The email that told Devi her best linguist was leaving arrived at 7:14 on a Monday morning, and it was three sentences long. Devi runs localization for a medical-device company that ships into thirty-one markets, and the linguist who resigned, a German reviewer named Anke, had caught a flipped contraindication in a patient leaflet two years earlier that would have triggered a recall if it had shipped. Anke did not leave for more money. Her note said, in effect, that she had spent nineteen years becoming a translator and now spent her days cleaning up after a machine for a rate that had been cut, on content the engine drafted before she opened the file, with a quota that assumed she would skim. "I did not train for this," she wrote, "and I am too good at it to do it badly, so I am going to do something else." Devi had rolled out an MT-first workflow eight months before. She had hit every throughput target in the business case. She was also, she realized reading that email, six months from losing the exact people whose judgment made the whole program safe. This lesson is about the contract Devi should have written before she rolled anything out: the explicit, honest agreement that reframes the linguist from the machine's cleanup crew into the owner of the quality the machine cannot own, and that keeps your best people instead of quietly driving them away. It is a leadership problem, not a linguistics problem, and it has a right answer.

Why Your Best Linguists Distrust the Rollout

If you are a localization decision-maker rolling out an MT-first program, the first thing to accept is that your best linguists are right to be suspicious, and understanding exactly why is the whole foundation of earning their trust back. Let us define terms first, because the rest of the lesson depends on precision. MT (machine translation) is any system that converts source text to a target language with no human writing the words. MTPE (machine-translation post-editing), sometimes called PE, is the workflow in which a human edits machine output instead of translating from a blank target. An MT-first rollout is the operational decision to pre-populate every segment (the sentence-sized unit of work inside the CAT tool, the computer-assisted translation environment, and the TMS, the translation-management system) with machine output before a linguist ever opens the file. A quality owner is the term this lesson will build toward: the person who is accountable for whether a file is correct and safe to ship, not merely the person who produced the words. The distrust you are facing is not technophobia and it is not resistance to change. It is a rational response to a specific set of signals your rollout sends, whether you intended them or not.

Sit with the linguist's point of view for a moment, because leaders who skip this step build change plans that fail. From where the linguist sits, an MT-first rollout usually arrives as three simultaneous losses, and each one is real.

The Rate Cut That Feels Like a Demotion

The first loss is money, and it is not subtle. MTPE typically prices at 50 to 75% of full human translation, roughly $0.05 to $0.15 per word, with light post-editing landing as low as $0.02 per word. When you switch a linguist from translation to post-editing at those rates, you have, from their side of the desk, announced that the work they do is worth half of what it was worth last year. The industry framing says this is fine because the hybrid workflow lifts throughput from the traditional ceiling of about 2,000 words a day to 5,000 words a day and beyond, so total income can hold. But notice what that framing quietly asks the linguist to accept: that their value is words per hour, and that they should make up a rate cut with volume. To a craftsperson who has spent two decades learning to catch the error that ends a relationship, being told to go faster for less is not a neutral productivity adjustment. It reads as a demotion, and it reads that way because, in the words-per-hour frame, it is one.

The Cleanup-Crew Feeling

The second loss is harder to put on a spreadsheet but it is the one that actually drives resignations. A translator translates: they read a source, they compose a target, they exercise craft, and the finished sentence is theirs. A post-editor working an MT-first file opens a document where every segment is already filled in, fluently, by a machine, and their job is to find and fix what is wrong. That is a different psychological experience even when the hourly pay is identical. It is the difference between building something and inspecting someone else's build for defects. Done all day, on quota, it feels like being demoted from author to janitor, and the feeling is sharpest for exactly your best people, because the better the translator, the more the blank-page composition was the part of the work they valued. You did not intend to tell your senior linguists that their craft no longer matters. But an undesigned MT-first rollout says it anyway, every morning, in the shape of the file they open.

The Risk Transfer Nobody Named

The third loss is the one that should worry you most as a leader, because it is both a fairness problem and a quality problem at once. When a linguist translates from scratch, the risk they carry is the ordinary risk of their own craft. When a linguist post-edits MT-first output on a quota that assumes skimming, they inherit a new and worse risk: the silent critical error, the fluent, grammatical, confident machine rendering that means the opposite of the source, the flipped dosage or the dropped negation or the inverted indemnity clause that reads perfectly and is catastrophically wrong. The engine produces those. The linguist's name is on the delivery. And the faster you push them to make the cut rate work, the more likely they are to skim past exactly the sentence that will trigger the recall. So an undesigned rollout transfers a larger liability onto the linguist while paying them less and telling them to go faster. Your best people see this trade clearly, because they are the ones who understand what a shipped Critical error actually costs. They are not leaving because they fear the machine. They are leaving because the deal, as offered, is a worse trade for them and a worse outcome for you, and they can do the math.

