AI at Work: Quona’s Series on Agentic AI

By Michel Zaidler and Fernanda Rezende

AI has moved from promise to performance — expanding the addressable market beyond software into labor and services, and enabling companies to scale faster with leaner teams. For fintech, this shift is particularly powerful: AI can unlock new products, reduce operational costs, improve risk decisions, and elevate customer experiences at scale. Across our Quona portfolio, we’re already seeing transformative gains from companies deploying AI.

Over the next few weeks, we’ll share a short series on how AI is reshaping the application layer. This first post explores what real moats look like in the age of AI — and how to build them.

Winning the Application Layer: Moats in Agentic AI

The narrative around AI-application startups has shifted significantly in recent quarters. What began with broad skepticism toward so-called “AI wrappers” has evolved into the rise of some of the fastest-growing and most valuable companies of the modern tech era.

At the same time, competition at the application layer has intensified. Incumbents are adopting AI more rapidly than in past platform shifts, and foundational model providers like OpenAI are moving upstream, blurring the lines between infrastructure and applications. Amidst this complexity, a core question remains: how can Agentic AI companies build real defensibility and lasting value?

We believe the application layer remains a fertile ground for building high-value, generation-defining companies. Below, we offer a framework for building defensibility at the application layer, drawn from extensive analysis of successful AI-native companies around the world.

1. End-to-End Solutions: A Thousand Miles Between a Great Model and a Great Product

Building effective agentic systems requires far more than access to a powerful model. It demands deep domain expertise, a sharp understanding of the needs of end users, and substantial investment in engineering, design, and infrastructure integration. As OpenAI’s Sam Altman aptly put it, “There are a thousand miles between a great model and a great product.”

Truly competitive AI-native solutions are not single agents, but sophisticated orchestration systems that coordinate multiple specialized agents tightly embedded within real-world workflows. Effective orchestration encompasses multi-agent collaboration, integration with external tools, rigorous evaluation frameworks, and coordination across diverse infrastructure components — including data stores, observability, memory, security, and, increasingly, identity and payment layers.

These systems should account for the inherent uncertainties of AI models, especially their nondeterministic behavior, where identical prompts can yield different outputs. Robust orchestration requires careful calibration between agent autonomy and system oversight. This includes implementing advanced error-handling strategies and fail-safe mechanisms to ensure resilience and reliability, even in complex or unpredictable environments.

2. Agent Specialization

General-purpose models often fall short in high-stakes scenarios where accuracy, reliability, and compliance are critical. Specialized agents, by contrast, thrive precisely because they narrow their functional scope, meticulously embedding domain-specific logic, precise operational rules, and detailed handling of edge cases. This intense focus translates into superior performance, reduced risk, and greater user trust, particularly in regulated industries such as finance and healthcare.

Specialized agents — whether aligned to a specific task or to a vertical — excel because they accumulate domain-specific expertise through continuous use and feedback-driven adaptation. Unlike generalist models that require extensive retraining to handle new tasks, specialized agents become progressively more adept through targeted iterations and incremental refinement, significantly increasing accuracy and efficiency over time.

3. Feedback Loops Through Usage and Experience

Feedback loops aren’t just a technical feature: they’re a core strategic asset. Agentic AI companies that embed continuous feedback mechanisms directly into their products (capturing user preferences, corrections, failure modes, and decision traces) unlock faster learning cycles and compound performance gains.

But data alone isn’t sufficient. Leading teams reinforce these loops with targeted evaluation systems that link feedback to tangible improvements in model behavior, user experience, and business outcomes. Evaluation-driven development is fundamental in applied AI.

Many companies bootstrap these loops through human-in-the-loop (HITL) systems. For example, Akido Labs, operating in healthcare, spent a year refining an “AI doctor” inside its own clinic network. The team used a closed-loop environment — augmented by human doctors and nurses — to observe, supervise, and iteratively improve the agent before deploying it more broadly. In financial services, Ramp applied a similar approach. The company embedded feedback capture directly into its expense workflows, allowing users to correct transaction categorizations and flag policy issues. These inputs feed into supervised pipelines that retrain internal models, improving classification accuracy and compliance over time.

4. Controlling workflows, Systems of Record (SORs) and Systems of Action (SOAs)

Controlling or deeply integrating into Systems-of-Record (SORs) and Systems-of-Action (SOAs) — core business systems like CRMs, ERPs and other vertical software — materially improves the utility of agentic systems.

