In a short clip released by OpenAI, I—Daniel Rabinovich, Chief Operating Officer at Mercado Libre—shared what GPT-5 is enabling at our company. As the largest e-commerce and fintech platform in Latin America, we face a scale and complexity that forces us to rethink how we design and operate intelligent systems. In that conversation I explained how a single model that merges conversational ability, reasoning, and tool use is changing the way we build agents and deliver services to millions of customers. This article expands on those remarks, reporting the facts, interpreting the implications, and outlining how this capability is reshaping internal operations, developer practices, and customer experiences across our marketplaces and financial products.
🧭 Executive summary
Mercado Libre operates thousands of internal tools and serves millions of users across Latin America. Historically, managing AI agents at this scale required juggling multiple specialized models: some dedicated to natural conversation, others to complex reasoning, and yet others to orchestrate external tools. GPT-5 offers a unified approach—one model that can converse naturally, perform multi-step reasoning, and call tools with the proper parameters. This combination reduces architectural complexity, shortens development cycles, and improves the way we support sellers and buyers through situations that would even challenge human operators.
In my remarks for OpenAI’s short clip, I stressed that our work involves helping sellers navigate situations "even complicated for human beings." I also pointed to a critical operational truth: "The ability to use the right tool with the right parameters is absolutely key to scale." With GPT-5, we can bring that capability into a more maintainable, cohesive form.
🤝 Background: Mercado Libre at scale
Mercado Libre is by far the largest e-commerce and fintech platform in Latin America. We host markets across multiple countries, handle payments, provide credit, manage logistics, and offer seller support. Our ecosystem includes marketplaces, payment processors, logistics networks, and a dense set of internal tools that employees and automated systems use every day.
At our scale, complexity is inevitable. Sellers ask for refunds, dispute chargebacks, request help with customs documentation, inquire about inventory synchronization, and need guidance on compliance and best practices. Consumers want transparent tracking, simple returns, and quick resolution for payment issues. All of this happens across many languages, jurisdictions, and regulatory environments.
To keep the platform both efficient and trustworthy, we rely heavily on automation, human teams, and intelligent agents that assist our workforce and customers. The real challenge is not only to automate but to automate correctly: to apply the right policy, at the right time, with the right tool, while preserving trust and safety.
🔧 The challenge of building agents at scale
Building an agent is a complex endeavor. It looks simple when you describe it in one sentence—create a system that can understand user intent, reason about context, and act through tools or workflows—but in practice it involves multiple distinct capabilities that must be integrated carefully.
First, the agent must engage in natural conversation. This requires fluency across languages and dialects, the ability to maintain context over multiple turns, and an empathetic tone that aligns with customer expectations. Second, the agent must reason: it needs to infer missing information, reconcile conflicting signals, and plan multi-step procedures. Third, the agent must use tools—querying databases, creating tickets, initiating refunds, or calling third-party APIs—while ensuring that parameters are accurate and actions are auditable.
Previously, our architecture often required different models for each capability. One model would excel at conversational fluency but lack deep reasoning. Another model would be optimized for planning or symbolic manipulation but not for natural language. Tool use was frequently handled by orchestration layers that invoked separate reasoning engines. Managing these components demanded careful engineering, expensive runtime coordination, and significant effort to keep behavior consistent across model boundaries.
When I say "building an agent is a complex endeavor," I mean it in both technical and organizational terms. Each additional model brings an integration surface, an operational burden, and a potential point of failure. At our scale, those costs are amplified. We have thousands of internal tools and workflows; coordinating them through disjointed models becomes untenable.
🤖 Why unified conversation and reasoning matters
The great thing about GPT-5 is that it combines the ability to have conversation and reasoning on the same model. That single statement encapsulates a profound shift for us.
Combining conversation and reasoning into one model eliminates the brittle boundaries between “understanding” and “planning.” Rather than interpret user intent in one module and pass structured intent to a separate reasoning engine, we can now let the same model maintain continuity between comprehension, inference, and action. This continuity reduces miscommunication, improves context preservation, and streamlines error handling.
Before GPT-5, you had to use two different models to do so. That meant bridging the semantic gap between natural language outputs and the structured inputs expected by reasoning modules. We engineered many translation layers to cope with that gap, introducing latency and sometimes unpredictability. With GPT-5, reasoning is native to the conversational fabric. The model can weigh options, simulate outcomes, and propose next steps while still speaking in a human-friendly way.
For customer-facing applications, that shift is especially meaningful. Consider a seller dealing with a complex logistics exception: the package is delayed, customs paperwork is incomplete, and the buyer is threatening cancellation. A previous architecture might run a conversational model to triage the seller’s problem, then hand off to a reasoning pipeline to compute remediation options, and then call a separate tools layer to initiate a refund or paperwork. With GPT-5, a single model can maintain the seller’s context, reason about constraints and trade-offs, and in the same flow determine which tools to call with precise parameters.
