Editor’s Note: We recently spoke with Evan Fire, Chief Operating Officer and Managing Principal at Pzena Investment Management, who shared his perspective on model-agnostic architecture, AI governance, and explainability in institutional investment management. His comments appear below.
Every few months, the conversation around artificial intelligence shifts toward a new winner.
One model introduces stronger reasoning. Another expands its context window. An open-weight alternative closes the performance gap. Enterprises rush to compare benchmarks, evaluate costs, and debate which large language model (LLM) should power their AI initiatives.
It is an understandable conversation. It is also the wrong one.
As enterprise AI matures, the organizations that gain the greatest long-term advantage will not be those betting on today’s leading model. They will be the ones building an AI architecture capable of evolving regardless of which model leads the market tomorrow.
The question facing enterprise leaders is no longer which AI model to use. It is how to build an AI platform that stays effective as models, technologies, and business requirements change.
That shift, from model selection to architectural flexibility, is one of the defining technology decisions of this decade. Across the industry, the conversation is already moving beyond individual models and toward agentic AI, governance, and the enterprise architecture required to scale AI responsibly.
Why Enterprise AI Needs Model-Agnostic Architecture
Nearly every software vendor now claims to offer an AI platform.
In many cases, what is presented as a platform is a user interface wrapped around a single AI provider.
That may accelerate initial deployment. It also creates vendor dependency.
Artificial intelligence is evolving at an extraordinary pace. New frontier models emerge regularly. Open-source innovation accelerates. Pricing changes. Capabilities improve. Regulatory expectations evolve. Building an enterprise AI strategy around a single provider means tying your roadmap to someone else’s roadmap.
That is not a platform. It is a feature with better packaging.
A true enterprise AI platform should outlive today’s leading model. That means thinking beyond individual models and toward a model-agnostic agent layer.
Rather than embedding enterprise workflows directly into a single LLM, organizations should build intelligent agents capable of orchestrating whichever model is best suited for the task at hand. As new models emerge, improve, or become more economical, the architecture evolves while the business workflows remain intact.
The hard part sits underneath. Swapping a model breaks prompts and invalidates evaluation baselines, so the agent layer has to own routing, evaluation, and data-egress rules, not just the API call. Firms that treat model-agnostic as a configuration setting find this out on the first swap.
Built properly, AI models become interchangeable components rather than permanent dependencies.
This flexibility is not simply about avoiding vendor lock-in. It is about preserving optionality in a market where change is the only constant. Architecture, not the model itself, is the real competitive advantage in enterprise AI.
During a recent discussion on enterprise AI strategy, Evan Fire, Chief Operating Officer and Managing Principal at Pzena Investment Management, put it this way:
“Choosing a model is not a technology decision anymore. You inherit that lab’s pricing, its outages, its safety posture, and its competitive future. That is the same list a board already runs on any critical vendor. We would rather make model choice a piece of infrastructure we can swap than a bet we have to defend.”
The Winning Architecture Combines Frontier and Open-Source AI
Being model-agnostic does not mean avoiding commercial AI models. It does not mean betting entirely on open source.
The strongest architectures recognize that different workloads demand different capabilities. The future is not frontier models versus open-source models. It is architectures capable of using both.
Frontier models provide the highest level of reasoning for complex analytical tasks. Open-source models are the better choice when organizations prioritize cost efficiency, private deployments, data residency, or customization.
Four questions drive the routing decision. How sensitive is the data? How fast does the answer need to come back? What does the task cost at production volume? Does the output need to be reproducible?
The strongest enterprise AI platforms will not force organizations to choose. They will orchestrate both, routing each workload to the model best suited for the job.
Fire framed the trade-off directly:
“Frontier models buy you the capability ceiling. Open source buys you cost control, data residency, and the ability to keep sensitive work inside your own walls. Routing each workload to the right one is not a compromise between them. It is the point.”
That is a fundamentally stronger architecture than committing every workflow to a single provider. It gives organizations the flexibility to adopt new capabilities as they emerge without redesigning the platform underneath.
