# The agentic AI naming trap: Why you must move from red oceans to blue > Source: > Published: 2026-08-04 10:15:00+00:00 Ask any founder with a chatbot open: generating names for a product has never been easier. A long list appears in minutes. But this is the great naming illusion of the [AI](https://www.fastcompany.com/section/artificial-intelligence) era: the belief that because generating names is now free, or nearly so, the naming challenge itself has been solved. The opposite is true. Generation was never the hard part. What generative AI has done is shift the bottleneck. Producing a hundred names may now easy, but finding names that are distinctive, legally available, and sonically appealing in every market you intend to enter, in every class of goods you’ll eventually sell, is a much harder task. And few naming waters are more congested right now than those around agentic AI — *assistant*, *copilot*, and *agent*: three blurring terms on a single axis of machine initiative, from answering when asked to pursuing goals on their own. The industry has no settled name for the group; “agentic AI” is the working umbrella, and the shorthand used here. So how crowded are these waters? No one can say exactly. The global technology group Prosus and the venture-data platform Dealroom have mapped more than 1,500 AI agents; River + Wolf’s own census of the category turned up 450-plus names clustered into just eighteen recurring territories shown in the graphic below. A generative model learns language from what has been written — and what has been written about agentic AI is the existing landscape of agentic AI names. Ask one to name your product and it begins fishing in one or more of the 18 territories illustrated here. Inevitably, its suggestions feel right because they resemble what’s already familiar. But resemblance is the raw material of trademark conflict: the AI tool that founders trust to carry them out of crowded waters is the anchor holding them in place. A useful frame here is W. Chan Kim and Renée Mauborgne’s red and blue oceans. They applied it to markets, but it transfers almost perfectly to naming, because a name is strategy compressed into a word. Name in the red ocean and you inherit your competitors’ sameness, collisions, and crowds; name in the blue and you claim space the category has yet to populate. What follows are three ways out of the crowd: two escapes to bluer water — the compound and the coined name, each in several varieties — and one deep cast into the red. Like founders in many industries, AI founders favor short, single English words (Glean, Dot, Pod), which is exactly why those words are largely exhausted. One way to get beyond this style of naming is to fuse two ordinary words into a distinct whole. The fusion can join two distant words, as in Moveworks, the enterprise AI assistant that resolves employees’ IT and HR requests on its own. You can also create a compound in which an adjective modifies a noun as with TinyFish, an enterprise AI platform that deploys autonomous web agents to navigate, research, and extract data from the live internet. The craft lies in the seam: the compound must feel inevitable rather than welded together. Also avoid building on overused category parts: *mind*, *bot*, *chat*. Trademark examiners often deem such parts “merely descriptive,” registrable only with a disclaimer that leaves that half of the name unprotected. Even OpenAI was refused a trademark on GPT for exactly this reason. A compound can also cross tongues, welding a word that is archaic, or little known to the target market, to a familiar one. Wraithwatch, the AI cyber-defense firm contracted to defend U.S. federal networks, fuses a 16th-century Scots word for ghost to plain English. The old word supplies depth, the living word familiarity, and the seam alchemizes the two into a name that feels distinct and approachable, with potentially less trademark risk. Remaking a sound entirely, though, is a craft of its own — which brings us to coined names, also known in trademark parlance as fanciful marks. *Freshly Minted* The bluest water belongs to the fully invented name, and its coins are minted two ways. The first is pure abstraction. Kodak meant nothing before Kodak — and that was its strength: pure sound, free to stand for anything the company would ever do. But such names are not easy; the phonetics must carry what the dictionary does not. If your promise is ease — the core promise of most agentic AI — choose sounds that flow: the liquid letters *l* and *r*, open vowels like the *a* in *calm*, and the soft sibilance of *s*. A name built on these sounds suggests ease before the customer knows what the product does. Still, abstraction is open territory for a reason: a name built from nothing familiar is hard to warm to and filling that emptiness has historically taken media budgets few challengers can afford. Hence the better bet is often the other side of the coin — a coinage that carries a trace of an existing word within it. *Minted with Meaning* True inventions from nothing — what linguists studying word formation call “root creations” or, more evocatively, “creations *ex nihilo*” — are rarer than the naming industry admits; most coinages are built from fragments of existing language. At River + Wolf we call these trace names. Nvidia is the classic case: the founders’ internal NV file label sent them hunting for a real word to inhabit, and they landed on the Latin *invidia* — envy — a meaning the name faintly radiates and their logo underscores. Gradium, which builds the voice layer for AI agents, works this way: the *grad-* of *gradient*, machine learning’s technical heartbeat, softened by the elemental *-ium* that makes the name sound mined, like titanium or uranium, rather than manufactured. Such names clear the distinctiveness bar while sparing customers the coolness of total abstraction. A trace name can also be clipped from other languages, especially the old myths, whose characters often anticipate today’s products. Many words from archaic languages resist the English tongue, but with intelligent clipping they can sometimes work. Ratatoskr, the name of the Norse squirrel who carries messages up and down the world-tree, is an unwieldy word to both eye and ear — but it clips handsomely to Tosk: clean, blunt, one syllable. Clip with care, though, because the shears can cut. *Tosk* is Norwegian for fool. If you use this approach, remember that every clip must survive linguistic hooks as well as trademark ones. But whether a pure-sound invention or one carrying a trace of an existing word, familiar or less so, every invented name faces the same hazard: land too close to an existing mark — in sound, appearance, or sense — and trademark trouble follows. The defense is built in at the mint: when a name runs near a soundalike or lookalike, its parts must pull toward a different meaning and a different market channel. Plaud, the AI note-taking recorder, sits one letter from Anthropic’s Claude, yet the two coexist on the trademark register — and not for lack of vigilance. Anthropic has shown it will act: in January, it pressed the viral open-source project Clawdbot to rename after its mascot borrowed the name of Anthropic’s own Clawd. That Plaud remains untouched suggests something else is at work — the names diverge where it counts. Claude is a French first name; Plaud reads as a clip of “applaud.” Examiners may also have reasoned that a recorder that employs AI is not the AI assistant itself. The divergence, it seems, did the work — so that no lawsuit had to. An additional escape requires no new territory at all: stay in the red water but cast deeper — even the reddest territories can sometimes turn blue at depth. Greek mythology, for example, is exhausted in the shallows; every chatbot offers the same tired Olympians, and the marquee names of other pantheons are going fast. But farther down lie treasures. When Apple built an internal system for making machine-learning models run swiftly on its devices, its team reached past the god Mercury and named the tool Talaria — the winged sandals that bore him along. A name about speed and delivery, and liquid on the tongue. In agentic AI, Arm’s Metis reaches further back — to one of the oldest figures in Greek myth. Metis was a Titaness, a generation older than the Olympians everyone names their chatbots after. Her name doubles as a common Greek noun: *mêtis* — cunning, wily, an adaptive intelligence. For a security framework, that’s pitch perfect. But however a name is found or formed, every mark faces the same legal challenges. Happily, in less-fished waters more names can surmount these challenges, and what survives may be more distinctive than the least-blocked version of what everyone else is already using. The lesson for the fleet of thousands searching crowded waters for agentic AI names, or any names, is simple: stop fishing where everyone is anchored. Fresh territories for naming still exist, but only for those willing to leave the red ocean behind — or, if they stay, to drop their hooks many fathoms below.