Your best linguists do not distrust MT. They distrust a rollout that cuts their rate, reframes them as cleanup crew, and transfers the machine's most dangerous failure onto their name, all without ever saying so out loud.

Here is the strategic point that turns this from a morale complaint into a leadership priority. The judgment your best linguists carry is the load-bearing control in your entire quality program. The engine drafts fluently and is structurally blind to whether it is correct; the only thing standing between a confident mistranslation and a shipped recall is a skilled human who reads the target against the source and catches it. If your rollout drives those humans out, you have not saved money. You have removed the control that made the speed safe, and kept the speed. That is the most expensive thing a localization leader can do, and it happens quietly, one three-sentence resignation email at a time.

The "AI Drafts, You Own the Quality" Contract

The remedy is not a raise, a pizza party, or a reassuring town hall. It is a contract: an explicit, written, honestly-priced reframing of the role that changes what the linguist is being paid for and what they are accountable for. The slogan is simple enough to put on a wall, "AI drafts, you own the quality," but the value is entirely in making it real rather than aspirational. Let us build the contract term by term, because a vague version of this idea makes things worse, not better. A leader who says "you own the quality" while still paying by edit-distance and measuring by words per hour has just added a burden without changing the deal, and linguists will hear the hypocrisy instantly.

The contract has four clauses, and each one has to be true in the pay, the metrics, and the workflow, not just in the messaging.

Clause One: The Machine Drafts, You Decide

The first clause names the division of labor honestly. The engine produces the first draft. It is fast, it is fluent, and it is not accountable for anything, because it cannot be. The linguist is not competing with that draft on speed and is not being paid to reproduce it faster. The linguist is being paid for the decision that the draft cannot make itself: is this correct against the source, is it on the approved term, is it safe, is it fit to ship. This is a genuine reframe, not a euphemism, and the test of whether you mean it is whether your metrics measure the decision or the typing. If you still rank linguists on throughput and edit-distance, you have not implemented clause one; you have just renamed post-editing. The decision is the product. The draft is raw material the machine happens to provide.

Clause Two: You Are the Quality Owner, With Real Authority

The second clause is the one that separates a real contract from a slogan, and most rollouts skip it. To own the quality, the linguist must actually have the authority that ownership implies, which means the authority to stop a delivery. A quality owner who cannot fail a file is not an owner; they are a rubber stamp who will be blamed when the stamp turns out to have covered a Critical error. So the contract must give the linguist a real gate: the right, under the ISO 5060 error-severity model, to score the output against a formal typology (Critical, Major, and Minor errors across accuracy, terminology, locale, and fluency) and to block delivery when they find a Critical, regardless of how clean the rest of the file looks or how tight the deadline is. This is not a courtesy. It is the whole point. The revised ISO 18587, the standard governing post-editing of machine-translation output and now expanded to cover AI and LLM (large language model) output, insists that the post-editor hold the same linguistic competence as a professional translator precisely because the go/no-go decision is real translation judgment, not a button-push. If you want the linguist to own the quality, you have to hand them the one authority that makes ownership meaningful, and then back them when they use it.

Clause Three: Fair Pay for the Decision, Not the Typing

The third clause is about money, and it has to be concrete or the whole contract collapses into cynicism. Paying MTPE at a flat 50 to 75% of the human rate, priced purely on the assumption that post-editing is faster typing, is the pricing that produces the cleanup-crew feeling and the resignation emails. It pays for the draft the machine already did and treats the judgment as free. The contract prices differently: it pays for the risk tier of the content and the accountability the linguist carries, not the keystrokes. Low-stakes, high-volume content that genuinely needs only light touch can be priced accordingly, and linguists accept that when it is honest. But the high-consequence content, the regulated leaflet, the indemnity clause, the financial disclosure, is priced as what it is: a skilled human taking on real liability to guarantee a file is safe. That work is worth more, not less, than a raw translation, because the stakes are higher and the machine has made the judgment scarcer relative to the flood of fluent drafts it produces. We will build the pricing structure in detail later, but the principle of clause three is non-negotiable: you cannot ask someone to own the quality and pay them as if they were just typing faster.

