The Physical World Doesn't Care About Your AI Demo

There is a massive gulf between a polished demo on a Tuesday and a live production system on a Friday.

I’ve spent the better part of the last decade building and scaling teams across Singapore and the wider APAC region, from the early days of ShopBack to enterprise transformations at GoPomelo. If there is one constant in software, it is that the physical world—and legacy enterprise infrastructure—does not care about your elegant code. It will break your system.

We are currently in a cycle where the noise around AI is deafening, but the signal is incredibly narrow. When you look at what is actually being deployed this week, a clear picture emerges of where the real bottlenecks lie.

The Physics of Retail Automation

Everyone wants to automate the physical world. It sounds great in a board deck.

But a recent AI-driven store experiment shows gaps in automation and management. You cannot simply hang a cluster of vision models from the ceiling, declare the store autonomous, and fire your shift managers. The physical world is entirely composed of edge cases. A customer puts a cold item on a warm shelf. A spill blocks a sensor. Inventory arrives damaged.

Software loops are predictable; human behavior in a retail environment is not. When the AI fails to account for these physical discrepancies, the burden falls entirely on the remaining management layer—who are now managing a brittle system instead of managing people. The hardest part of automation isn't the inference—it is the operational fallback when the model inevitably hallucinates physical state.

If your deployment strategy assumes a frictionless environment, you are building a toy, not a business.

The Infrastructure of the Agentic Web

Move away from the physical world and back into the cloud, and the problems shift from physical edge cases to digital plumbing.

We have moved past generic chatbots. We are now building agents that need to take action. The Model Context Protocol (MCP) has become the standard for this, and operators are rushing to build AI agent tools with MCP to connect LLMs to local data. But building a single tool is a demo. Running hundreds of them across a fragmented enterprise is a nightmare.

How many times have we seen a pristine demo completely shatter when it hits a legacy enterprise IAM setup?

This is exactly why we are seeing a sudden surge in the MCP gateway market to solve auth sprawl and tool routing at scale. When you have dozens of agents trying to ping internal databases, CRM APIs, and HR systems, authentication becomes your primary point of failure. Identity management for autonomous agents is fundamentally different from identity management for human employees.

You cannot let an agent run wild with super-admin credentials, but you also cannot have it constantly failing to execute because of expired tokens. The real enterprise moat in 2026 isn't having the smartest model. It is having the infrastructure to route context securely. If your agents cannot reliably and securely access your proprietary data, they are just expensive typing assistants.

Narrow, Deep, and Agentic

Time is the ultimate constraint. You have a finite amount of it to allocate across your career, your family, and your finances. Software should buy you time.

Generic enterprise chatbots do not buy you time. They just change the interface by which you waste it. The deployments that actually matter right now are hyper-vertical. Take finance. Trintech is advancing the financial close with agentic AI built specifically for finance.

They are not selling a copilot to help accountants write better emails. They are deploying agentic systems designed to handle the grueling, high-stakes reconciliation work of closing the books. This is where compounding returns happen. If an agent can reconcile ledgers autonomously—and flag only the anomalies for human review—you have fundamentally changed the unit economics of a finance department.

You don't need a superintelligence to close the books. You need a highly restricted, domain-specific agent that understands the exact schema of your ERP and the regulatory rules of your jurisdiction.

Stop chasing the mirage of a universal AI that can run your entire company. The operators who will win this cycle are the ones doing the unglamorous work. They are fixing their data pipelines, wrestling with agent authentication sprawl, and deploying narrow models to solve deeply boring, highly expensive problems.

The future doesn't belong to the best prompters. It belongs to the best plumbers.