How to price translation services in the age of AI

Most translation pricing AI guidance stops in the same place: choose a model, per-word or hourly or project-based, then defend it to the client. After enough quoting conversations with agencies, we think the model is the smaller half of the problem. The harder half is that a word stopped being a stable unit of cost. Two jobs of identical length now land far apart in profitability, and no rate card built on volume will tell you which of the two you are about to lose money on. What follows is about repricing the quote rather than repricing the word.

Where translation pricing AI assumptions break down

The per-word rate was a decent proxy for a long time. Drafting dominated the hours, drafting scaled with length, so length predicted cost well enough to build a business on. Post-editing broke that quietly. Verification does not scale with word count. It scales with how much damage a wrong sentence can do, and with how hard the wrongness is to see.

One agency we work with ran two 8,000-word English-to-German jobs in the same week for the same logistics client. The first was internal onboarding slides, reviewed by a bilingual team lead who signed off in about ninety minutes. The second was a service agreement for the client's legal department. Same length, same language pair, same AI pass. The agreement consumed eleven hours of a senior legal translator's time, because every clause had to be read against the English original and two defined terms had come back rendered inconsistently across sections. On a single per-word rate, the slides carried a healthy margin and the agreement finished near break-even.

That gap is the thing your price list has to see. Nothing in the raw output explains it. Both files came back from the model at roughly comparable quality, and what differed was the cost of being wrong, which set the amount of human attention the deliverable genuinely needed. A rate card that averages across both is cross-subsidising risky work with easy work. That holds up fine until a client starts sending you only contracts, and then the averaging that kept you comfortable for years becomes a structural loss nobody budgeted for. The agency in that example caught it after four months rather than four weeks, because monthly revenue looked entirely normal the whole time.

Price the components, not the word count

A translation quote used to be one line because the work was one activity. It isn't one activity anymore. Five distinct things happen on a document job, they have different cost drivers, and clients follow them separately better than most agencies expect.

Source preparation is the first. Getting to a usable file: a DOCX import that doesn't fight you, tracked changes resolved rather than translated, text extracted from images, a forty-tab workbook narrowed to the sheets that actually need XLSX workbook translation. This gets priced per file or per hour. Never per word, because the word count of a bad file tells you nothing about the cost of fixing it.

Terminology setup is the second: building or updating the glossary before anything is translated, one-time for a new client and light maintenance afterwards. Then the translation pass, which is the part that genuinely got cheap, and should be priced as such. Then verification, where a named human becomes accountable for the output, with cost driven by risk tier rather than length. And finally sign-off, which is the question of whose name goes on the delivery and what you are warranting when it does.

An agency we spoke with this spring moved to this structure for new clients only and kept the old per-word card for its back catalogue. The first visible effect was not revenue. It was that project managers stopped guessing. A request for "the usual" arriving as a scanned PDF now produces a preparation line with a number beside it, and the client either approves it or sends a better file. Previously that work was invisible, unbilled, and done at eleven at night by whoever was still awake.

This works best when your jobs vary a lot in shape. If every file from every client looks the same, the extra quoting effort buys you very little.

Tie the price to the risk the client is buying down

Once the components are separate, verification becomes the line that moves, and the honest driver of verification cost is consequence. We have found three tiers enough for most agencies, and a fourth is usually a sign someone is overthinking it.

Internal comprehension sits at the bottom. Someone needs to know what a supplier email or a status report says. Errors get caught by the reader, who has context and will ask. Light post-editing, a QA report to catch numbers and omissions, no second pass. Customer-facing content sits in the middle: anything a client's customers or staff will read as finished work, where a clumsy sentence costs credibility but not money. Full post-editing by a subject translator, terminology checked against the glossary. Regulated or contractual content sits at the top, where a wrong sentence has a price attached to it. Second-linguist review, every number and defined term verified, and a documented sign-off.

The reason this is easier to sell than a rate increase is that the client is choosing something real. Same source file, three prices, three different deliverables. A procurement manager who balks at a higher per-word rate will often choose the top tier without argument for a safety data sheet, then put the quarterly internal newsletter on the bottom tier and feel clever about it. Both of you come out better.

Where it gets uncomfortable is the middle of the range, and we won't pretend otherwise. Marketing copy for a launch is customer-facing by definition, but a bad adaptation can cost real revenue, and clients rarely want to pay top-tier prices for it. We usually see that resolved by pricing marketing adaptation as its own service with its own rate, outside the tiering, since the work is closer to copywriting than to post-editing. If you try to force it into a risk tier, you will underprice it.

Charge for source-file condition instead of absorbing it

Here is the line item almost nobody bills, and it has grown in the last two years rather than shrunk. When drafting was the expensive part, a messy source file added maybe ten percent to a job. Now that drafting is cheap, file condition can be the single largest cost on the invoice, and it is still being treated as goodwill.

A real case: an equipment manufacturer sent a 61-page maintenance manual for translation into French and Spanish. Quoted per word, two languages, standard turnaround. The file turned out to carry nine pages of part-number tables with mixed numeric and text cells, a custom warning style that collapsed on import, header text sitting inside field codes, and one English text box buried in a grouped drawing object that no automated pass will reach. Preparation and post-assembly cleanup took longer than the post-editing on either language. The agency delivered on time and made almost nothing.

