PUB SYS LIVE
[SYS.04]
AI 2026.04.13

From Proof of Concept to Production: Why AI Projects Fail in Complex Enterprises and How to Fix That

Most AI initiatives do not fail in the model. They fail in the transition from proof of concept to production.

In regulated environments, this transition is not a deployment step. It is a system constraint problem. Governance, compliance, and operational risk define what can be deployed.

In financial market infrastructure, the stakes are structural. Trading platforms, market surveillance, clearing engines, and market data systems operate under strict latency constraints, regulatory oversight, and continuous availability requirements.

The conditions that make AI interesting in a lab are precisely the conditions that make it difficult to operationalize in production.

Most organizations discover this only after the proof of concept is declared successful.

Where Proofs of Concept Break

Data quality versus production data reality

A proof of concept typically runs on clean, curated, often static datasets. Production systems emit live, high-frequency, noisy data. Surveillance engines ingest millions of events per session. Order management systems produce streams with microsecond timestamps that require precise normalization.

When a model trained on clean historical data meets production feeds, behavior degrades in ways that are not always predictable and rarely visible until failure occurs.

Integration with legacy systems

Most financial infrastructure was not built for model inference. Core trading systems, risk engines, and clearing platforms carry legacy architectures that predate modern API standards.

A proof of concept runs alongside production systems. A production deployment runs through them.

Integrating an inference layer into a real-time matching engine or a post-trade workflow requires latency budgets, failure handling, fallback logic, and change management processes that no proof of concept team typically scopes in advance.

Non-deterministic model behavior

In market surveillance systems such as SMARTS, a model that cannot explain why a pattern was flagged cannot be operationalized.

The output must support analyst review and regulatory scrutiny, not just model accuracy.

Latency constraints

Many AI frameworks were not designed for microsecond environments.

In trading systems, latency budgets are measured in single-digit microseconds for co-located infrastructure and in low milliseconds for adjacent systems.

Inference calls that introduce unpredictable tail latency require architectural changes to the surrounding system, which a proof of concept does not anticipate.

Why Governance Blocks Scale

Auditability requirements

Regulators such as FINMA require that firms can reconstruct and explain decisions made by automated systems.

An AI model must produce audit-ready outputs: what input it received, what it decided, on what basis, and when.

Most proof of concept implementations do not instrument for this.

Model explainability

Explainability is not a feature. It is a precondition for deployment.

If the only output is a probability score without traceable reasoning, this prevents deployment in any consequential workflow.

Change management and release processes

Financial infrastructure operates under strict change control.

An AI model that retrains or changes behavior is effectively a new software release and must be treated as one.

Few organizations have adapted their release processes to account for model versioning and behavioral regression.

Regulatory constraints under FINMA and DORA

AI systems relying on external model APIs or vendor infrastructure introduce third-party risk.

These dependencies must be assessed, documented, and tested for resilience.

Many proofs of concept rely on infrastructure that cannot meet production regulatory requirements.

Organizational Failure Points

Separation between innovation and production teams

The team that builds the proof of concept is rarely the team that runs production.

Without integration from the outset, the handover becomes a failure point.

Lack of ownership

A proof of concept has a sponsor. It rarely has an owner.

Without a defined production owner, the work stops at experimentation.

Vendor-driven pilots

Many AI pilots succeed because the vendor carries them.

When the engagement ends, the internal team lacks the capability to sustain the system.

What Actually Works

Treat AI as system integration

An AI component is a system component.

It must be designed with interfaces, failure modes, latency constraints, and observability from the start.

Build production constraints into the proof of concept

A proof of concept that ignores production constraints does not prove viability.

It proves that the model works in isolation.

Align engineering, risk, and compliance early

In regulated environments, these functions are inputs to the design, not reviewers at the end.

Define ownership before starting

If no one owns the system in production, the proof of concept should not start.

Conclusion

The problem is not model capability.

The problem is deploying models inside systems where failure has regulatory and operational consequences.

Organizations that succeed do not treat AI as innovation.

They treat it as infrastructure.