These integrations unlock two essential capabilities: (i) access to rich, contextual data, and (ii) the ability to trigger actions in real time. For instance, an AI sales assistant becomes significantly more effective when it can draw on deal histories, customer segmentation, and account notes within a CRM. Likewise, a collections agent directly integrated into a loan management system can initiate outreach the moment a borrower becomes delinquent.

This underscores the strategic importance of systems of record and systems of action — the control points where context lives and decisions get executed. For established platforms, this is an advantage: existing platforms can activate rich, longitudinal data and defend share through integration and bundling, even if their agentic features start out less advanced.

At the same time, these control points are more open than they’ve been in years. AI-native entrants can wedge into high-ROI workflows, prove value quickly, and expand outward to replace — or redefine — the system of record itself. The next generation of SORs will be multimodal and continuous: designed to ingest unstructured, real-time data (often passively) and to power agentic actions, not just store static fields. Early AI-native CRMs and workflow platforms already reflect this shift.

5. Persistent Memory: Driving Personalization and Stickiness

Over the last twelve months, the leading foundation-model providers have treated long-term memory as the next competitive frontier. Substantial progress has already been made. Models have evolved from stateless interactions to systems that can recall past conversations and use them to generate more personalized and relevant responses.

Although memory still has limitations, the trajectory is promising. Over time, this will enable AI applications to become increasingly context-aware, delivering deeply personalized interactions. Personalized workflows, retained preferences, and adaptive tone will all contribute to switching costs.

Furthermore, this advanced memory capability opens avenues for deeper predictive insights and anticipatory actions. By recalling historical interactions and behavioral patterns, agents can proactively suggest next steps or flag potential issues before they arise. For example, a financial advisor assistant could detect that a user tends to overspend in the final week of each month and proactively suggest budget adjustments — or surface reminders tied to recurring obligations like loan repayments or tax deadlines.

6. Model Specialization

Developing specialized models remains non-trivial. It requires highly technical talent, high-quality domain data, and often meaningful compute resources. In practice, many fine-tuned models still underperform general-purpose models.

That said, the most capable AI application teams are increasingly building smaller, domain-specific models tailored to focused use cases. A growing body of research and real-world deployment evidence shows that specialized models can outperform generalist ones across key dimensions: lower latency, reduced inference costs, and higher task-specific accuracy.

This shift is supported by broader technological advancements, particularly in post-training methods such as reinforcement learning, supervised fine-tuning, and knowledge distillation, which enable model refinement at a lower cost than full pre-training. Synthetic data is also becoming a powerful tool to supplement proprietary datasets, enabling teams to simulate rare or hard-to-collect edge cases.

Several innovators in financial services are already moving in this direction:

● Stripe (payments) launched what it describes as “the industry’s first foundation model for payments,” trained on tens of billions of real transactions. Unlike its previous narrow machine learning models, this is a general-purpose model capable of powering multiple use cases — fraud detection, dispute resolution, authentication, and more. Built on transformer architectures, the model delivers vastly improved performance. According to Stripe, attack detection accuracy jumped from 59% to 97%.

● Nubank (consumer finance) is building its own foundation models from scratch to analyze financial behavior at scale. A centralized library of base models and fine-tuning pipelines enables different product teams to customize model outputs for specific use cases across the organization.

● Digits (accounting) benchmarked GPT-4.5 against its own domain-specific model on a dataset of nearly 18,000 financial transactions. While GPT-4.5 achieved 66% classification accuracy, Digits’ in-house model reached 93% — with an order-of-magnitude improvement in latency.

Wrapping Up: Architecting Moats Intentionally and AI-Native Implications

Defensibility in Agentic AI isn’t accidental — it’s the result of deliberate strategy and relentless execution. It stems from strategic choices around specialization, user experience, and structural alignment that compound over time into durable, hard-to-replicate advantages. Traditional moats like distribution and brand still matter — and, when combined with AI-native design, can further extend this edge.

So, what does it mean to be AI-native? At its core, it’s about building with the assumptions and constraints of an AI-first world from day one. This orientation shapes everything: infrastructure, product interfaces, go-to-market motion, organizational structure, and monetization strategy. For truly AI-native companies, this alignment is often their strongest advantage.

In our next post, we’ll unpack what being AI-native looks like in practice, and how today’s AI-first businesses differ in structure, strategy, and pace from more traditional technology companies.