⚙️ Tool use: the fulcrum of scalable automation
I have often said—and I repeat here—that "the ability to use the right tool with the right parameters is absolutely key to scale." At high volume, small mistakes in parameter selection, or a lack of precision when invoking tools, lead to outsized operational costs and degrade trust.
Tool use in production requires three core competencies:
- Accurate intent mapping: Understanding what the user wants and mapping that intent to the correct tool.
- Parameter correctness: Supplying precise, validated parameters to the tool—dates, amounts, identifiers, codes—so that the tool executes exactly as intended.
- Auditable actions: Ensuring every action taken by an agent is logged, reversible when appropriate, and consistent with policy.
With GPT-5, these competencies emerge more naturally. The model’s integrated reasoning helps it disambiguate intent and derive missing parameters. For example, if a seller asks to "reschedule a pickup for next week," the model can infer timezone, local holidays, and the appropriate pickup window from context, and then call the scheduling tool with validated parameters. It can also explain its reasoning in natural language if asked, increasing transparency.
This transparency is important for trust. When an agent explains why it chose a particular tool or parameter, a human operator can more easily validate or correct the decision. That reduces escalation rates and allows automation to handle a larger share of interactions safely.
📈 Operational gains and risk reduction
Unifying conversation, reasoning, and tool use has concrete operational benefits. At a high level, those benefits fall into three categories: development velocity, runtime efficiency, and operational resilience.
- Development velocity: Building with a single model reduces the complexity of pipelines, which shortens development time. Teams can focus on embedding domain knowledge and safety policies directly into a single model-driven flow rather than engineering bridges between specialized models.
- Runtime efficiency: A unified model reduces latency since fewer model-to-model handoffs are needed. This can lower compute costs and improve user experience—the agent responds faster with fewer intermediate steps.
- Operational resilience: Fewer components means fewer points of failure. Troubleshooting is simpler when the chain of responsibility is clearer, and rollbacks are easier because changes are consolidated into a smaller set of artifacts.
At the same time, risk reduction is not automatic. A more capable model widens the range of actions the agent could perform, so governance and monitoring must scale as well. We need robust guardrails around sensitive operations—financial transactions, account modifications, or regulatory compliance workflows. Unification simplifies some aspects of governance but raises the bar for how we instrument models and enforce policies.
🧩 Practical example: helping sellers with complex cases
To make these ideas concrete, let me walk through a practical example—one I mentioned briefly: helping sellers deal with complicated situations that are even complicated for human beings.
Imagine a seller receives a return claim that appears to be fraudulent. The buyer disputes the condition of the item, but the seller has evidence—photos, package tracking, and serial numbers—that suggest a mismatch. The case involves the marketplace, payment authorization, and possibly a logistics carrier. Human teams can handle such cases, but they often require cross-referencing multiple systems and reconstructing a timeline of events.
With GPT-5, an agent can:
- Engage the seller conversationally, eliciting missing details and clarifying ambiguous statements.
- Reason across multiple data sources to reconstruct the timeline, weighing evidence such as tracking timestamps, photo timestamps, and payment authorization logs.
- Decide whether the case should escalate to a human reviewer or whether the system can resolve it automatically by initiating a refund, opening a carrier claim, or reinstating seller protections.
- Call the necessary tools with validated parameters—for instance, generating a carrier claim with the correct shipment ID, or creating a refund with the correct amount and reason codes.
- Log the entire decision chain in an auditable format and explain the decision to the seller in plain language.
That chain—from natural conversation to evidence-based reasoning to safe tool invocation—happens in a single flow with GPT-5. The agent’s explanations improve seller confidence, and the auditable logs give our compliance teams the data they need to validate outcomes.
📚 Developer experience: fewer moving parts, more focus
From a product and engineering perspective, a unified model changes the way teams work. Previously, developers needed expertise in multiple specialized models and the glue that connected them. Now, teams can concentrate on three things:
- Prompt engineering and domain knowledge: Crafting prompts that accurately reflect business rules, edge cases, and the tone we want agents to adopt.
- Tool definitions and interfaces: Designing clear tool APIs with robust validation so the model can call them reliably.
- Monitoring and safety: Building observability, guardrails, and human-in-the-loop processes around the model's actions.
This focus has two practical implications. First, training cycles become shorter because there are fewer integration points to test. Second, cross-functional teams—product managers, legal, compliance, and engineers—can collaborate more directly on the agent’s behavior since the end-to-end flow is more cohesive.
However, this does not mean that prompt engineering is a silver bullet. Prompts must be crafted with care to encode policies, disambiguate intents, and provide fallbacks when the model lacks necessary data. Robust validation in the tools layer remains essential; the model should never be the only line of defense for critical parameters like monetary amounts or legal codes.
