Human-validated AI translation: how to offer workflows clients actually trust

Clients started asking for human-validated AI translation about two years ago, and most of them still cannot tell you what they mean by it. They know they do not want raw machine output on a signed contract. They also know they are not paying 2019 per-word rates. Somewhere between those two positions sits a service, and the agencies that have written it down are taking work from the ones still putting "we use AI responsibly" on a capabilities slide. We talk to agency owners about this constantly. The ones losing these deals rarely have weaker linguists. They have never turned the promise into a procedure a buyer can read.

So this is about the procedure. What the client is actually worried about, what to put in writing, how to price the review step so it does not eat your margin, and when to tell a client this is not the service they need.

What clients mean when they ask for human-validated AI translation

Two different worries hide behind the same request.

The first is legal exposure. A procurement officer at a manufacturer does not care about BLEU scores. She cares that if a torque specification in a translated maintenance manual is wrong and a technician gets hurt, there is a named person who signed off on that document. Her legal team has told her that "the vendor used AI" is not an answer she can give in a deposition.

The second worry is quieter and far more common: embarrassment. A regional sales director will read the translated deck in front of a customer. If a product name has been translated instead of kept in English, he looks unprepared. Nobody gets sued. Someone gets a difficult email, and the account manager remembers it at renewal time.

Those two buyers want different things from you. The liability buyer wants traceability: who reviewed what, against what source, on what date. The embarrassment buyer wants domain familiarity and consistent terminology, and will usually accept sampled review if the reviewer knows the product line.

Agencies lose bids by answering the wrong one. One agency we know sent a four-page methodology document, process diagrams included, to a buyer whose actual question was whether the same reviewer would handle all twelve monthly batches. The answer was yes, and it was on page three. The buyer signed with a smaller competitor who put it in the first line of an email.

Ask which worry you are solving before you write the proposal. One question usually settles it: what happens inside your organisation if this document has an error in it? The two kinds of answer sound nothing alike. One gets you a traceability paragraph and a signature block. The other gets you a named reviewer who already knows the product line.

Name the review step, not the adjective

"Human-validated" means nothing until you say who looked at what. A reviewer who skimmed 10% of the target text against nothing at all is human validation. So is a bilingual reviser who checked every segment against the source and logged every change. Clients cannot tell those two apart from the outside, which is precisely why some vendors sell the first and describe the second.

The industry already has the vocabulary for this. ISO 18587 covers post-editing of machine translation output and distinguishes light post-editing, aimed at comprehensible and accurate text, from full post-editing, aimed at output comparable to human translation. If you have never mapped your own tiers onto that distinction, the exercise takes an afternoon and makes your quotes much easier to defend. We went through the practical differences in light vs full post-editing.

ISO 17100 is the other standard buyers cite, usually without knowing that its scope excludes workflows where the first draft comes from machine translation and is then post-edited. That mismatch causes real trouble in RFPs. A buyer asks for 17100 compliance on a job you intend to run as MTPE, and if you answer yes without qualifying it, you have made a claim the standard itself does not support. The better answer describes your revision step, names the standard that does cover it, and offers a human-first tier for the documents where the buyer genuinely needs 17100. Whether the certification pays for itself at small scale is a separate question, and we looked at it in is ISO 17100 certification worth it.

Write the definition out, one paragraph per tier. Who performs the review. What they compare the text against. What share of the text they read. Which error categories they are accountable for. What the client receives at the end. If you cannot write that paragraph for a tier you are quoting, the tier does not really exist yet; it is a word on a price list.

Two tiers is usually enough. We have seen agencies build four, then spend every scoping call explaining the difference between the middle two.

What belongs on a human validated translation certificate

A human validated translation certificate is worth exactly as much as its specificity. We have read ones that say "this AI-assisted translation has been checked for accuracy by a qualified linguist." In a dispute, that sentence protects nobody, and that includes you.

A usable certificate identifies the file rather than the project: filename, word or page count, language pair, and a document date or version so there is no argument later about which file was reviewed. It describes the process in plain language, machine translation included, because a client who discovers that later stops believing the rest of the document. It names the reviewer and states the credential that makes them the right reviewer for this text. It states the scope without euphemism: full bilingual revision against the source, or sampled review at a stated percentage. Then it says what is not being claimed, because no certificate should imply notarisation, sworn status, or legal validity unless you actually hold it.

One agency we spoke to added a single line to its template: "Terminology was checked against the client glossary dated [date]." That line did more for their renewals than the ISO logo in the footer. It gave the client's in-country reviewer something concrete to check against, which turned vague arguments about quality into specific conversations about the glossary. Those conversations are cheap to have. The vague ones are the ones that resurface in a renewal meeting six months later.

Keep the certificate to one page. Anything longer reads as hedging, and buyers skim it regardless. The two lines they do read are the reviewer's name and the scope sentence, so those should be the hardest-working sentences on the page. Put them near the top, not in a footer after the methodology.

