Olufunsho Peters is the CEO of INFINION Technologies.
Every month, an executive team walks into a boardroom convinced they are six weeks from an AI breakthrough. They have a dazzling demo and the pilot works. Yet, the project stalls the moment it touches a real customer record or transaction.
Technology is not the problem. AI initiatives in regulated industries don’t fail because the models are weak. They fail because executives apply a startup mindset to an environment engineered for the opposite of speed.
RAND research shows that more than 80% of AI projects fail to deliver intended business value because they are treated like software upgrades. Below are the three issues that surface most often, along with a framework called the Governance-First Deployment Model that can help correct course.
1: Treating AI Pilots As Standalone IT Experiments
Leadership will often greenlight a pilot insulated from core operations and regulatory oversight. Of course, it performs beautifully, and often, the success of the pilot makes leadership attempt to scale it overnight.
A model that behaves flawlessly in a sandbox now collides head-on with data privacy statutes or real-time audit requirements in production. That’s when it hits the wall. What looked like a six-week rollout becomes a year of remediation.
Imagine a financial institution that builds a fraud-detection model trained on scrubbed, anonymized historical transactions. It flags fraud impressively in testing. But once wired into live transaction monitoring, the compliance team discovers the model can’t produce an audit trail explaining why it flagged a given customer, something that is a compliance requirement. The launch pauses indefinitely at a greater cost than if compliance had been in the design conversation from the start.
The workaround is to design for production and compliance from day one, not as an afterthought. Risk, legal and compliance leaders belong in the room from the moment of conception, not the approval stage.
2: Neglecting The Data Foundation Beneath The Algorithm
Executives are often drawn to the visible, exciting part of the story. Understandably so, but no machine learning model can ever outperform the data architecture beneath it. Too many organizations deploy sophisticated models on top of fragmented legacy silos, quietly assuming the algorithm will compensate for messy inputs. It won’t. An AI system cannot automatically fix bad data. Instead, it will amplify the issues that result.
Take, for example, a telecommunications provider rolling out an AI-driven billing anomaly detector after a merger. Customer records from the two legacy systems were never fully reconciled—some accounts carry duplicate IDs, others have mismatched address and usage histories. The model, trained on this patchwork, starts flagging thousands of legitimate customers as billing anomalies while missing real discrepancies buried in the messy records. Support lines flood with complaints, and what was meant to reduce billing disputes ends up manufacturing new ones.
How do you fix the foundational data? Reframe data governance as core infrastructure, not administrative overhead. Before a single model goes into production, leadership teams should be pressed on three questions: Can this data be traced to its source? Who has access, and why? Is the metadata clean enough to survive an audit? If the answer to any of these is uncertain, then that’s where to start.
3: Underestimating The Human Side Of The Deployment
The most persistent resistance to enterprise AI rarely comes from the infrastructure. It comes from the people who have spent careers mastering the compliance and risk frameworks the AI is now meant to support. Professionals often respond to new tools with either skepticism or, worse, an over-reliance that erodes the human judgment the organization depends on when leadership positions AI as a replacement or skips the investment in training people to be rigorous about interrogating its outputs.
An energy utility authority deploys an AI system to predict equipment failures across its grid. Yet the field engineers, who may be veterans, can’t trust the tool, so they quietly keep working off their own instincts until a predicted failure it correctly flagged goes uninvestigated and causes an outage. In the post-incident review, the deeper failure isn’t the model or the engineers; it’s that no one had defined when an AI alert required mandatory escalation versus when human judgment could override it. The tool and the workforce were never actually integrated and were left to operate in parallel.
To avoid this, enterprise organizations have to first build deliberate guardrails that define exactly where automation ends and human sign-off begins. Upskilling and change management need to extend to how to recognize model drift, question an anomalous output and override the system when needed. The organizations that get this right treat human oversight as a requirement.
The Governance-First Deployment Model
In regulated industries, designing and architecting for resilience beats being the first to ship or go to market in environments where trust is the product. The organizations that deploy AI successfully share three habits. First, they involve risk and compliance from the get-go. They treat data integrity as part of infrastructure, and they build human oversight into the architecture rather than around it.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
