Four Things I've Been Thinking About This Week: Vocabulary Moats, Voice Input, Build vs Buy, and the Death of the Middle

The real advantage in AI building isn't technical skill — it's the accumulated language of knowing how things work.

Four observations from the trenches of AI-accelerated building this week. The moat isn't code — it's vocabulary. Voice changes everything. Build vs buy has flipped. And the middle of the internet is about to get eaten.

I keep a running list of things that are bugging me — patterns I'm noticing, half-formed opinions, stuff that doesn't quite have a home yet. Every now and then, enough of them cluster together that they're worth writing down.

This week, four things.

TL;DR

The real advantage in AI building isn't knowing how to code or how to prompt — it's having the vocabulary to describe what you want precisely. Voice input is massively underrated and changes the quality of what you build. The build vs buy calculation has quietly flipped for solo operators. And MCP is about to do to the middle of the internet what the internet did to the middle of the high street.

---

1. The moat is vocabulary, not code

Everyone keeps saying "domain expertise is the new moat." I agree, but I think people are getting the mechanism wrong.

It's not that you need to understand your industry to build useful software. That's obvious and not very interesting. The actual mechanism is more specific than that: the moat is vocabulary.

Here's what I mean. Someone who's spent ten years around product development knows what a sidebar is, what a webhook does, why you'd use an edge function, what "state management" means in context. They don't need to write code — they need to describe things. And the gap between someone who can describe what they want with precision and someone who says "I want this thing to come out from the side of the screen" is the entire gap between a working product and three days of frustration.

This maps to something I keep seeing. According to a comprehensive vibe coding guide from Kumar Gauraw, the skill that determines your results with AI coding tools isn't technical — it's communicative. The ability to describe what you want with precision. That's vocabulary. That's years of accumulated understanding of how systems relate to each other.

A doctor who can say "I need a patient intake form that validates NHS numbers, routes to the correct care pathway based on triage category, and flags safeguarding concerns" will get a better product from AI than a developer who doesn't understand healthcare. Not because the doctor can code. Because the doctor has the words.

This compounds. Every year of experience in a domain adds descriptive precision that AI can act on. It's not about prompting technique or knowing which model to use. It's about whether you can describe the thing you want in enough detail that the machine can build it.

I've been building production software for 18 years now, and honestly, the biggest advantage I have isn't technical skill. It's that I've sat through enough product meetings, enough design reviews, enough architecture discussions, that I know what to ask for. That's the moat. And it's one that people who've spent years in their own industries already have — they just don't realise it yet.

2. Voice input changes everything (and almost nobody is doing it)

I've been dictating my prompts for months now. The difference is hard to overstate.

When you type, you cut corners. You start a detailed description of what you want, get halfway through, think "this is taking ages," and just hit enter with something half-baked. The AI builds something half-baked. You spend 20 minutes trying to fix it. The whole cycle was caused by typing being too slow for how your brain works.

When you speak, you say things you would never type. You describe edge cases. You explain why something matters. You add context that would've taken five minutes to type but takes thirty seconds to say. And the output is dramatically better because of it.

The numbers back this up. Speaking clocks in at roughly 150 words per minute versus 40-80 for typing. That's not a marginal improvement — it's a 3-4x multiplier on input bandwidth. And since the quality of AI output depends entirely on the quality of your input, that multiplier flows straight through to what gets built.

Both Claude Code and OpenAI's Codex shipped native voice input in the same week in early March 2026. That's not a coincidence. The tools are catching up to what early adopters already figured out: the bottleneck in AI-assisted building isn't the model. It's the input. One Medium post put it well — as we shift from writing code to directing AI agents, the most efficient way to write natural language is with your voice.

If you're building with AI and you're still typing every prompt, try dictating for a week. I'm betting you won't go back.

3. Build vs buy has quietly flipped for solo operators

Something I keep running into: people building their own internal tools instead of buying SaaS, and ending up with something better for a fraction of the cost.