The fix is procedural rather than clever. Quote in two stages: a firm price for the translation and verification components, and a preparation estimate that stays provisional until someone has actually opened the file. Twenty minutes of inspection before the quote goes out is enough to see most of it. Count the tabs in a workbook. Check whether a PPTX presentation translation is going to run into text baked into images on half the slides. Look for tracked changes nobody resolved. Then put a number against what you found, with a short note saying what triggered it.

Clients push back on this less than agencies fear, partly because it gives them something to act on. An operations lead who learns that sending exported DOCX instead of scanned PDFs removes a recurring line from every invoice will usually go and fix the export. Some of them cannot, because the source lives in a system they don't control, and those clients simply pay the preparation line. Either outcome beats absorbing it.

What belongs on the invoice and what does not

There is a temptation, once you have unbundled the work, to show the client everything. We would advise against showing them your cost structure.

Name deliverables and responsibilities on the invoice: preparation, glossary, translation, verification level, sign-off. Do not name the share of the work an AI pass did, do not quote an engine, and do not break out a "machine translation" line against a "human" line. The moment those two numbers sit side by side, you have invited a negotiation about the ratio instead of a conversation about the deliverable, and you will lose it, because the client has no way to judge what the right ratio is and will push the cheap number up every renewal. What you are selling is a document somebody will stand behind. That is the unit.

The related discipline is harder: do not pass the whole efficiency gain through. Agencies that cut prices by the full amount AI saved them spent 2024 and 2025 teaching their clients that translation is a commodity, and several are now trying to walk that back against procurement teams who have the old quotes on file. Passing some of it through is reasonable and often necessary to keep an account. Passing all of it through sets your ceiling permanently.

One practical habit worth copying from an agency we know: they keep an internal column on every quote recording the hours actually spent, separate from the price. Nothing fancy, just a number typed in after delivery. After two quarters they could see which clients and which document types had drifted below their floor, and they renegotiated four accounts with evidence rather than a feeling. Most agencies have this data scattered across timesheets and nobody ever joins it to the quote.

Where per-word pricing is still the honest answer

Component pricing is not a universal upgrade, and plenty of work is still priced best by volume.

High-repetition clients are the clearest case. If a client sends quarterly updates to the same technical documentation and your TM covers seventy percent of it before anyone starts, per-word pricing with match-rate discounts describes the economics accurately and both sides already understand the arithmetic. Replacing that with components would make the quote worse, not better. The same goes for genuinely small jobs: a 400-word certificate does not need five line items, and a flat minimum charge is kinder to everyone.

Procurement systems are the other real constraint. A good number of enterprise and public-sector buyers run portals that accept a rate per word and nothing else. You can sometimes get a preparation line accepted as a separate service code, and sometimes you cannot, in which case the practical move is to blend your tiering into differentiated per-word rates by content type rather than fight the portal. It is less precise. It still beats one rate for everything.

There is also a scale threshold worth naming. Component pricing adds quoting overhead per job, and below roughly fifteen or twenty jobs a month that overhead is a meaningful share of a small team's week. Freelancers in particular are usually better served by a two-tier rate and a firm minimum than by a five-component quote, and our own reading of value-based translation pricing is that it rewards volume of quoting practice as much as it rewards the structure itself.

How to move an existing client onto a new rate card

The structure is the easy part. Migrating an account that has been buying at one rate per word since 2022 is where most of these projects die, usually because someone announces a new price list in an email and waits for the fallout.

Shadow quoting first. For one quarter, quote every job the way you always have and send that, while recording what the component structure would have produced internally. Nobody outside your team sees the second number. At the end of the quarter you will know three things you currently only suspect: which clients were subsidising which, how far off your old rate was on the risky work, and whether your preparation estimates are any good. That last one matters, because the fastest way to lose credibility on a new price list is to quote a preparation line and then blow through it.

Then migrate on an event rather than on a date. A new document type, a new language pair, a contract renewal, a new stakeholder on the client side. Any of those is a natural moment to say the pricing has been restructured, because something actually changed in the relationship. A mid-contract repricing with no trigger reads as an increase hunting for a justification, whatever the structure behind it.

What to say, roughly: the drafting step costs less than it did and the price reflects that, while checking and sign-off are now the larger share of the work, so they are priced separately and you can choose the level. Most clients find this more legible than a flat rate, because it matches what they already believe about AI. What not to say: anything that sounds like you are charging them for the AI. Charge for the deliverable and the accountability.

Grandfather the accounts that will not move. Some clients will stay on the old card indefinitely and that is survivable, as long as you know which ones they are and what they cost you.

What to do with your last twenty invoices

Pull the last twenty delivered jobs. For each, write down the price you charged, the hours actually spent across preparation, post-editing and review, and the risk tier you would assign in hindsight. Then divide price by hours. You are looking for the spread between your best and worst effective hourly rate, and in nearly every set we have looked at it is wider than a factor of three.

Twenty is enough to see the pattern and few enough that you will actually finish the exercise in an afternoon.

The jobs at the bottom of that list are your new price list. Not the average, and not the ones you like quoting. Whatever those jobs have in common, whether it is file condition, risk tier, or one particular client who sends contracts and pays newsletter rates, is the thing your current card cannot see, and it is the first line you should add. If the exercise shows no meaningful spread, your volume pricing is describing your work accurately and you can stop reading here, which is a result worth having too. Before you touch the card, settle how you will handle the post-editing levels themselves, because the price only holds if the MTPE level is decided per project rather than per habit.