
Roughly three quarters of organisations with AI in use say it is delivering measurable financial impact. Only 13 per cent have scaled it the way the original business case said they would. That gap is where scaling AI and enterprise architecture meet, and most organisations are still treating it as a project management problem.
The numbers come from a BearingPoint study released on 1 October 2026, which surveyed 1,050 C-suite executives and senior leaders in August. TechTarget covered it on 7 October. It is a consultancy survey, so read it as a direction of travel and not a precise measurement. The direction, though, matches what we see.
Why do AI pilots fail to scale?
Because the pilot never had to prove the hard part.
A pilot is small by design. It runs on hand-picked data, inside one team that already knows the workarounds, against a single workflow. Nobody asks whether the identity system can handle two thousand users or whether the core platform can take a write-back at volume. Those questions only arrive when the pilot succeeds, which is exactly the wrong moment.
Picture a claims triage pilot with twenty experienced handlers. It works well. Then the business tries to roll it out to four hundred people across three states, and the AI suddenly needs to read policy data from a system that only updates overnight, push decisions back into a platform that was never built for it, and respect permissions that differ by region. Nothing about the model changed. Everything around it did.
Ipek Ozkaya, a technical director at Carnegie Mellon University's Software Engineering Institute, put it bluntly in the TechTarget piece: "At the end of the day, an AI implementation is a software implementation." What this suggests is that the discipline behind scaling AI looks far more like architecture than experimentation.
Almost nobody checks scalability first
Fewer than a third of organisations formally assess scalability before launching an AI initiative, according to the same study. The two barriers most often reported were complex regulatory frameworks and legacy system integration.
This is where most organisations struggle. Scalability gets treated as something to sort out after the pilot earns its funding. By then the team has promised a date, the vendor has been signed and the architecture questions arrive as bad news instead of design inputs.
Abstraction is not the same as modernisation
There is a popular shortcut: put an API or abstraction layer over the old system and let the AI talk to that. Sometimes it is the right call. Tony Marron, who leads global engineering capability at Liberty Mutual, told TechTarget that abstraction does not replace modernisation and that the two go hand in hand.
The practical test is what the AI needs from the system. APIs tend to suit read-heavy or decision-support use cases. Low latency, transaction write-back and high-volume data transfer usually point towards modernising the system underneath. Ozkaya warned that poorly managed dependencies produce "cascade failures, unexpected latency spikes and silent correctness bugs". The last one is the dangerous one. Silent means nobody notices until a customer does.
Five checks before the next pilot goes wide
TechTarget's list of pre-scaling questions is a good starting point. Adapted for an architecture team, it looks like this:
- Data and context: is the data current, connected and governable, or hand-fed for the pilot?
- Integration: can the AI reach core systems without fragile workarounds?
- Identity and authority: do permissions still make sense at ten times the users?
- Performance: will the platform meet response time, write-back and volume needs?
- Operations: can evaluation, monitoring and human oversight be reused for the next deployment?
Identity and authority deserve extra attention in Australia. The Privacy Act's automated decision making disclosure requirements start on 10 December 2026, and you cannot disclose what you cannot map. For enterprise architecture in Australia, AI is now a regulatory mapping problem as much as a technical one.
What This Means for Your Organisation
What we see across organisations is that the architecture function gets invited in after the pilot succeeds. By then the useful decisions have already been made by default. The better pattern is to bring architecture in before the pilot is chosen, so the first use case is picked partly because it is representative of what the platform will need to carry.
It also helps to keep a running record of what each pilot quietly relied on: the manual data fix, the shortcut permission, the person who knew which field to ignore. That list is your real scaling backlog.
Key Takeaways
- Only 13 per cent of organisations in the BearingPoint study scaled AI as planned, and fewer than a third assessed scalability before launch.
- A pilot proves the idea works in favourable conditions, not that the enterprise can carry it.
- Abstraction layers help, but they do not replace modernisation when AI needs low latency, write-back or high data volumes.
- Run the five checks on data, integration, identity, performance and operations before the pilot is promoted.
How Trusenta Can Help
- Enterprise Architecture gives your team one place to map the applications, integrations and data flows an AI use case depends on before you scale it.
- AI Integration Services connects AI to core systems without the brittle workarounds that sink rollouts.
- AI Strategy helps you choose use cases that the architecture can carry.
Conclusion
Scaling AI is rarely held back by the model. It is held back by the data, integrations and permissions around it, which is the territory enterprise architecture has always covered. Organisations that treat the next pilot as an architecture decision first will probably find the gap between "it works" and "it works as planned" a lot smaller.
