Most operators I talk to have bought a few AI tools this year. A drafting assistant. A research summariser. Maybe a transcription tool if they're feeling ambitious. Each one solves a real problem. Each one gets justified on its own merits.

None of them talk to each other. And almost nobody treats that as the actual problem.

This is where I think most expert-led businesses get the AI decision backwards. They frame it as a tooling question: which chatbot, which summariser, which app. It's not a tooling question. It's a pricing and resourcing question, and most owners never get that far because they're still comparing feature lists.

Here's the mechanic that gets missed: in a real operation, the output of one step has to become the input of the next step without a human doing the translation. Research findings feed the draft. The draft feeds the compliance check. The compliance check feeds whatever the client actually reads. If any one of those handoffs requires a person to copy text out of one tool and paste it into another, you haven't built an operation. You've built a set of islands with a human ferry running between them.

The expensive part isn't the subscriptions. It's that the ferry is usually the most senior, most expensive person in the business: the owner, or the one associate who understands how all the pieces fit together. Every time that person copies an output from tool A into tool B, they're doing unpaid systems integration work, and they're doing it manually, every single time, forever, because nobody ever designed the handoff.

I call this the copy-paste tax. It doesn't show up on an invoice. It shows up as the reason nothing compounds. You do good work in January and February and March, and by April you're doing the same good work at the same pace, because the coordination cost never went away. It just got absorbed into someone's calendar.

The fix isn't "buy a platform that does everything." Platforms that promise to do everything usually do most things badly and none of them well, and you're back to the same integration problem one layer up. The fix is deciding, deliberately, which handoffs in your business are worth engineering and which ones are fine to leave manual because the volume doesn't justify the build.

That's a resourcing decision. It belongs in the same conversation as pricing and team design, not in a separate "AI strategy" folder that only the tech-curious person in the business opens.

So here's what I'd actually do this week, and it's not a checklist, it's one exercise: pick the single workflow in your business that runs most often, and trace every point where a human currently moves an output from one system into another by hand. Don't fix it yet. Just write down every handoff. That list is your actual AI roadmap. It's more honest than any tool comparison spreadsheet, because it's built from where your business already loses time, not from where a vendor's marketing wants your attention.

Anthropic's Model Context Protocol (MCP) is worth understanding properly, not because it's new, but because it's trying to solve the exact handoff problem above at the infrastructure level: a standard way for an AI application to connect to external data sources and tools without every integration being bespoke glue code.

My honest take: it's a genuinely useful standard for the "connector" problem, the part where an agent needs to read from your calendar, your database, your file store. It is not a solution to orchestration. Knowing how to fetch context isn't the same as knowing when to hand control from one agent to another, or how to supervise a multi-step process that's gone off the rails. Treat MCP as plumbing, not as the operating system itself. Anyone selling it to you as the whole architecture is selling you the pipe and calling it the house.

Tony