The truth problem: Why verifiable AI is the next strategic mandate

A few years ago, a model we had integrated for customer analytics produced results that looked impressive, but no one could explain how or why those predictions were made. When we tried to trace the source data, half of it came from undocumented pipelines. That incident was my “aha” moment. We didn’t have a technology problem; we had a truth problem. I realised that for all its power, AI built on blind faith is a liability.

This experience reshaped my entire approach. As artificial intelligence becomes central to enterprise decision-making, the “truth problem,” whether AI outputs can be trusted, has become one of the most pressing issues facing technology leaders. Verifiable AI, which embeds transparency, auditability and formal guarantees directly into systems, is the breakthrough response. I’ve learned that trust cannot be delegated to algorithms; it has to be earned, verified and proven.

The strategic urgency of verifiable AI

AI is now embedded in critical operations, from financial forecasting to healthcare diagnostics. Yet as enterprises accelerate adoption, a new fault line has emerged: trust. When AI decisions cannot be independently verified, organisations face risks ranging from regulatory penalties to reputational collapse.

Regulators are closing in. The EU AI Act, NIST AI Risk Management Framework and ISO/IEC 42001 all place accountability for AI behavior directly on enterprises, not vendors. A 2025 transparency index has found that leading AI model developers scored an average of 37 out of 100 on disclosure metrics, highlighting the widening gap between capability and accountability.

For me, this means verifiable AI is no longer optional. It is the foundation for responsible innovation, regulatory readiness and sustained digital trust.

The 3 pillars of a verifiable system

Verifiable AI transforms “trust” from a matter of faith into a provable, measurable property. It involves building AI systems that can demonstrate correctness, fairness and compliance through independent validation. In my career, I’ve seen that if you cannot show how your model arrived at a decision, the technology adds risk instead of reducing it. This practical verifiability spans three pillars.

1. Data provenance: Ensuring all training and input data can be traced, validated and audited

In one early project back in 2017, we worked with historic trading data to train a predictive model for payment analytics. It looked solid on the surface until we realized that nearly 20 percent of the dataset came from an outdated exchange feed that had been quietly discontinued. The model performed beautifully in backtesting, but failed in live trading conditions.

This incident was a wake-up call that data provenance is not about documentation; it is about risk control. If you cannot prove where your data comes from, you cannot defend what your model does. This principle of reliable data sourcing is a cornerstone of the NIST AI Risk Management Framework, which has become an essential guide for our governance

2. Model integrity: Verifying that models behave as intended under specified conditions

In another project, a fraud detection system performed perfectly during lab simulations but faltered in production when user behavior shifted after a market event. The underlying model was never revalidated in real time, so its assumptions aged overnight.

This taught me that model integrity is not a task completed at deployment but an ongoing responsibility. Without continuous verification, even accurate models lose relevance fast. We now use formal verification methods, borrowed from aerospace and defense, that mathematically prove model behavior under defined conditions.

3. Output accountability: Providing clear audit trails and explainable decisions

When we introduced explainability dashboards into our AI systems, something unexpected happened. Compliance, engineering and business teams started using the same data to discuss decisions. Instead of debating outcomes, they examined how the model reached them.

Making outputs traceable turned compliance reviews from tense exercises into collaborative problem-solving. Accountability does not slow innovation; it accelerates understanding.

These principles mirror lessons from another domain I have worked in: blockchain, where verifiability and auditability have long been built into the system’s design.

What blockchain infrastructure taught me about AI verification

My background in building blockchain-based payment systems fundamentally shaped how I approach AI verification today. The parallel between payment systems and AI systems is more direct than most technology leaders realize.

Both make critical decisions that affect real operations and real money. Both processes transact too quickly for humans to review individually. Both require multiple stakeholders, customers, regulators and auditors to trust outputs they cannot directly observe. The key difference is that we solved the verification problem for payments more than a decade ago, while AI systems continue to operate as black boxes.

