For as long as I have been in this trade, the unit of account has been the word. You count source words, multiply by a rate, and that is the invoice. It is clean, it travels well across languages and tools, and project managers can build quotes in their sleep. The problem is that the word count was never really measuring the thing we sell. It was a proxy for effort, and that proxy has quietly stopped working.
When the bulk of a translator's labor was typing and term-hunting, words and effort tracked each other reasonably well. Two thousand words was roughly twice the work of a thousand. That assumption is now broken in both directions. A machine-translated paragraph can take longer to fix than to write from scratch. A single sentence in a regulatory disclaimer can carry more liability — and demand more thought — than the entire marketing brochure it sits beside. The number of words on the page tells you almost nothing about either.
The proxy collapsed, but the math didn't update
The industry's response to machine translation has mostly been to keep the per-word model and bolt a discount onto it. Post-editing rates are quoted as a fraction of the full human rate — sixty percent, fifty, sometimes less — as if effort scaled linearly with raw output quality. It doesn't. The relationship between machine output and human effort is not a smooth slope; it is a cliff. Above a certain quality threshold the editing is genuinely light. Below it, you are doing forensic work: reverse-engineering what the model misunderstood, then rebuilding the sentence anyway. The per-word post-editing discount averages these two worlds into a single number that is wrong for both.
This hits low-resource languages first and hardest. For a pair with abundant training data, the machine output sits closer to that light-editing zone often enough that a flat discount is merely unfair. For Turkmen, the output frequently lands in the forensic zone, where the morphology is mangled, terminology is invented on the fly, and the suffix chains have to be unpicked before anything else can happen. Charging a post-editing discount on that is not a discount — it is a subsidy I am paying the agency for the privilege of cleaning up a tool that made my job harder. I suspect a lot of colleagues in similar pairs feel this and can't quite name it.
The per-word model also flatters the wrong work. It rewards volume and punishes restraint. The most valuable thing a senior linguist does on a high-stakes file is often to not change something — to recognize that an awkward-looking phrase is correct and load-bearing, and leave it. There is no line item for judgment exercised in stillness. Yet that judgment is precisely what the machine cannot supply and what clients are, increasingly, actually paying for.
Price the consequence, not the keystroke
The alternative I keep coming back to is pricing by risk. Not effort, not output, but the cost of getting it wrong. This is how most mature professional services already work. A lawyer's hour on a routine filing and an hour on a contested merger are not the same hour, and nobody pretends they are. Translation is converging on the same reality, and we should meet it deliberately rather than be dragged.
In practice this means tiering work by consequence. A product description where a clumsy sentence costs a shrug can run through machine translation with a light human pass, priced low and delivered fast — and I have no quarrel with that. A pharmaceutical instruction, a court submission, a safety procedure, a brand's flagship campaign: these are not the same product at a different word count. They are a different product. The risk of error is categorical, and the price should reflect the size of the downside, not the size of the file.
For project managers this is more useful than it sounds, because it gives you a vocabulary for the conversation you are already having internally. When a client pushes back on rate, "it's more words" is a weak answer. "This content carries regulatory exposure and needs a named human on the chain" is a strong one. Risk-tiering lets you charge confidently where it matters and concede gracefully where it doesn't, instead of defending a flat number across both.
What this asks of us
The shift is uncomfortable for translators too, because it requires us to articulate our value in terms other than productivity. If you compete on words per hour, the machine wins and the race is already over. If you compete on the ability to assess where a text is dangerous, to recognize what cannot be automated in your language, and to take professional responsibility for the outcome, the machine is just a tool you supervise.
I don't think the per-word invoice disappears overnight; it is too deeply wired into how agencies operate. But the profession that survives the next decade will be the one that stops selling typing and starts selling judgment under risk. The agencies that figure out how to price that — cleanly, defensibly, on the line item — will keep the clients worth keeping. The ones still slicing percentage points off a post-editing word rate will be competing on a number that no longer means anything.