🔐 Trust, safety, and compliance concerns
More capable models magnify both opportunity and responsibility. When an agent can perform complex financial operations or resolve disputes autonomously, we must be rigorous about trust and safety.
Key areas of focus include:
- Access control: Ensuring an agent can only call tools and perform actions appropriate to the user’s role and verified identity.
- Parameter validation: Applying strict checks on any parameter supplied by the model before execution—amount ranges, account IDs, jurisdiction codes.
- Auditability: Storing a clear, retraceable record of decisions, inputs, and the model’s reasoning so that human teams can review and, if needed, reverse actions.
- Human-in-the-loop: Defining thresholds where the model must escalate to a human—for example, high-value refunds, regulatory exceptions, or ambiguous evidence.
- Continuous monitoring: Tracking model behavior in production and running red-team scenarios to uncover failure modes and adversarial inputs.
For companies like ours that move money and manage sensitive user data, these safeguards are non-negotiable. We architect the system so the model suggests or drafts actions while endpoint systems verify and enforce policy before committing irreversible changes. This layered approach balances automation with control.
🔭 What this means for customers and sellers
For customers and sellers, the immediate benefits are improved speed and better outcomes. Questions get resolved faster. Sellers receive clearer guidance that takes into account the nuance of their situation. Buyers see more consistent handling of disputes and a smoother experience for returns and refunds.
Beyond speed, a unified model enhances quality. Because reasoning and conversation happen together, agents can surface uncertainties, propose alternatives, and adapt to unexpected scenarios. Instead of a rigid script, users interact with a system that can explain its decisions and learn from corrections.
That said, we must manage expectations. Automation will not replace human judgment in all cases. Instead, it augments it—allowing human specialists to concentrate on the most complex cases and strategic initiatives while routine and semi-structured problems are addressed efficiently by agents.
📦 Logistics, payments, and cross-functional integration
Mercado Libre's business model bundles marketplaces, payments, and logistics. The value of a unified conversational and reasoning model multiplies when it can operate across these domains.
Consider integrated flows where an agent must:
- Check payment authorization and refund eligibility with the payments system,
- Query shipping status from the logistics platform,
- And apply marketplace policies that define seller protections and buyer rights.
A single model that reasons across those contexts can produce coordinated actions—a refund in the payments system that follows an automated carrier claim in logistics, accompanied by an explanatory message to both buyer and seller. Those coordinated, multistep actions historically required careful choreography across teams and systems; now they can be proposed and, where appropriate, executed in a more seamless way.
🧠 The technical underpinnings: why GPT-5 feels different
From a technical standpoint, GPT-5 feels different because it blurs the lines between tasks we previously separated: natural language understanding, symbolic reasoning, and tool orchestration. The unification arises from architectural choices and training strategies that emphasize multi-task learning and grounded tool use.
Consolidating these capabilities into a single model reduces friction in the pipeline. Inputs remain in natural language for longer, enabling the model to leverage latent context that might otherwise get lost in translation into structured representations. The model can internally simulate the steps it will take, check constraints, and generate both a natural-language explanation and a structured API call.
This internal simulation is powerful. It allows the model to "think through" sequences of actions before selecting one. That kind of deliberative reasoning reduces impulsive or incorrect calls to external systems. When the model does propose an external action, it also provides the reasoning that led to that choice, which can be presented to a human reviewer or used for automatic validation.
🔁 Integration strategy: how we adopt unified models
Adopting a unified model like GPT-5 is not a lift-and-shift. It requires a deliberate integration strategy that balances speed with safety.
Our approach follows several practical steps:
- Start with low-impact flows: Pilot unified agents on tasks that are helpful but have limited downside if things go wrong—informational queries, basic seller guidance, and draft generation of support messages.
- Instrument heavily: Add logging, explainability, and monitoring from day one. Track not just outcomes but the model’s suggested actions and reasoning.
- Define clear escalation: Set thresholds for human approval based on impact, complexity, and regulatory sensitivity.
- Iterate with cross-functional teams: Involve product, legal, ops, and safety teams in prompt design and policy specification.
- Embed validation checks: Ensure any model-proposed parameter passes through deterministic validators before execution.
- Scale gradually: Expand to higher-impact flows only after measured success and robust guardrails.
This staged approach allows us to capture the benefits of unified models while managing the operational and legal risks that come with more autonomous actions.
📊 Measuring success: metrics that matter
To quantify the impact of unified models, we track a mix of business, technical, and governance metrics.
Business metrics:
- Time to resolution for support cases
- Seller satisfaction scores after agent interactions
- Reduction in escalation rates to human teams
Technical metrics:
- Latency from user query to tool invocation
- Rate of invalid tool parameters generated by the model
- Percentage of actions that require human override
Governance metrics:
- Audit coverage for automated actions
- False positive and false negative rates for policy enforcement
- Frequency of security incidents tied to model actions
Combining these metrics gives a holistic view of whether the unified model improves outcomes while maintaining safety and compliance.