When we built payment infrastructure, immutable blockchain ledgers created an unbreakable audit trail for every transaction. Customers could independently verify their payments. Merchants could prove they received funds. Regulators could audit everything without accessing private data. The system wasn’t just transparent, and it was cryptographically provable. Nobody had to take our word for it.

This experience revealed something crucial: trust at scale requires mathematical proof, not vendor promises. And that same principle applies directly to AI verification.

The technical implementation is more straightforward than many enterprises assume. Blockchain infrastructure or simpler append-only logs can document every AI inference, what data went in, what decision came out and what model version processed it. Research from the Mozilla Foundation on AI transparency in practice confirms that this kind of systematic audit trail is exactly what most AI deployments lack today.

I’ve seen enterprises implement this successfully across regulated industries. GE Healthcare’s Edison platform includes model traceability and audit logs that enable medical staff to validate AI diagnoses before applying them to patient care. Financial institutions like JPMorgan use similar frameworks, combining explainability tools like SHAP with immutable audit records that regulators can inspect and verify.

The infrastructure exists. Cryptographic proofs and trusted execution environments can ensure model integrity while preserving data privacy. Zero-knowledge proofs allow verification that an AI model operated correctly without exposing sensitive training data. These are mature technologies, borrowed from blockchain and applied to AI governance.

For technology leaders evaluating their AI strategy, the lesson from payments is simple: treat AI outputs like financial transactions. Every prediction should be logged, traceable and independently verifiable. This is not optional infrastructure. It is foundational to any AI deployment that faces regulatory scrutiny or requires stakeholder trust at scale.

A leadership playbook for verifiable AI

Each of those moments, discovering flawed trading data, watching a model lose integrity and seeing transparency unite teams, shaped how I now lead. They taught me that verifiable AI is not just technical architecture, it is organisational culture. Here is the playbook that has worked for me.

• Start with an AI audit and risk assessment. Our first step was to inventory every AI use case across the business. We categorized them by potential impact on customers, operations and compliance. A high-risk system, like one used for financial forecasting, now demands the highest level of verifiability. This triage allowed us to focus our efforts where they matter most.
• Make verifiability a non-negotiable criterion. We completely changed our procurement process. When evaluating an AI vendor, we now have a checklist that goes far beyond cost and performance. We demand evidence of their model’s traceability, documentation on training data and their methodology for ongoing monitoring. This shift fundamentally changed our vendor conversations and raised transparency standards across our ecosystem.
• Build a culture of skepticism and accountability. One of our most crucial changes has been cultural. We actively train our staff to question AI outputs. I tell them that a red flag should go up if they can’t understand or challenge an AI’s recommendation. This human-in-the-loop principle is our ultimate safeguard, ensuring that AI assists human judgment rather than replacing it.
• Invest in the right infrastructure. Building verifiable AI requires investment in data pipelines, lineage tracking and real-time monitoring platforms. We use model monitoring and transparency dashboards that catch drift and bias before they become compliance violations. These platforms aren’t optional — they’re foundational infrastructure for any enterprise deploying AI at scale.
• Translate compliance into design from the start. I used to view regulatory compliance as a final step. Now, I see it as a primary design input. By translating the principles of regulations into technical specifications from day one, we ensure our systems are built to be transparent. This is far more effective and less costly than trying to retrofit explainability onto a finished product.

The path forward: From intelligence to integrity

The future of AI is not only about intelligence, it’s also about integrity. I’ve learned that trust in AI does not scale automatically; it must be designed, tested and proven every day.

Verifiable AI protects enterprises from compliance shocks, builds stakeholder confidence and ensures AI systems can stand up to public, legal and ethical scrutiny. It is the cornerstone of long-term digital resilience.

For any technology leader, the next competitive advantage will not come from building faster AI, but from building verifiable AI. In the next era of enterprise innovation, leadership won’t be measured by how much we automate, but by how well we can verify the truth behind every decision.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?