Clause Four: Transparency About Where AI Is Used

The fourth clause is trust hygiene, and its absence is corrosive in a specific way. Linguists must be told, plainly and in advance, exactly where AI touches the workflow: which content is machine-drafted, which engine, whether an automatic quality-estimation score is routing their attention, whether an LLM is being used to pre-fill terminology, and critically, what the machine is forbidden to touch. Nothing poisons trust faster than a linguist discovering, mid-project, that content they were told was human-first had actually been machine-drafted, or that their edits were being fed back to train an engine they were not told about, or that a "quality score" they were being measured against was a black box. Transparency is not a nicety here; it is the precondition for the linguist being able to do the judgment work well, because you cannot read a draft skeptically if you do not know it is a draft, and you cannot trust a score you are not allowed to understand. The contract says, in writing: here is every place the machine is in this pipeline, here is what it drafts, here is what it must never touch, and here is exactly how your work is measured and used.

The contract is four clauses that must be true in the pay and the workflow, not just the messaging: the machine drafts and you decide, you own the quality with real stop-the-line authority, you are paid for the decision and the risk rather than the typing, and you are told exactly where AI touches the work.

Involve Linguists in Designing the Workflow

A contract handed down from above, however fair, still lands as something done to the linguist rather than with them, and that difference decides whether it earns trust or merely reduces complaints. The single highest-leverage move a localization leader can make in an MT-first rollout is to put the linguists who will run the workflow in the room where the workflow is designed, before it is designed, not after it is decided. This is not consultation theater. It is the recognition that your senior linguists know things about where the engine breaks that no engine vendor, no consultant, and frankly no manager knows, because they are the ones who have spent years catching exactly those failures.

Consider what a good reviewer can tell you that you cannot learn any other way. They can tell you which content types the engine drafts acceptably and which it mangles in this specific language pair. They can tell you that the German output blows the UI's character budget, that the engine reliably swaps the client's approved device name for a common synonym, that dates and negations are where the Criticals hide, that the automatic quality-estimation score is trustworthy on marketing copy and dangerously optimistic on regulated text. That knowledge is the raw material of a good pipeline. A rollout designed without it will route the wrong content to the wrong effort tier, set quotas that force skimming on exactly the files that cannot be skimmed, and measure quality with a metric that misses the failures that matter. And the linguists will watch you build it wrong, having not been asked, and conclude, correctly, that the program does not value what they know.

What Involvement Actually Looks Like

Involvement has to be structural, not symbolic, or it backfires worse than no involvement at all, because a consultation that ignores what it heard is more insulting than no consultation. Make it concrete:

• Let linguists define the risk tiers with you. The classification of content into what the machine may draft, what needs full human post-editing, and what is MT-forbidden should be built with the people who know the consequences of getting it wrong. Their intuition about which leaflet section can kill someone is more reliable than any generic rule.
• Let them set the effort-per-tier and the realistic pace. If a reviewer tells you that regulated content cannot be safely cleared at 5,000 words a day, that is not resistance to productivity; it is the single most valuable risk data you will get, and a quota that ignores it is a decision to ship Criticals. Set the pace to the content, informed by the people who read it.
• Let them choose the checks that go into the gate. The linguists know that numbers, negations, dosages, and obligations are where the silent critical error lives. Build the verification steps around their hard-won pattern knowledge, so the gate catches the failures the engine actually produces rather than the ones a template imagined.
• Give them a standing channel to flag engine failures. A running log where linguists report where the engine drifted, hallucinated, or broke a placeholder is both a trust signal and an operational asset. It tells them their expertise shapes the system, and it gives you the data to improve the engine grounding, the termbase, and the routing.

There is a deeper reason this works beyond the obvious morale benefit. When linguists help design the workflow, the workflow stops being the machine's process that they clean up after and becomes their process that uses a machine. That shift in ownership is the psychological inverse of the cleanup-crew feeling. It is the difference between "the company installed a machine and now I mop up after it" and "we built a pipeline that drafts fast so I can spend my judgment where it matters." Same tools, opposite relationship to the work, and the difference is entirely in who was in the room when it was designed.

Fair Pricing and Workload in Practice

Trust is built or broken in the invoice and the quota more than in any speech, so the strategist has to get the economics of the contract concrete. This is where good intentions most often die, because a leader can genuinely believe in "AI drafts, you own the quality" and still run a pricing model that quietly says the opposite. Let us build a fair structure, and be honest about the tension inside it, because there is a real one and pretending otherwise is its own kind of dishonesty.

