Sixty eight percent of companies say at least some of their AI initiatives ran over budget in the past year, and a third say it happens most of the time or always. Only nine percent say more than three quarters of their AI initiatives have delivered a measurable financial return. Put those two figures side by side and the easy conclusion is that AI is simply too expensive, or that vendors are overpromising. The more specific and more fixable explanation in the research is that the overruns concentrate in implementation, not in the model itself.
Where the Budget Actually Goes Wrong
AI implementation costs currently range from around 5,000 dollars for a small pilot to 250,000 dollars or more for an enterprise deployment, and that range is mostly explained by how much integration work a given project actually requires, not by which model gets selected. A budget built around model licensing costs and a rough estimate of engineering time rarely accounts for the data readiness work, the legacy system connections and the compliance requirements that only become visible once a project is genuinely underway.
This is where most overruns are quietly generated. Not in a single dramatic cost blowout, but in a series of smaller, unplanned integration tasks that were never scoped because the original budget assumed the model was the expensive part.
Why "The Model Was More Expensive Than Expected" Is Rarely the Real Story
Recent research on this exact question is fairly direct about where the actual challenge sits: integrating AI capabilities into production workflows with the right architecture, testing, security controls and cost observability, not the model itself. A model's price is known upfront and rarely changes mid project. Integration scope, by contrast, is frequently underestimated at the outset and then discovered incrementally, which is exactly the pattern that produces a budget that looked reasonable at the start and did not hold.
The Architecture Decisions That Prevent Overruns Before They Happen
The organisations avoiding this pattern are, in practice, the ones that map integration scope and data readiness before committing to a budget and a timeline, rather than after. That means identifying every system the AI solution needs to connect to, assessing how clean the data actually is rather than how clean it is assumed to be, and building in the compliance and security review as a planned cost rather than a discovery made partway through the project. None of this is complicated in principle. It is simply work that has to happen before the budget is set, not after.
Why Some Organisations Are Already Seeing 55 Percent ROI While Others See None
The gap between the nine percent seeing meaningful returns and everyone else is not primarily explained by which organisations picked a better model. Teams that followed established AI implementation best practices to what researchers describe as an extremely significant extent reported a median generative AI return on investment of 55 percent. That is not a marginal difference. It is the clearest evidence available that the technology itself is capable of returning real value, and that the gap between success and failure sits almost entirely in how the implementation was actually done.
What This Means for Australian Organisations Budgeting AI Work Now
For Australian organisations, the integration scope that tends to get underestimated is often larger than the global average, given the legacy ERP, banking core and clinical systems that AI implementation in Australia routinely has to connect to, and the sovereign data hosting requirements a growing share of the market treats as non-negotiable. Budgeting AI implementation cost in Australia without accounting for that additional scope is one of the more predictable ways to end up inside the 68 percent reporting overruns rather than the smaller group seeing genuine returns.
What This Means for Your Organisation
What we see across implementation engagements is that the organisations avoiding cost overruns are rarely the ones that negotiated the best price on a model licence. They are the ones that did the unglamorous work of mapping integration scope and data readiness honestly before committing to a budget, rather than discovering the real scope of the project midway through it.
Key Takeaways
- Sixty eight percent of enterprises report AI initiatives running over budget, and only nine percent see meaningful financial returns on most of their AI initiatives.
- AI implementation costs range from roughly 5,000 to 250,000 dollars or more, a range mostly explained by integration complexity rather than model choice.
- The real challenge, per recent research, is integrating AI into production workflows with the right architecture, testing, security controls and cost observability, not model pricing.
- Teams following AI implementation best practices to a significant extent reported a median genAI ROI of 55 percent, evidence the technology delivers value when implementation is done properly.
- Australian organisations face a larger than average integration scope given legacy system connections and sovereign data requirements, and should budget accordingly.
How Trusenta Can Help
Custom AI Development scopes integration and data readiness before the budget is set, closing the exact gap this post identifies as the source of most overruns.
AI Integration Services connects AI to legacy ERP, banking core and clinical systems as a planned, costed part of the project rather than a mid project surprise.
Enterprise Architecture maps existing systems and data readiness before development starts, so the true integration scope is known before the budget is committed.
Conclusion
The honest reading of this year's cost and ROI research is not that AI implementation is unpredictably expensive. It is that the organisations treating integration scope and data readiness as a planning exercise, done properly before the budget is set, are the ones seeing real returns, and the organisations skipping that step are the ones supplying the overrun statistics everyone else is now reading about.