Choosing an LLM Is Becoming an Enterprise Risk Decision
Selecting an LLM is one of the most consequential vendor decisions an enterprise can make.
Commit your AI strategy to a single provider and you inherit more than its technology. You inherit its pricing decisions, service availability, product roadmap, safety posture, release cadence, and competitive future.
Leadership teams now evaluate questions that extend well beyond engineering.
- How exposed are we if pricing changes dramatically?
- What happens if a provider experiences an outage?
- How do we address data residency requirements across jurisdictions?
- Can sensitive information remain within approved environments?
- How will evolving regulations affect our AI strategy?
- What happens if a better model emerges six months from now?
These are governance questions every bit as much as technology questions.
Organizations have navigated similar decisions before when selecting cloud providers, cybersecurity vendors, and other forms of critical infrastructure. AI is joining that category.
The firms that recognize this early will not simply be better prepared for technology change. They will build AI strategies capable of adapting alongside evolving regulations, market conditions, and business priorities.
AI Must Produce Decisions That Can Be Defended
As AI becomes embedded in investment management and other highly regulated industries, another reality is emerging. Enterprise AI cannot simply produce convincing recommendations. It must produce recommendations that withstand scrutiny.
For institutional investment managers, every AI-generated insight eventually faces the same question: why?
If an investment committee, regulator, auditor, or client asks why an AI recommended a trade, flagged a compliance issue, or surfaced a particular insight, “the model thought it was right” is not an acceptable answer.
Fire drew the line between enterprise AI and consumer AI:
“We are fiduciaries. An answer that sounds right is worth nothing if we cannot show where it came from. Every recommendation has to carry its provenance: the data behind it, the reasoning, and enough of a trail that another professional can reproduce it and defend it to a client or a regulator.”
Decision intelligence means every recommendation traces back to the data that informed it, the reasoning behind it, and the evidence supporting it. Every output should be explainable. More importantly, every output should be defensible.
Every recommendation should answer:
- What data informed this conclusion?
- How was this recommendation generated?
- Can another professional reproduce the result?
- Can compliance explain the underlying reasoning?
- Can the organization defend this decision to regulators, auditors, clients, or investment committees?
Without transparency, confidence erodes.
This is where the conversation shifts from generative AI to decision intelligence. Decision intelligence is not defined by how convincing an AI response sounds. It is defined by whether every recommendation can be defended. In regulated industries, explainability is not a luxury. It is a requirement.
The organizations that succeed will not simply deploy AI. They will deploy AI that people trust. As AI systems become more autonomous, governance, auditability, and traceability matter as much as intelligence itself.
Looking Beyond Today’s Models
Today’s leading AI models will not remain today’s leading AI models forever. Foundational technologies evolve faster than most organizations can rebuild around them.
The companies that thrive will not necessarily have access to better models than everyone else. They will have built architectures capable of taking advantage of whatever comes next. That architecture will be:
- Model-agnostic rather than vendor-dependent.
- Flexible enough to incorporate new innovations.
- Governed through strong security, oversight, and policy controls.
- Explainable, so every recommendation can be understood and defended.
- Resilient enough to adapt as technology, regulation, and business priorities evolve.
New frontier models will emerge. Open-weight alternatives will mature. Pricing structures will change. Regulations will evolve. None of that should require an organization to rethink its AI strategy. Competitive advantage will not come from choosing the right AI model. It will come from making sure your enterprise is not dependent on any single one.
About INDATA
At INDATA, this principle is shaping the evolution of Nexus AI: a governed AI platform designed to connect investment workflows, enterprise data, and multiple model options without forcing firms to anchor their long-term strategy to a single provider. The objective is not merely to add AI capabilities, but to give firms a secure, adaptable foundation for using them responsibly.
Disclaimer: Pzena Investment Management has been an INDATA client since 2005. The comments and perspectives shared by Evan Fire in this article reflect his own views on enterprise AI architecture and are provided for industry discussion and informational purposes.