Price the Risk Tier, Not a Flat MTPE Discount

The single most important pricing decision is to stop pricing MTPE as one flat discount off the human rate and start pricing it by risk tier. A flat "MTPE is 60% of translation" rate is the pricing that produces resentment, because it applies the same logic to a throwaway marketing string and a contraindication, treating the judgment as equally cheap in both. It is also operationally wrong, because those two files require completely different amounts of human work. The fair and defensible structure prices three things separately:

• Light-touch tier, for high-volume, low-consequence content where the engine drafts well and the human is doing a genuine light pass. Here the MTPE discount is honest, because the work really is lighter. Linguists accept the lower rate on this tier when it is transparently reserved for content that warrants it.
• Full post-editing tier, for content where the engine helps but the human is exercising real translation judgment across the file. This is priced much closer to the full human rate, because the work is closer to full translation, and the ISO 18587 revision's insistence that the post-editor hold full translator competence is the standards-based justification for pricing it that way.
• Human-owned tier, for high-liability and MT-forbidden content, priced at or above the full human rate, because the linguist is taking on real liability and guaranteeing a file is safe. Pricing this tier as if it were a discount off translation is not a saving; it is you underpricing the exact work that protects you from a recall.

The Honest Tension You Must Name

Here is the tension a serious leader has to hold rather than paper over. The whole business case for MT-first rests on capturing cost savings from speed, and the whole trust case rests on paying linguists fairly for judgment. Those two pull against each other, and anyone who tells you they resolve cleanly is selling something. The resolution is not that there is no trade-off; it is that you are honest with the linguists about where the savings genuinely come from and where they do not. The savings are real on the light-touch tier, where the engine truly reduced the work. The savings are not real on the human-owned tier, and pretending they are, by demanding MTPE rates on content that legally requires full human translation, is both a liability and a betrayal of the contract. The fair position, and the defensible one, is: we capture the speed where the speed is real, we pay full value where the judgment is load-bearing, and we never fund the business case by underpaying the work that keeps us out of court. A linguist can trust a leader who says that, because it is true and they can see that it is true.

Workload Is a Quality Control, Not a Throughput Dial

The quota is where the contract is tested every single day, and getting it wrong undoes every fair thing in the pricing. The instinct in an MT-first rollout is to treat the throughput ceiling, 5,000-plus words a day, as the target, and to set quotas against it. On high-consequence content, that instinct is a decision to ship Critical errors, because the pace that makes the numbers work is the pace at which a human skims, and skimming is the exact reading mode that misses the silent critical error. So the workload has to be set by risk tier, not by a uniform productivity target. Light-touch content can move fast because the consequence of a missed error is low. Regulated content moves at the pace of careful reading, because the consequence of a missed error is a recall, and the whole reason you keep the human in the loop is to catch it. Framing the quota this way is also a trust signal: it tells the linguist that you understand the work, that you are not asking them to gamble their name to make a number, and that when they slow down on a leaflet you will back them rather than flag them. The quota, set honestly, is one of the strongest statements a leader can make that the contract is real.

Fair pricing prices the risk tier, not a flat MTPE discount; it captures speed savings only where the engine truly reduced the work; it pays full value for the judgment that keeps you out of court; and it sets the quota by consequence, because a throughput quota on regulated content is a decision to ship Criticals.

A Worked Change Approach That Earns Buy-In

Return to Devi, because principles are only useful if they change what a leader does on a real Monday, and she has an urgent problem: she has already rolled out badly, Anke has resigned, and the rest of her senior pool is watching to see whether Anke was right. Here is the change approach she runs, and it is deliberately concrete, because the difference between a rollout that earns trust and one that hemorrhages talent is almost entirely in the execution details.

Step One: Name the Loss Out Loud

Devi starts by doing the thing most leaders avoid: she acknowledges, to the whole linguist pool, that the initial rollout got it wrong. She names the three losses plainly, the rate cut that felt like a demotion, the cleanup-crew experience, the risk transferred onto their names, and she does not minimize them. This is counterintuitive for a leader mid-rollout, because the instinct is to project confidence and defend the decision. But the linguists already know the losses are real; pretending otherwise only confirms that leadership does not understand the work. Naming the loss out loud is the price of admission to a real conversation. It costs Devi nothing but ego, and it buys her the credibility to propose the fix. A change approach that opens with "everything is going great" to people living the losses daily is dead on arrival.

