Artificial Intelligence (AI) is evolving quickly, but for organizations looking to turn its potential into meaningful business value, the bigger questions aren’t necessarily about the technology itself. Where can AI make the greatest impact? How should organizations determine which processes are ready for it? And what does it take to move from experimentation into real-world operations?
We asked three members of the S4A team to weigh in from their respective areas of expertise. From defining the right business outcomes, to rethinking workflows for an agentic future, to navigating the realities of enterprise implementation, their perspectives point to a common theme: successful AI adoption starts with the business, not the tool.
Deep Bhandari, AI & BAIP Advisor
Q: What business outcomes should organizations be focused on when evaluating AI, rather than simply chasing the latest trend?
A: Too many organizations evaluate AI by asking “what can this model do?” The better question is “what decision or outcome gets better, faster, or cheaper if we get this right?” AI tools are converging fast — the differentiator isn’t which model you rent, it’s whether you own the context, data, and judgment that make any model useful inside your business. So the outcomes I push clients to focus on are concrete and measurable: reduced decision latency, fewer costly errors from ambiguous or conflicting data, and faster time-to-value on the processes that actually move the P&L. If an AI initiative can’t point to one of those, it’s a demo — not a strategy.
German Aravena, VP of Product Development
Q: You recently completed some training around Agentic AI put on by the Harvard Data Science Review – what were some of your biggest takeaways?
A: My biggest takeaway was this: we are at a genuine paradigm shift, and it requires us to rethink workflows from the ground up.
The training introduced a framework called AGENT (Audit, Gauge, Engineer, Navigate, Track) that forces you to start with the right question: what is the actual desired outcome of this workflow, and where does the current process fall short of delivering it?
What I found most valuable was the sequencing logic: before deploying any agent, you map your business processes, assess each step for impact, repeatability, and complexity, and then deliberately decide which steps are candidates for agent execution versus human authority. Low-criticality, high-repeatability steps can be handed to agents earlier. High-stakes decisions stay with humans until trust is earned.
The hesitation I often hear is that AI isn’t perfect – it can make mistakes. True. But so do humans. Agentic AI gives humans better, more informed decision-making capabilities: more data, clearer trade-offs, faster analysis. For low-risk, low-criticality work, we can confidently hand tasks to AI. For higher-stakes decisions with real environmental and budgetary impact, humans remain in control, with every decision tracked, audited, and logged. That’s why mapping your processes and assessing risk and criticality upfront is so critical. Together, humans and AI can accomplish a lot.
I’ll admit, it’s easy to get caught up in all the excitement and new announcements. But I think we all have an underlying question: when will this actually matter to us and our customers? What I observed at Dell’s AI Day a couple of months back gave me a sense that we might be approaching that moment. As companies start investing in on-premise LLM infrastructure, we could see a significant wave of agentic AI adoption. When that happens, we need to be ready. That’s why S4A is investing in learning these concepts now, building internal capability, and positioning ourselves (both through services and products) to help our customers move fast once that wave arrives.
Carlos Jallës, AI Engineer
Q: What should organizations keep in mind when moving from AI experimentation to real-world implementation?
A: When organizations move AI from proof-of-concept to real-world operation, the model itself becomes the least of their worries. The real work lies in integration — making AI communicate with existing systems like SAP, respecting compliance rules, approval workflows, and messy master data.
Cost must be treated as a design decision from the start: a flow that’s trivial for a thousand records monthly can wreck a budget at enterprise scale, so model selection (smaller models, retrieval, fine-tuning) is architectural, not just technical.
Traceability is mandatory in regulated, audit-heavy environments — a clear record of inputs, applied rules, and human sign-off, with human-in-the-loop designed from day one.
Finally, AI maintenance differs from software maintenance: systems degrade silently through concept and data drift, demanding continuous monitoring and a defined “health plan.”
The value, ultimately, lives in the bridge between algorithm and operational reality — not the tool.