AI can generate content in several languages at once, so why the heck are we still writing everything in English first? And if producing the words is getting cheaper, why are we still charging for them?
Part 1, “Multilingual AI Is Really a Power Question”, looked at whether localization is still an industry or becoming a mere enterprise capability, and who really gets to own multilingual AI. This time, the fault lines run through the workflow and the business model.
Fault line 3: Is the source-then-translate workflow obsolete?
The question is whether this remains the best way to produce multilingual content, or simply the way organizations have learned to approve it.
Position A: The traditional English-first waterfall limits what non-English content could be if set free.
Give multilingual models a structured brief and they can produce usable drafts in many languages. Quality varies, particularly in lower-resource languages, but a polished English article needn’t be the starting point.
An English source brings more than approved facts. Like it or not, it brings an English structure: the order of the argument, the emphasis, the assumptions about what needs explaining, and what can remain unsaid.
Transcreation aside, the translator's job becomes adapting decisions that have already been made. Unless, of course, the source is some bare-bones Simplified English text that neither pleases nor offends anyone.
Start with a shared brief covering the facts, audience, goal, constraints, and claims. Each market then develops its version using local terminology and a tone suited to its readers.
Keeping every market downstream of a frozen English source can mean adding a new capability and then preventing it from changing the workflow.
Position B: The unloved waterfall still solves a coordination problem that parallel generation does not.
But the English sourcesss isn't just content, precious, I hear Gollum hiss. It's the thing Marketing signs off on, Legal reviews, Product checks, and Compliance certifies.
Organizations still need to know what has been approved, which claims may change, and what happens when Product updates a detail after several markets have finished their versions. Parallel generation doesn't make those questions go away. It can make them harder.
The alternative also assumes things that are hard to achieve: a well-defined shared model of products and concepts, reliable market-specific guidance, and enough reviewers to judge versions that are no longer close translations of one approved text.
Without them, an enterprise risks replacing one review chain with many, while losing the version control that kept them connected.
So what’s the disagreement, really?
Positions A and B differ, but they both share an assumption that is perhaps worth questioning: that the thing being approved has to be the English text.
What if it isn't? Approve the brief, not the prose. Legal, Compliance, and Product sign off on the structured intent: the facts, the claims, the constraints, the terms that must not move. Markets then generate freely within those boundaries. When Product changes a detail, it changes in the brief, and every dependent market version gets flagged.
Version control now has to work at the level of claims and concepts instead of sentences, and that infrastructure hardly exists in most organizations today.
The key question is: which decisions genuinely need one approved source, and which are being forced through English simply out of habit?
Fault line 4: Is per-word pricing dead?
Another popular obituary we’ve seen written for well over a decade.
Position A: Per-word pricing no longer reflects the work.
Per-word pricing made sense when a word represented a roughly comparable unit of human effort. In an AI-assisted workflow, it doesn’t. A word generated by a model and (hopefully) reviewed is not the same commercial proposition as a word researched, rewritten, or transcreated by a real human.
For buyers, the frustration is obvious. If automation reduces the work, they expect the bill to change. Charging more simply because more words pass through the system may look seriously uncool.
For providers, cutting the rate is not much of a strategy either, especially if the remaining work involves more judgment, risk, and specialist review; work that doesn't go away.
The proposed replacements include hourly rates, subscriptions, capacity-based arrangements, platform access, and outcome-based pricing.
All of these aim to charge for something closer to the value delivered rather than the length of the text. (Meanwhile, AI vendors keep reshuffling their own pricing between tokens, seats, and credits, so we aren't short of examples of how unstable “value-based” pricing can be.)
Which is fair enough until someone asks what, exactly, they're buying.
Position B: Per-word pricing will survive because the alternatives are hard to specify.
Per-word pricing has a quality its critics sometimes overlook. People understand it. A buyer can estimate a budget. A provider can quote. Procurement can compare bids without first negotiating a new theory of value.
It has also survived disruption before. Translation memory didn't kill the word; it produced fuzzy-match grids. Machine translation didn't kill it either; it produced post-editing rates. Each time, the unit stayed and the price fragmented into tiers.
And the alternatives come with their own awkward questions. A subscription buys access to what, with which limits? Capacity means how much work, at what complexity? An outcome means what, and how much of it can the language provider actually control?
Nobody wants to pay for an outcome that can't be measured. Nobody wants to be paid for one that depends on decisions somebody else makes.
From this perspective, per-word pricing isn't disappearing. It's perhaps shifting into the segments where its predictability remains useful.
So what’s the disagreement, really?
Everyone and their mother agrees that per-word pricing is under pressure. The disagreement is whether AI is just another discount tier, like TM and MT before it, or something different in kind, and whether a workable replacement is ready.
We no longer sell words. We sell outcomes. But which outcomes? Who measures them? Who carries the risk? And what happens when the client changes the brief?
In the beginning was the word. It may also be there at the end.
Here is where I think the two fault lines meet. If each market generates from a brief, you can still count the output words. But what are you paying the provider to do: generate them, check the claims, or take responsibility for the finished version? The workflow changes what the work involves. The pricing debate has to account for that.
The word isn't dead. But it is no longer the only thing holding the industry together, and we should stop pretending it is. And yes, I do feel fine.