Pricing the review tier without eating your margin

Here is the uncomfortable part. Review effort does not track word count. It tracks source quality, domain, and how well terminology was settled before anything got translated. Per-word pricing on review is convenient for quoting and routinely wrong by a factor of three in both directions.

A 12,000-word supplier agreement we heard about came back from the AI step in decent shape with one problem repeated forty times: a defined contractual term rendered two different ways. Fixing it took a reviser twenty-five minutes, one terminology decision and a careful find-and-replace. Priced per word at a revision rate, that job had been quoted for roughly four hours. The agency made money and nobody noticed why.

The reverse case is the one that hurts. A 6,000-word set of technical bulletins, originally scanned and run through OCR, produced output that read beautifully and had quietly dropped a negation in a safety instruction. Fluent output is harder to review than clumsy output, because nothing on the page tells the reviser where to slow down. That file needed full bilingual reading at about 1,200 words an hour. At a post-editing per-word rate it lost money, and the reviewer who caught the negation was the only reason it did not cost far more than that.

The fix is unexciting and it works: measure. On your next ten jobs, log revision minutes per thousand words next to the domain, the source format, and whether a glossary existed before translation started. Inside two months you have multipliers built from your own production rather than a rate card someone posted in a forum. Quote per word if that is what the client needs, but derive the rate from measured minutes, and keep an hourly fallback clause for files where the source turns out worse than the sample promised.

One caveat, and it matters. This only works if revisers log honestly, which means not treating a slow job as a performance problem. We have watched a measurement effort produce beautifully clean and completely useless data for exactly that reason.

Where trust gets built, and the two ways it breaks

Trust does not come from the quality of your output. It comes from the match between what you said you would do and what the client can verify you did.

The first failure mode is selling full review and delivering sampling. It survives for a while, sometimes years. Then a client's in-country reviewer reads an entire file, finds an error pattern in a section nobody sampled, and asks a completely reasonable question. There is no recovering that account, and there is no telling of the story where you come out looking merely inefficient rather than dishonest.

The second failure is subtler and much more common. An agency wins a program of monthly batches, delivers each one competently, and never sets up a shared glossary or a named reviewer. By batch four the client notices that the same component appears under three different names across the set, concludes that the AI is unreliable, and asks for a quote on full human translation. The translation step was never the problem. Nobody owned consistency across the program.

Both are process failures rather than talent failures, which is good news, because process is cheaper to fix. Assign a named reviewer per account and tell the client their name. Keep one glossary per client instead of one per project, and update it after each batch rather than at the start of the next. When a term decision changes, say so, so that a deliberate improvement does not read as drift.

If you want one diagnostic question, use this: could the client audit your claim without your help? If the answer is no, you are asking to be trusted rather than showing your work, and that holds up only until the first careful reviewer arrives. Handing a client the means to check you feels exposed the first time you do it. In our experience it is also the fastest way to stop being asked for a discount every quarter.

When this is the wrong service to sell

Some jobs should not run this way at all, and saying so early is better business than discovering it in week three.

Sworn, certified, and notarised translation for official use often carries form requirements set by a court, a ministry, or the receiving authority, and some jurisdictions restrict who may produce such documents. A review step does not convert an AI-translated file into a sworn translation. Check the receiving authority's rules rather than your own comfort level, and check them for the specific country, because they differ sharply.

Marketing copy that has to persuade is a different craft. A client who needs a headline that works in the target culture needs transcreation and a copywriter, and an accurate validated rendering of the English line will land flat. We have seen this go wrong often enough to suggest splitting the quote: validated translation for the body text, separate creative work for the lines that carry the pitch.

Then two practical blockers. Low-resource language pairs, where output quality varies enough that review stops being review and becomes retranslation at a post-editing rate, which is a quiet way to lose money all year. And confidentiality constraints, where a client's contract or sector rules forbid sending source text to a third-party model at all. In that second case the honest answer is a human-only workflow at human-only pricing. Clients respect that answer more than most agencies expect, particularly the ones whose own legal teams have been asking the same question internally and have not landed anywhere yet.

What to do before your next proposal goes out

Pick one client program you already run and write the tier definition for it. Who reviews, against what, how much of the text, which error categories, and what the certificate will state. One paragraph, this week.

Then three small things. Rewrite your certificate template so the reviewer's name and the scope sentence sit at the top, and delete any phrase you could not defend if someone read it back to you in a dispute. Start logging revision minutes per thousand words on every job that passes through an AI step, with domain and source format recorded beside it. And put your exclusions in writing, sworn translation and transcreation included, so that turning down the wrong job becomes a policy instead of an awkward phone call.

The agencies winning this work are not the ones with the best models. They are the ones whose answer to "who checked this" fits in a sentence and survives being checked.