Step Two: Bring the Linguists Into the Redesign

Next, Devi convenes her three most respected senior reviewers, the ones the others trust, and puts the workflow redesign in their hands, not hers. Together they build the risk tiers, define what the engine may draft and what is MT-forbidden, set the realistic pace per tier, and choose the checks that go into the quality gate. Devi's role shifts from architect to sponsor: she sets the constraints and the goals, and she lets the people who know where the engine breaks build the machine around them. This does two things at once. It produces a better pipeline, because it is grounded in real failure knowledge. And it converts the most influential skeptics into co-authors, which matters enormously, because the senior reviewers the rest of the pool trust are exactly the people whose buy-in propagates. When Anke's peers help design the workflow, the workflow stops being the thing that drove Anke out and becomes the thing they built.

Step Three: Rewrite the Deal in Writing

Devi then puts the contract in writing and makes it specific. Not a values statement, an actual document: here are the three tiers and what each pays, here is your authority to fail a file when you find a Critical and here is my written commitment to back that call against any deadline pressure, here is exactly where AI touches this pipeline and what it is forbidden to touch, here is how your work is measured (by the defensible quality of your decisions, not by edit-distance or words per hour), and here is the confirmation that your edits are not being used to train an engine without your knowledge and consent. Writing it down matters because trust rebuilds on kept promises, and you cannot keep a promise that was never made specific enough to check. A vague "we value quality" cannot be verified and therefore cannot rebuild trust. A written deal with named rates, named authority, and named transparency can be, because the linguist can watch you honor it or fail to, and the watching is how trust returns.

Step Four: Back the First Stopped File, Visibly

The contract becomes real at exactly one moment: the first time a linguist uses their new authority to stop a delivery, and the leader has to be ready for it and get it right. Some weeks in, a reviewer fails a file on a shipping deadline because they found a Critical error in a regulated section. The account manager is furious; the client is waiting; the pressure to overrule the linguist and ship is intense. This is the test of the entire change approach, and it is worth more than every memo Devi has written. She backs the linguist, publicly, and she makes sure the pool sees that she backed them: the file was held, the Critical was fixed, the standard was honored over the deadline. That single visible decision does more to earn trust than a year of messaging, because it proves the authority in clause two is not decorative. The inverse is equally true and equally fast: the first time a leader overrules a linguist's Critical call to hit a deadline, the contract is dead, everyone knows it is dead, and the resignations resume. You get one chance to demonstrate that "you own the quality" means what it says, and it comes when it is least convenient.

Step Five: Show Them the Career Move Is Up

Finally, Devi connects the daily contract to a longer arc, because linguists who see only a fairer version of the same job will eventually still leave for a more interesting one. She shows them that owning the quality is not the ceiling of the new role but the entry to it: the quality owner who runs the error gate can grow into the evaluator who owns the scoring program, the terminologist who builds the termbase the engine is grounded on, the quality lead who designs the tiers, the person who can stand in front of a client or an auditor and say "here is the throughput, here is the risk tier each content type got, here is the ISO 5060 error score with zero Criticals, all defensible under the revised ISO 18587." That sentence is a credential the raw MT vendor can never say, and the linguist who can say it has climbed to ground the engine cannot reach. Showing the pool that the contract is a first step up a real ladder, rather than a gentler version of being stuck, is what turns retention into commitment. Anke left because she saw a demotion with no future. The linguists who stay are the ones who can see the future, and it is the leader's job to make it visible.

The change approach that earns buy-in: name the loss out loud, put the redesign in the linguists' hands, write the new deal down with named rates and named authority, back the first stopped file visibly whatever it costs, and make the move up the value chain a real and visible ladder.

What It Costs You to Get This Wrong

It is worth being clear-eyed about the stakes, because a localization leader under throughput pressure can be tempted to treat linguist trust as a soft concern that competes with the hard numbers in the business case. It is the opposite. Linguist trust is the load-bearing element of the hard numbers, and losing it is one of the most expensive failures available to you, precisely because the cost is hidden until it is catastrophic.

