Artificial Intelligence

Practical perspectives on AI, from emerging capabilities to implementation and business impact

I’ve spent the last few years back in Medellín, Colombia, working in tech, specifically scaling software development and engineering operations for US companies expanding into the region. From this vantage point, I can tell you with absolute certainty that the AI conversation has completely shifted.

We are no longer debating whether AI matters. We are furiously figuring out how to integrate it into our operations without breaking what already works. This isn’t about hype anymore; it’s about infrastructure. And that reality changes everything about how we approach adoption.

The Current State: Tiers, Thinking, and Multimodal

When I first started paying close attention to AI around 2023, large language models (LLMs) felt like a novelty. Today, they are table stakes. But the most interesting development isn’t their mere existence. It is how rapidly the landscape is consolidating into practical capability tiers.

The Capability Split: A clear separation has emerged between frontier models (like OpenAI’s GPT-4o, Google’s Gemini, and Anthropic’s Claude) and highly efficient, open-weight models (like Llama, Mistral, and Qwen). Organizations now have a strategic choice. They can pay a premium for absolute top-tier reasoning, or they can run highly capable models locally with total data sovereignty.
The Speed-for-Accuracy Tradeoff: Reasoning models like o1 and DeepSeek-R1 have proven that giving models more time to “think” about hard problems produces far better results than raw scale. For many engineering workflows, we don’t need the fastest model; we need the one that gets the complex answer right.
Multimodal as a Requirement: Multimodality is no longer a buzzword. In 2026, frontier models process text, images, audio, and video natively. For tech operations, this means a single AI agent can watch a production incident unfold across monitoring dashboards, read the corresponding logs, and make informed routing decisions in real-time. This isn’t theoretical. Companies are building this infrastructure right now.

AI Agents: Where the Value Actually Lives

LLMs are just tools; agents are where the business value actually lives.

An AI agent is fundamentally different from a chatbot. It is autonomous software that understands an objective, plans a sequence of steps, executes them, and adapts when it encounters an error. We are finally seeing these move from laboratory experiments into production systems:

Finance teams are automatically processing invoices and reconciling accounts with 90%+ accuracy, cutting operational costs by 70%.
Supply chain operations are flagging supplier anomalies before they cascade into crises.
Field service teams are automating the scheduling and routing of technicians.
Healthcare providers are streamlining patient intake and claims processing.

The Reality Check: Gartner estimates that 40% of agentic AI projects will fail by 2027. Why? Because legacy systems cannot support modern AI execution demands. This isn’t a technology shortcoming; it is an architecture problem. Most enterprise systems simply weren’t built with the assumption that software would need to be partially autonomous. How we bridge that gap is the defining challenge of the next few years.

The Adoption Reality: Where Engineering Teams Get Stuck

I talk to engineering leaders constantly, and the failure pattern is incredibly consistent. They roll out an AI coding assistant. Developers use it. Code gets generated faster. And then, the productivity gains completely evaporate in the code review queue.

This is the dirty secret of AI in software development. AI didn’t solve the bottleneck; it just moved it downstream. You can generate code ten times faster, but if you have the same number of human reviewers, your actual throughput hasn’t changed.

Beyond the bottleneck, teams are colliding with three major hurdles:

Data Security: When developers paste proprietary business logic into external AI prompts, that data leaves your ecosystem. You need clear, technically enforced policies about what can be shared. Cultural guidelines aren’t enough.
The Skills Gap: Studies show 40% of enterprises lack the internal AI expertise to meet their goals. The solution isn’t to mass-hire AI experts. It is to train your existing engineers to view AI as an augmentation tool rather than a replacement threat.
The Measurement Crisis: While 61% of companies increased engineering budgets in 2025, only 20% actually use engineering metrics to measure AI’s impact. If you aren’t measuring, you don’t know if you’re getting value or just moving faster in the wrong direction.

Impact on Software Development Teams

I’ve watched teams adopt AI in three distinct ways: cautiously, at full-steam, or stuck in the middle. Only the cautious approach consistently works.

Winning teams start small. They target a narrow, specific workflow, like reducing onboarding time for new engineers or triaging first-response production alerts. They give the AI a limited scope, measure the impact, iterate, and then expand.

Conversely, struggling teams try to transform everything overnight. They deploy GitHub Copilot, integrate Claude into their wikis, and launch internal agentic workflows simultaneously without adjusting their underlying organizational processes. They end up with disjointed tooling that developers actively distrust.

Psychological safety is paramount here. Framing AI as “a tool to help you work faster” drives significantly better adoption than framing it as “a tool to prevent human error,” which triggers replacement anxiety. Furthermore, AI cannot fix organizational rot. If your developers lose 10 hours a week to context switching, unclear requirements, and approval wait times, generating code faster will only compound the chaos.

The LATAM Context: Punching Above Our Weight

Latin America is accelerating AI adoption faster than most anticipate. From Medellín, I see engineering teams integrating AI into production daily. The region accounts for 14% of global visits to AI solutions, despite having only 11% of global internet users. Brazil is leading the charge, closely followed by Mexico and Argentina.

However, we face real structural constraints. LATAM receives only 1.12% of global AI investment, and the brain drain of local AI specialists has accelerated.

This creates a fascinating asymmetry. LATAM companies are desperate to adopt AI but struggle to find deep local expertise. But for tech operations teams in the region, this is an opportunity. We already know how to operate with constrained resources. We know how to build resilient systems without unlimited capital.

Corporate adoption of AI agents in LATAM is expected to grow 327% over the next two years. The organizations that figure out how to implement these agents effectively will capture an enormous competitive advantage. They will achieve this by iterating quickly and building sustainable practices rather than waiting for perfect local talent pools.

Practical Advice for Leaders

If you are guiding an organization through this transition, keep these principles in mind:

Start with skepticism. Question where AI actually delivers value versus where you are just automating a process that didn’t need to exist in the first place.
Pick your first win carefully. Find one workflow where AI demonstrably saves time or unblocks people. Land a visible win in 90 days rather than chasing a 9-month transformation roadmap.
Build for security first. If data governance is an afterthought, you will never gain the organizational trust needed to scale.
Measure business impact. Does the tool improve time-to-market? Does it reduce errors? If you can’t answer those questions, you just have an expensive toy.
Treat it as organizational change. AI is not a technology problem to be solved solely by your tech team. It requires alignment on processes, culture, and goals.

Looking Ahead

We are moving away from monolithic models and toward composed systems. Tomorrow’s agents will use a heavy reasoning model for complex analysis, hand off routine tasks to a cheaper, faster model, and keep sensitive data entirely local.

The technology is now good enough and cheap enough that access is no longer the differentiator. Execution is. Can you integrate this into your legacy systems? Can you measure its real impact?

For those of us building in LATAM, our history of working within limitations is our greatest asset. The AI transformation is already here. The only question is whether you will adopt it thoughtfully or chaotically. Choose wisely.