🔬 Ongoing research and future directions
Although GPT-5 brings capabilities that simplify many agent-building tasks, research and engineering work continues. Key areas of exploration include:
- Improved grounding: Ensuring the model reliably references authoritative sources and reduces hallucination, especially in legal and financial contexts.
- Fine-grained policy conditioning: Allowing the model to internalize and apply complex regulatory rules across jurisdictions.
- Explainable decision trails: Enhancing the model’s ability to produce concise, auditable rationales for decisions that map cleanly to internal compliance requirements.
- Robustness to adversarial inputs: Hardening models against malicious attempts to manipulate behavior or trick the system into unsafe actions.
- Human-model collaboration: Designing interfaces and workflows that let humans guide, correct, and teach models in real time.
These areas are active research fronts across industry and academia. For a company operating at our scale, progress in these domains directly translates into better user experiences and reduced operational risk.
🌎 Regional implications: why Latin America matters
Latin America presents unique opportunities and constraints that shape how we deploy unified models. Multi-lingual support is essential—Spanish and Portuguese dominate, but local variations and indigenous languages also exist. Connectivity constraints in some regions demand efficient inference and robust offline fallbacks. Regulatory environments differ by country, which increases the importance of fine-grained policy conditioning and local legal expertise.
At the same time, the potential upside is enormous. Many markets in the region are undergoing rapid digitization. Better automated support and intelligent agents can increase financial inclusion, help small merchants scale online, and reduce friction in cross-border trade. A unified model simplifies the engineering needed to deliver these benefits at scale because it reduces the need for many bespoke integrations across languages and local systems.
📝 A few concrete lessons we've learned
From our early experience with GPT-5, several lessons stand out:
- Start small, iterate fast: Early pilots on low-risk flows allow you to learn quickly and improve prompts and validations.
- Instrument every action: Logging, explainability, and policy checks are not optional; they are integral to trust and governance.
- Design tools with validation in mind: Clear interfaces and strict parameter checks prevent many operational errors.
- Keep humans in the loop: For high-impact decisions, human oversight remains essential. Automation should reduce human workload, not eliminate accountability.
- Invest in cross-functional collaboration: The best outcomes arise when engineers, product managers, legal, and operations design agents together.
📣 Broader industry implications
The unification of conversation, reasoning, and tool use in models like GPT-5 will ripple through industries. Financial services, customer support, healthcare, and public services can all benefit from agents capable of coherent, evidence-based action. The consequences include faster service, more personalized experiences, and the capacity to automate complex workflows that were previously too brittle or expensive to handle at scale.
At the same time, regulators and industry bodies will need to adapt. Auditable decision trails, standards for parameter validation, and norms for human oversight will become central to responsible deployment. Companies that embed these practices early will be better positioned to scale responsibly.
🔗 Closing thoughts: the start of a new chapter
GPT-5 represents more than a new model; it represents a new approach to building intelligent systems. By uniting conversation, reasoning, and tool use in a single model, we can simplify architectures, improve agent coherence, and deliver better experiences for customers and sellers. That said, capability is not a substitute for governance. The "right tool with the right parameters" remains the fulcrum of safe automation, and we must keep investing in validation, monitoring, and human oversight.
As Mercado Libre continues to adopt these technologies, my focus is on maintaining a balance: pushing the frontier of automation to increase efficiency and inclusion, while preserving the trust and safety that underpins our marketplace and fintech services. The early results are promising—GPT-5 helps us reduce friction, bring complex case resolution within reach of automation, and empower our teams to handle scale more effectively. But we will move deliberately, measuring outcomes and refining our approach as we learn.
If you work in product, engineering, or operations and are considering similar transformations, my practical advice is simple: instrument heavily, design tools with validation in mind, and keep humans in the loop where impact is high. When you do those things, unified models like GPT-5 become powerful enablers rather than sources of uncontrolled risk.
📬 Final notes and acknowledgments
I want to acknowledge OpenAI for advancing models that allow companies like ours to rethink agent design. The work to integrate conversational intelligence, reasoning, and tool use into a coherent flow is a major step forward for practical AI deployments. As we continue to deploy and study these systems, I look forward to sharing our learnings with the broader community so we can collectively raise the bar for safe, useful, and inclusive automation.
"Mercado Libre is by far the largest e-commerce and fintech platform in Latin America." — Daniel Rabinovich
"We're helping sellers deal with complicated situations that are even complicated for human beings." — Daniel Rabinovich
"The ability to use the right tool with the right parameters is absolutely key to scale." — Daniel Rabinovich
Thank you for reading. I’m Daniel Rabinovich, and I’ll continue reporting on how these technologies evolve at Mercado Libre and across the industry.
