AI Word Salad

How to tell a genuinely new idea from an old one in a new coat

"Our agentic platform delivers autonomous skills within a governed loop, with human-in-the-loop guardrails and enterprise-grade orchestration."

Every word in that sentence is real. Several of them mean something. Read together, they mean approximately nothing, and yet some version of it has probably reached your inbox this quarter. Possibly this week.

We are not claiming immunity, by the way. We are claiming we noticed.

The vocabulary arrived faster than any agreement about what it means, which is normal for a young field and inconvenient for anyone trying to buy something. So here is the sorting system we ended up building for ourselves. It has three buckets, and almost every word falls cleanly into one.

Bucket 1: new word, genuinely new thing

Some of this vocabulary exists because there was nothing else to call the thing, like:

• Large language model is a specific object with a specific meaning.
• Prompt, context window, and token describe a machine that did not exist a few years ago, and no older word fits.
• Hallucination is an ugly word for a genuinely new failure mode: confidently wrong, fluently expressed, at a scale and smoothness no previous software could manage. We needed a word for that. It earns its place.

Learn the ones above. They are not just some marketing. They are a useful working vocabulary of real technology, and understanding them well is the difference between really mastering the new thing, or just tagging along.

Bucket 2: new word, old thing, excellent new coat

• Automation. We have had this word around since the Jacquard loom in 1804. It is a perfectly good word. It does not become a different word by being spoken near a language model.
• Skills. Functions, with a job title.
• Assistant. The wizard, returned from the 1990s, considerably better dressed.
• Loop. A 'while' statement wrapped in a new word.
• Orchestration. Workflow, which was itself "business process management", which was itself something else. This one has been renamed roughly every eight years since the invention of the org chart.

None of this is just BS, exactly. Renaming and reframing a known word or narrative is how we, as an industry, adopt and digest change by using familiar terms in new packaging.

But be cautious: if a vendor says its main differentiator is that its software has loops, you have learned nothing about them except that it runs more than once.

Bucket 3: old word, and the meaning quietly moved

This bucket costs people money, and it is worth thinking about.

These are words you already know. You have used them for years. You have a definition in your head, and it is a good definition, but their precise meaning is unclear or not shared:

• "Human in the loop" Sounds like: a person decides. Might mean exactly that. It might also mean a person reviews a sample afterwards, or that a dashboard exists where somebody could, in principle, look. All three get described with the same four words.
• "Guardrails" Sounds like a promise about what a system will never do. Usually describes a setting, which is to say something somebody can change. Guardrails on a road are not a configuration option. Guardrails in software very often are. Worth establishing which kind is on offer.
• "Approval" The sneakiest of the lot, because it sounds unambiguous. Approve what, exactly? The thing the model produced? A summary of it? A confidence score attached to it? "A human approved it" covers all three and distinguishes none of them.
• "Governance" Used to describe a structure of accountability. Increasingly describes a dashboard.

Somewhere in the last two years, the industry began using them for something new: close enough that nobody notices the swap, far enough that you and the others leave the meeting having understood different things.

The question that makes the words irrelevant

You don't need to learn every vendor's private AI language. You need one question that clarifies what they actually mean:

What happens to what your AI produces?

Does it arrive as a suggestion someone retypes, or as an artifact your process already knows how to review, approve, version and audit?

Ask it of every vendor on your shortlist.

Ask it of us. The answers separate quickly, and they separate on something that matters rather than on who has the better marketing words.

A good follow-up is even shorter: what actually runs? Not what was proposed, but what executes.

In a lot of AI tooling, what the model wrote is what runs. That is a legitimate design choice. It is also a completely different one from the alternative, and none of the vocabulary above will tell you which you are looking at.

Notice: next time you read a new AI product launch announcement, do they say what a person still decides?