Before AI, nothing had a "moat". Not even a "caveat".
No product had a "north star", a "flywheel" or a "wedge".
Then AI arrived, and suddenly every founder had a thesis, every meeting had a takeaway, and every product had a "surface area". All written by AI, and unfortunately most of it not even proof-read by a human.
Meanwhile, the product is still a chatbot with a landing page.
And every roadmap now needs "agentic workflows", "context engineering", "guardrails" and "evals", because apparently shipping software never included anything like that before. All of it, it seems, was discovered by AI.
A short history of words AI "discovered"
None of this vocabulary is new. Most of it is older than the people now using it in pitch decks.
- Moat. Warren Buffett has been writing about businesses protected by a "moat" in his Berkshire Hathaway shareholder letters since the 1980s.
- Flywheel. Jim Collins described the flywheel in Good to Great, and Amazon was already mapping its own flywheel in 2001, coming out of the dot-com bust.
- Attack surface. Microsoft's security team was publishing how to measure and reduce a product's attack surface in 2003.
- Caveat. Latin. It has been warning buyers for a few centuries.
The newer AI buzzwords are mostly old practices with a new label:
| The AI-era word | What we called it before |
|---|---|
| Evals | Tests, acceptance criteria, quality assurance |
| Guardrails | Validation, permissions, limits, error handling |
| Context engineering | Documentation, requirements, a clear brief |
| Agentic workflows | Automation, with a person checking the result |
| Surface area | Scope, or attack surface if we meant security |
The AI versions are not identical. Testing a model is harder than testing a function, and an agent that acts on its own needs tighter limits than a script. But the discipline underneath is the same discipline good engineering teams have practised for decades.
The problem is not the words, it is the missing thinking
I have nothing against vocabulary. A good term saves a paragraph.
The problem starts when the word replaces the thinking. "We have a moat" is easy to write. Explaining what exactly a competitor could not copy, and for how long, is not. When a deck is full of moats and flywheels but cannot answer that question, the words are doing the work the strategy should be doing.
And when the text was generated by AI and nobody read it before sending, it shows. Investors, customers and candidates read a lot of these documents now. They recognise the same phrases in the same order, and they draw a conclusion about how much thought went into the business behind them.
Meanwhile, the product is still a chatbot with a landing page
A new vocabulary can make a thin product sound substantial. It cannot make it substantial.
"Guardrails" and "evals" only mean something if the team actually tests the model on real cases and knows how often it gets them wrong. "Context engineering" only works if there is documentation worth retrieving. An "agentic workflow" without anyone checking its output is just automation that fails quietly.
The companies that do this well rarely need the buzzwords. They can show the test results, the documentation and the numbers.
A translation test for AI jargon
Next time one of these words comes up in a meeting or a deck, ask the plain question behind it:
| When someone says | Ask |
|---|---|
| Moat | What would it cost a competitor to copy this in six months? |
| Flywheel | Which step feeds the next, and what are the numbers at each step? |
| North star | Which single number, who owns it, and what did it do last quarter? |
| Guardrails | What exactly can the system not do, and how do we test that? |
| Evals | What is the success rate on real cases, and who wrote the cases? |
| Context engineering | Is the documentation written, current and owned? |
| Agentic workflow | Which steps run without a person, and who checks the result? |
If the answer is clear, the word was useful. If the answer is another buzzword, you have learned something too.
Use the words. Just mean them
There is nothing wrong with saying moat, flywheel or guardrails. They are useful shorthand, which is why they lasted long before AI.
Just make sure there is something behind them. And read it before you send it.