When the Words Don’t Exist Yet
Why AI governance is harder than communication advice can explain.
In the late Republic, Cicero set himself the project of translating Greek philosophy into Latin. He set out to render Plato and the Stoics in his own language, and he discovered fairly quickly that Latin could not hold the concepts he was trying to import. The vocabulary did not exist. In the Academica he wrote — half-anxiously, half-ambitiously — that he had been forced to manufacture the words. He coined qualitas to render Plato’s poiotes. He coined moralis from mos, the Latin for custom, to render the Greek ethikos. To these he added evidentia, humanitas, quantitas, essentia — by modern accounts, more than two hundred Latin equivalents of Greek philosophical and rhetorical terms, many of which survive in English as quality, moral, individual, vacuum, property, definition, infinity, science. The Romans got their philosophical vocabulary because one man manufactured it in Latin, sentence by sentence, while writing about it.
Cicero did not describe his work as translation. He described it as construction.
Most modern advice for people working at the seam between technical teams and business leadership does not seem to know that Cicero had this problem. The advice treats the cognitive load as a communication issue. Slow down. Find better metaphors. Bring people along. The advice is everywhere, and it isn’t wrong. It is also about two thousand years late, and it misreads the actual work.
The work of sitting between two worlds is not principally about explaining one to the other. It is about building the vocabulary that lets either side describe the territory in the first place. The cognitive load that comes with the work — the exhaustion, the reaching for words that aren’t quite right, the tedious circumlocution because the precise term hasn’t been settled yet — is not evidence of a failure to communicate. It is evidence of construction underway.
The Romans understood the position well enough to give it a god. Janus, two-faced, looked simultaneously at the past and the future, and in some traditions at the inside and the outside, the known and the unknown. He presided over thresholds and transitions. The month of January is his. Janus was not a metaphor for the experience of holding two views at once — he was an acknowledged cosmological role, important enough that the calendar opened with him. The Romans accepted that someone had to sit at the threshold between worlds, and they gave that someone a name.
They did the same for the institutional version of the role. The Latin word pontifex — the title borne by Roman priests, including the Pontifex Maximus — literally means bridge-builder. Pons (bridge) and facere (to make). The role had a name, an office, and a temple. The state recognized that the seam between worlds was load-bearing, and it appointed someone to carry it. The Romans were precise about this, and they were not romantic about it. They simply understood that a bridge does not build itself, and the person who builds it deserves an institutional title.
Move the picture forward two thousand years. AI governance, in any institution serious about AI, sits in the same position Cicero sat in. Two domains that do not natively share a vocabulary.
The technical world has its terms: model weights, retrieval-augmented generation, zero-data retention, prompt injection, agentic loops, fine-tuning, MCP servers, data residency. Each one carries a long technical tail and assumes a body of operational knowledge to be useful.
The legal and business world has its terms: privilege, fiduciary duty, materiality, retention obligations, regulatory exposure, duty of competence, professional secrecy. Each one carries its own long tail, and it assumes a different body of operational knowledge.
Neither vocabulary natively contains the other’s concepts. There is no Latin for poiotes and no business term for prompt injection. The person doing AI governance is, structurally, doing what Cicero did. They are manufacturing the language that one side can use to describe the other. The phrase AI governance itself is a coinage of the last few years, and it isn’t fully settled — different institutions mean different things by it. The vocabulary is being built, in real time, by the people occupying the position.
This is why communication advice misses the work. Communicate better assumes the words already exist and the speaker is choosing the wrong ones. The actual problem is more like Cicero’s. The words don’t exist yet. The practitioner is coining them. They are testing whether guardrail lands, whether AI risk lands, whether agentic system lands, whether the legal team will accept prompt-level controls as a meaningful category. They are doing what Cicero was doing in the Academica, and they are doing it under board-level scrutiny rather than over a quiet correspondence with Atticus.
The cost is structural and ancient
The cognitive load that comes with the position is not a personal weakness, and it is not unique to the modern moment. Cicero complained about it openly. In his letters to Atticus he expresses anxiety about whether Latin can carry the work, whether his coinages will hold, whether the Romans will accept words that did not exist yesterday. Cicero’s anxiety in those letters was diagnostic. He was naming the structural reality of vocabulary construction, and the cost it produced. The work was ambitious, and the ambition came with a cost the work itself created.
What changes when the position is understood this way is mostly the optimization function. The work is generative work. It produces vocabulary that did not exist before, and that vocabulary does work long after any given meeting ends. Cicero’s qualitas persisted for two thousand years. Boards in the present moment are not going to remember any individual conversation about AI governance, but the categories that get coined now — the names for risks, controls, accountability structures, evaluation methods — those will persist. The careful manufacture of the right word is the most durable form of leadership available in liminal positions.
That recognition reframes what the work actually is. The temptation in the position is to optimize for fluency — for speed, for clarity, for the smooth handoff between domains. Fluency is useful, but it is downstream of the actual work. Upstream is the vocabulary. The careful, sometimes tedious, often unfinished construction of words that one side can use to describe the other accurately. Pontifex labour, in the original sense.
What we have not given the role
The Romans did one thing better than the present moment is doing. They named the position and they put it inside an institutional structure. Pontifex was an office. Janus was on the calendar. The role of bridge-builder between worlds was understood to be public, recognized, and load-bearing.
The same recognition has not been given to AI governance. The people doing the work hold a hundred different titles, most of them generic — head of innovation, AI lead, director of emerging technology, occasionally something like responsible AI. None of these names the actual work, which is closer to vocabulary engineer or bridge-builder for a new domain. The cost of the absent name is the same cost Cicero observed in his letters. When the role is unnamed, the work is invisible. The load gets interpreted as personal weakness, and the standard advice — given to invisible work that wasn’t supposed to look like work in the first place — is fluency. Fluency does not build the bridge.
The systemic leadership question — how do you bring others with you — has an older answer than the modern leadership literature gives. You do not bring people with you by communicating better. You bring them with you by manufacturing the language that lets either side describe the territory you stand on. Cicero did it for philosophy. The Romans built the office for it. The work is not new. The institution has not yet caught up to the work.
That is the work behind the work in liminal positions. The slow construction of the words that don’t exist yet.
The Work Behind the Work is the umbrella under which all of this thinking lives — six interdependent capability questions that AI structurally cannot answer for you. Subscribe at andrewlewis.ca.