You probably have a few really simple manual tasks that feel ripe for AI automation, but you can’t make progress because of where the data lives or where the work happens.
This is a challenge all businesses will face as they lean into AI: Automation isn’t hard because AI isn’t capable – it’s hard because we’ve built our businesses on decades-old software that only knows how to serve humans.
The good news is that LLMs like Claude are beginning to adapt to this mismatch – and as they add in agentic capabilities, they’re quietly redefining what work will look like when tools are built for machines.
Why simple automation fails
One of my clients wants to automate their invoicing and procurement process – and rightfully so, because it’s manual as hell.
In their current process, a purchase request lands in a system, an approval email goes out via that system, and a human opens that email, downloads the attached PDF, and types the purchase order numbers into another system.
On paper it looks like easy automation – it’s repetitive, rules-based, and should be low-hanging fruit for tools like Power Automate or Copilot Studio.
But in practice, every “simple” step becomes a tangle:
- Each customer portal is a bespoke system
- Permissions to the system vary
- APIs are inconsistent or nonexistent
- Data lives in PDFs, emails, or screens – not structured endpoints
What sounds like straightforward automation quickly turns into weeks of integration work – setting up access, APIs, and custom connectors for systems that were never designed to work together. It starts to feel easier and more defensible to just build something custom.
How Claude Code is solving the automation problem
A year ago, I saw agents as basically an architecture – not a big deal and certainly not a near-term step-change in AI’s capabilities. But Claude Code changed my mind.
Most AI coding tools before Claude Code were basic autocomplete or “comment-to-code” generators. Useful, but not revolutionary – largely because the real work isn’t happening in the editor.
AI needs access to the environment around the editor to have a real impact – to install libraries, run tests, manage files, and configure systems. Access to this environment has historically been outside a model’s reach. Anthropic realized that without access to the terminal, the agent would always be dependent on a human to move the work forward. So they gave it access.
This allows Claude to actually operate a computer, not just generate text. And once that barrier fell, developers could use the Code agent to do real work – wire up environments, debug, iterate, and deploy. The productivity spike was huge, and that was only possible because AI jumped over the human operator.
That’s the blueprint for the future and the way around the automation problem. These agents are the first prototype of systems that have been rewired to give AI control.
Next up to bat: The re-plumbing of the internet
This isn’t just speculation. Google’s Universal Commerce Protocol (UCP) is proof of that.
UCP is an open, standardized way for AI agents and commerce systems to “speak the same language” so they can discover products, manage carts, and complete purchases directly with merchants. In practice, it acts like a universal translator between AI surfaces (such as Gemini or AI Mode in Search) and retailers’ backends, enabling end‑to‑end shopping flows without custom one‑off integrations for each store.
On the surface, it’s about letting shopping agents transact for you. Under the covers, it’s a massive effort for Google to protect its search and shopping revenue in an agent-first world.
Traditional search is becoming more of a fallback when AI fails, and Google recognizes this. UCP is a hedge – a bid to ensure commerce still flows through Google’s channels even when agents do the heavy lifting.
So again, we’re re-wiring infrastructure for machine actors. And this is no small thing. Read the documentation – Google had hundreds of people working on UCP. They see the paradigm shifting and they’re re-plumbing the internet to prepare for it.
What this means for leaders
Today’s automation struggles are a symptom of human-first infrastructure. Claude and Google are demonstrating what happens when that infrastructure is rebuilt for agents, and that shift will define which organizations can scale automation and which remain stuck stitching SaaS tools together.
The replumbing of infrastructure for AI is already starting, and you need to have at least a basic understanding of what this means for how you do business.
Here’s my advice for you:
- Don’t look for short-term solutions in software. Most AI products being sold today are built on top of the same human-first systems you already have. They look impressive in demos, but they inherit all the same constraints – brittle integrations, partial access to data, and outputs you can’t use without human effort. That’s why so many AI initiatives stall after the pilot phase.
- Invest deliberately in internal capability. Whether you call them automation specialists, product-minded operators, or something else, you need people inside your organization who can wire together systems, experiment with low-code and no-code tools, and prototype agent-like workflows without waiting for perfect infrastructure. Not all of what they create will be winners, but the knowledge they’ll gain is far more valuable.
- Settle for adding AI into the narrow, unglamorous problems. The biggest wins right now don’t come from automating entire processes end to end. To get the productivity gains you’re looking for short-term, identify the places where humans are acting as routers – copying data between systems, reconciling information, preparing inputs for decisions – and reduce that friction.