Trace the failure chain. You roll out MT-first without the contract. Your best linguists, the ones whose judgment catches the silent critical error, read the losses correctly and start leaving, quietly, one at a time, for other work. You backfill with less experienced people, or with the linguists who stayed because they had fewer options, working a quota that assumes skimming. The engine keeps producing fluent, confident output that is sometimes catastrophically wrong. The judgment layer that used to catch those errors is now thinner, faster, and less invested. And then one of those errors ships: a flipped dosage in a leaflet, an inverted indemnity in a contract, a wrong figure in a financial disclosure. The recall, the regulatory finding, the lawsuit, the lost account: those costs dwarf the entire savings from the rollout, and they arrive because you removed the control that made the speed safe. "The engine wrote it" is not a defense; the accountability was always human, and the human you needed had already quietly left.

This is why the contract is not a morale program. It is a quality control expressed as an employment relationship. The judgment of your best linguists is the mechanism that turns a fast, risky MT-first pipeline into a fast, safe one, and that mechanism is made of people who have to choose to stay, do the demanding judgment work well, and use their stop-the-line authority when it is inconvenient. You cannot buy that with a rate cut and a quota. You earn it with an honest contract, a workflow they helped design, pay that matches the risk they carry, transparency about where the machine is, and the visible willingness to back their judgment when it costs you a deadline. Get that right and you keep the control that makes your program defensible. Get it wrong and you have bought speed by dismantling the only thing that made speed anything other than a faster path to the error that ends a relationship.

Anke's three-sentence email was not a morale problem. It was an early warning that Devi's quality program was losing its most important component, and she was lucky it came as a resignation she could learn from rather than as a shipped Critical she could not undo. The leaders who thrive in the MT-first era are the ones who read that warning correctly: that the machine made fluent drafts nearly free, which made human judgment the scarce and precious thing, which means the linguists who supply that judgment are not a cost to minimize but the asset the whole program is built on. Treat them that way, in the contract and not just the slogan, and they will own the quality the machine cannot. Treat them as cleanup crew, and they will leave, and the machine will keep drafting confidently wrong sentences with no one skilled enough left to catch them.

Key Takeaways

• Your best linguists distrust MT-first rollouts for rational reasons. An undesigned rollout delivers three real losses at once: a rate cut (MTPE at 50 to 75% of the human rate) that feels like a demotion, the cleanup-crew experience of inspecting a machine's build instead of authoring, and a transfer of the silent critical error's liability onto their name while they are told to go faster for less. That is the worst trade in the business, and skilled people can do the math.
• The remedy is an explicit contract, not a slogan. "AI drafts, you own the quality" has to be true in the pay, the metrics, and the workflow. Its four clauses: the machine drafts and you decide, you own the quality with real stop-the-line authority, you are paid for the decision and the risk rather than the typing, and you are told exactly where AI touches the work.
• Stop-the-line authority is what separates ownership from a rubber stamp. A quality owner who cannot fail a file is not an owner; give the linguist the real, ISO 5060-backed authority to score against Critical, Major, and Minor severities and block delivery on a Critical regardless of deadline, and back that call when they use it.
• Involve linguists in designing the workflow, structurally not symbolically. Your senior reviewers know where the engine breaks in a way no vendor or manager does. Let them define the risk tiers, set the realistic pace per tier, choose the checks in the gate, and log engine failures. This turns "the machine's process I clean up after" into "our process that uses a machine," the psychological inverse of the cleanup-crew feeling.
• Price the risk tier, not a flat MTPE discount. Price light-touch, full post-editing, and human-owned tiers separately. Capture speed savings only where the engine genuinely reduced the work, and pay at or above the full human rate for high-liability and MT-forbidden content, because that is the judgment that keeps you out of court, not a place to find savings.
• Name the honest tension instead of papering over it. The business case wants savings from speed; the trust case wants fair pay for judgment. They pull against each other. The defensible resolution is honesty: capture speed where it is real, pay full value where judgment is load-bearing, and never fund the business case by underpaying the work that prevents a recall.
• Set workload by consequence, because a throughput quota on regulated content is a decision to ship Criticals. The 5,000-plus words-a-day ceiling is the pace of skimming, and skimming misses the silent critical error. Let light-touch content move fast and regulated content move at the pace of careful reading, and back the linguist who slows down on a leaflet.
• The contract is a quality control, not a morale program, and the first stopped file is the whole test. Losing your best linguists removes the control that makes MT-first speed safe, and the resulting shipped Critical error dwarfs the rollout's savings. Back the first inconvenient Critical call visibly, write the deal down with named rates and named authority, and show the move up the value chain, or the resignations resume and the accountability, still human, is on you.