I'm not talking about replacing Salesforce at a 500-person company. I'm talking about a solo M&A advisor who built a deal management system for about £500 instead of paying hundreds per month for a CRM that doesn't fit. A mate with a beach building business whose entire operations backend — bookings, quotes, supplier emails, client comms — got built in a day by giving Claude access to a Dropbox full of existing documents.

These aren't edge cases anymore. According to Retool's 2026 Build vs Buy report, 35% of teams have already replaced at least one SaaS tool with something custom-built, and 78% plan to build more this year. The top categories under pressure are workflow automations and internal admin tools — exactly the kind of thing solo operators and small teams are replacing first.

The insight that keeps coming up isn't about cost, though. It's about fit. When you're building a business, you only need the specific bits you need at the specific time you need them. A CRM that costs £200 a month forces process conformity — it makes you work its way. A custom-built system moulds to how the business actually works and grows as your needs grow.

The old build vs buy calculation assumed building was hard and expensive. That assumption broke. For internal tools, personal workflows, and anything where the process is unique to your business, building is now cheaper, faster, and produces something that actually fits.

This is, incidentally, one of the patterns I see most often with service business founders exploring software. The first instinct is always "which tool should I buy?" The better question, increasingly, is "what would I build if building were easy?" Because it is.

4. The middle of the internet is about to get eaten

The internet ate the middle of the high street. Tesco survived (commodity). Mont Blanc survived (luxury). Everything in between — Topshop, Debenhams, BHS — got hollowed out. The internet offered the same thing without the friction, and the middle had no defensible reason to exist.

I think MCP and conversational AI are about to do the same thing to the middle of the internet.

Think about how you use Airbnb. You open the site. You enter dates. You filter by price, location, bedrooms, amenities. You scroll through listings. You click into one, read the reviews, check the map, go back, compare three more. The whole thing is a series of screens, boxes, buttons, and filters.

Now think about what happens when you can just say "I need a flat in Granollers for four people in August that has a pool and is near a train station" and get back three options with pricing. No screens. No filters. No buttons. Just the answer.

I wrote about this dynamic already in SaaS Isn't Dying, The UI Layer Is. But the high street analogy makes it more concrete. As one enterprise analysis from Cribl noted, MCP promises to become something like USB-C for enterprise software — and for SaaS vendors with established user interfaces, that can feel like a trap door opening under their engagement metrics. You're not losing your product. You're losing your UI as the place where value is perceived.

The business logic survives. The data survives. The matching algorithm, the pricing engine, the recommendation system — all of that still matters. What dies is the UI layer sitting in between. The screens, the navigation, the filters, the settings pages.

Now — I should be honest. I've had this conversation with actual developers who push back on this. They argue that browsing is part of the experience. That people want to scroll through pretty pictures of Airbnbs, not just get handed three options. That's a fair point. Live TV still exists for a reason. But I'd counter that the middle of the high street also had an experience — going to Topshop was fun — and it still got eaten.

The things that survive will be commodity (the cheapest, most frictionless option) and luxury (the experience is the point). Everything in between that's basically just a UI on top of a database and some business logic? That's vulnerable.

For service businesses building software products, the implication is clear: don't build another screen-heavy SaaS. Build the business logic layer. Build the methodology. Build the MCP server that lets AI access your expertise directly. That's where the durable value lives.

---

What connects these four things

If I zoom out, the thread running through all four observations is the same: the interface layer is collapsing.

The interface between you and code? Collapsed — you describe what you want and AI builds it. But the quality of that description depends on vocabulary (observation 1), and voice makes that description dramatically better (observation 2).

The interface between you and business tools? Collapsing — you build your own instead of buying someone else's UI (observation 3).

The interface between users and services? About to collapse — MCP means you don't need the screens and buttons anymore (observation 4).

The value is moving away from interfaces and towards the things behind them: domain knowledge, business logic, methodology, data. If you're building anything right now, that's where I'd put my attention.

---