Enterprise Value Compounds in Rows and Columns

The Tabular Reality Check

Most AI pitches start with a magical chatbot writing poetry or a generative model spinning up a slick marketing image in seconds. I’ve sat through dozens of these vendor demos across APAC, and my first question is always the same: how does this handle our messy, deeply nested relational databases?

Silence usually follows.

We are easily distracted by prose—but enterprise value compounds in rows and columns. The reality is that large language models are fundamentally mismatched for raw, unstructured tabular data. They hallucinate relationships, lose context across wide tables, and fail at basic arithmetic reasoning when the schema gets complex.

That is why SAP's RPT-1: Foundation models crack tabular data is one of the few recent developments that actually matters for operators. When a foundation model successfully parses and computes tabular data natively, you aren't just summarizing text anymore. You are predicting supply chain failures before they hit your P&L. You are identifying margin compression in real-time across thousands of SKUs. Order-of-magnitude improvements in operational efficiency do not come from writing better emails. They come from computing the unsexy, structured data that runs the actual business. If your AI strategy ignores the database layer, it isn't a strategy—it's a toy.

The Agentic Integration Swamp

We have officially moved from the "copilot" phase to the "agent" phase of the hype cycle. The promise is intoxicating—deploy an autonomous agent and watch your headcount costs plummet while output scales infinitely.

But what happens when that agent hits your heavily customized, ten-year-old on-prem ERP? It breaks.

The gap between a slick Twitter demo and a production-grade deployment is measured in integration hours. We are currently seeing the hard reality of this in the automation space. If you look at the documentation around Agent Interoperability use cases, you see where the real bottlenecks lie. It does not matter how smart your agent is if it cannot hand off a task to another agent, or securely authenticate into a legacy API.

Interoperability is the invisible ceiling on AI ROI. If your agents cannot talk to each other and your existing stack, you haven't bought automation—you’ve bought a very expensive silo.

Time is our ultimate constraint. We divide it across three domains: career, family, and finance. Wasting engineering cycles building bespoke bridges between incompatible AI agents is a guaranteed way to burn that time with zero compounding return. Before you buy into an agentic workflow, ask the vendor to show you the error logs when the agent encounters an undocumented API change. Real engineering happens in the error handling, not the happy path.

PR Wars vs. Ground Truth

Look at the broader ecosystem right now. On one end, you have frontier labs playing a high-stakes game of geopolitical PR. We see reports detailing Anthropic's efforts to win the AI image war, pushing narratives about models being "too powerful for the public."

It is brilliant marketing disguised as caution.

But step out of Silicon Valley and into an operator's shoes in Singapore or Jakarta. Do you really think the procurement team cares if a model is theoretically "too powerful"? No. They are looking at automation AI tools priced below $25 and trying to figure out if they can reliably parse invoices without hallucinating a zero onto a vendor payment.

The cognitive dissonance is staggering. The frontier labs are selling existential risk to regulators and investors, while the market is desperately hunting for basic, reliable task automation that doesn't require a PhD to maintain. Do not let the PR narratives of the big labs dictate your internal roadmap. You do not need artificial general intelligence to automate expense reconciliation. You just need a cheap, narrow model that works every single time.

The defining trait of a successful technical leader in 2026 isn't the ability to deploy the smartest model. It is the discipline to deploy the most boring one that actually solves the problem. Stop chasing the demo. Start fixing the data.