-old on-premises deployment with no webhooks and a closed API, your AI-generated lead scores are not going to get where they need to go without significant middleware work.
Question 7: How will you handle model versioning, monitoring, and retraining?
This is MLOps territory, and it is where many organisations underinvest. A model that was accurate at deployment will drift as the underlying data distribution changes. You need tooling and processes to detect that drift and respond to it. Tools like MLflow, Weights and Biases, or cloud-native options like AWS SageMaker Model Monitor can help, but they require someone to configure and operate them.
Governance: Who Is Responsible When Things Go Wrong?
AI governance is not just about ethics statements on a website. It is about operational accountability.
Question 8: Do you have a clear owner for AI decisions and outcomes?
Not a committee - a person. Someone who is accountable when the model makes a consequential error. In the absence of a dedicated AI governance role, this typically falls to the CTO or CIO, but it needs to be explicit.
Question 9: Do you have a process for reviewing AI outputs before they affect customers or operations?
Human-in-the-loop design is not always practical at scale, but for high-stakes decisions - credit approvals, medical triage support, hiring recommendations - you need a defined review process. The Australian Human Rights Commission's work on AI and human rights provides a useful framework for thinking about where human oversight is non-negotiable.
Question 10: Are you prepared for the AI Safety and Governance Framework requirements emerging at the federal level?
Australia's Department of Industry, Science and Resources has been developing voluntary AI ethics principles that are increasingly being referenced in procurement and regulatory contexts. Organisations in regulated industries should be tracking this closely, as voluntary frameworks have a habit of becoming mandatory ones.
Talent: Do You Have the People to Build and Run This?
Technology does not implement itself. The talent question is often the most honest constraint in an ai readiness checklist.
Question 11: Do you have the internal capability to evaluate AI vendors and solutions critically?
Buying an AI product is not the same as having an AI capability. If no one on your team can read a model card, interrogate a benchmark, or identify when a vendor is overselling, you are at significant risk of purchasing something that does not do what you need it to do.
Question 12: Do you have a plan for change management and staff upskilling?
AI tools that change workflows require people to change how they work. This is not automatic. A logistics company that deployed an AI-assisted route optimisation tool found that drivers ignored the recommendations because they did not trust them and had not been involved in the rollout. The tool was technically sound. The change management was not.
Question 13: Are you clear on what you will build versus buy versus partner on?
Most Australian organisations should not be training foundation models from scratch. The compute costs are prohibitive and the data requirements are enormous. But fine-tuning, prompt engineering, and retrieval-augmented generation are realistic in-house capabilities. Being clear about this boundary stops you from over-hiring or over-scoping.
Strategic Alignment: Does This Connect to Business Outcomes?
The final domain of any serious ai readiness checklist is the one that gets skipped most often - whether the AI initiative is actually connected to something the business needs.
Question 14: Can you articulate a specific problem this AI project will solve, with a measurable outcome?
Not "improve efficiency" - that is not a success criterion. Something like: reduce average claims processing time from 4 days to 1.5 days, or increase first-call resolution rate from 62% to 75%. If you cannot state it that specifically, the project is not ready to start.
Question 15: Have you mapped the dependencies and risks that sit between your current state and that outcome?
This is where the previous 14 questions come together. If question 2 reveals your data quality is poor, that is a dependency. If question 6 reveals your CRM cannot accept API calls, that is a risk. The point of an ai readiness checklist is not to produce a green-light or red-light verdict - it is to build a clear picture of what needs to happen before you can succeed.
A useful exercise here is to map your answers onto a simple 2x2: high readiness versus low readiness on one axis, high business value versus low business value on the other. Projects in the low readiness, high value quadrant are not no-goes - they are investments with prerequisites. Projects in the low readiness, low value quadrant probably should not happen at all.
What to Do Next
Work through these 15 questions with your leadership team, not just your technical staff. The data questions need input from whoever owns your systems of record. The governance questions need input from legal and compliance. The talent questions need input from HR and your functional business leads.
Once you have honest answers, you will likely find yourself in one of three positions:
- High readiness - You have the foundations in place and can move quickly to scoping and vendor selection or build planning.
- Partial readiness - You have gaps in specific domains that need to be addressed before a full AI project can succeed. Prioritise those gaps as pre-project workstreams.
- Low readiness - The fundamentals are not there yet. This is not a failure - it is useful information. Focus on data infrastructure, governance basics, and building internal capability before committing to a major AI initiative.
If you want to work through this assessment in a structured way, Exponential Tech runs a half-day AI readiness workshop for leadership teams that produces a prioritised gap analysis and a 90-day action plan. It is a practical starting point before you commit budget to a build.
Reach out through exponentialtech.ai to discuss whether that is the right next step for your organisation.
