When we first published this post a decade ago, the language business looked a lot different than it does today. Machine translation existed but most programs still ran on human translation, priced by the word, delivered by specialists who knew their subject matter. The per-word rate was a reasonable proxy for what a project would cost, and our advice reflected what buyers needed to know at the time.
Since then, AI has driven the cost of producing translation down while driving the cost of verifying it up. Expert judgment matters more now than ever. The per-word rate that once anchored purchasing decisions no longer reflects the level of oversight, accountability, or risk management you’re buying.
Here’s what we’d tell a buyer today.
1. Your first cost decision happens before translation starts
Word count endures as the main driver of translation costs, and most of that cost is determined before a project ever reaches a language services provider (LSP). Content written clearly and consistently with reuse in mind costs less to translate and maintain. In an AI-assisted workflow, source quality also directly affects how much reviewer time a project requires, and reviewer time is now one of the most significant expenses in a well-run language program.
2. Not all content needs the same treatment
Translation programs cover a wide range of content, from high-volume internal communications to regulated technical documentation. The more specialized the subject matter, the more a program depends on linguists with domain knowledge and a documented process for validating their work. As AI increases the volume of content a program translates, knowing which content requires that expertise gets harder to ignore.
3. Know what your translation memory, glossaries, and style guides are worth
Translation memory stores previously translated content and applies it to new projects, which reduces cost and improves consistency over time. Glossaries and style guides do the same work for terminology and tone. In an AI-assisted program, they also tell the system what words to use, how to use them, and what the approved translations are. They belong to you. Your provider should be maintaining them actively on your behalf, and if you ever change providers, they should be taken with you.
4. Quality means more than accuracy
The localization industry has spent decades measuring quality by the right words. However, a translation can pass every review and still contain a term that a regulatory body doesn’t recognize, or assumptions about the audience that don’t hold in the target market. When you evaluate a provider, ask how their quality process addresses regulatory terminology and audience fit, not just translation accuracy.
5. Ask specific questions about how AI is being used
AI is a part of most LSP production workflows today, but governance around it is not standard. A well-run program documents where AI is used, maintains a clear policy on whether client content is used to train or fine-tune any model, and keeps audit trails that let you verify what happened to your content at each stage. Human review should be validated and tied to specific content types, not offered as a general quality claim.
6. High volume and specialist oversight are different purchases
AI has driven down translation costs, and a growing number of providers compete primarily on that basis. For the right content, AI is a perfectly reasonable option. For regulated, technical, or brand-critical material, the relevant question is not how cheaply a provider can produce volume but what specialist oversight looks like and what it costs. Make sure you know which one you’re buying.
7. Quality claims are easier to make than to verify
For decades, translation quality meant reviewing output and certifying the process. In an AI-assisted workflow, output volume has increased to the point where reviewing everything is neither practical nor cost-effective. The ISO certification your provider holds tells you they have a process. It doesn’t tell you how that process performs at scale across languages, content types, and AI-assisted production.
8. Switching providers costs more than it used to
Changing LSPs has always involved a transition period, but the scope of what needs to transfer has grown significantly. Your translation memories, glossaries, and terminology assets need to migrate to the new provider and remain usable on arrival. If your current program uses integrations, connectors, or AI-assisted workflows, those need to be reconfigured, which takes time and requires your new provider to understand your old systems. Before you switch, confirm exactly what transfers and who is responsible for quality during the handover.
9. Fixing machine translation can cost more than translating from scratch
Paying an expert to clean up poor machine output often takes longer than producing a fresh translation from the start. If a provider’s linguists spend their time rewriting clumsy automated sentences just to make them readable, you are paying expert linguist or subject matter expert (SME) rates for a slow editing process. The real cost isn’t what you pay to generate the initial words—it’s what you pay for the human hours required to fix them.
10. Price should reflect risk and accountability
A low per-word rate covers the cost of generating text, but it doesn’t cover the cost of verifying that the text is correct. When a quote for high-stakes content carries a premium, that extra charge pays for the additional layers of human validation required to guarantee the work. A transparent pricing model separates these numbers so you know exactly how much of your budget goes toward text production and how much goes toward quality assurance.
Know What You’re Really Buying
Buying localization today is really about buying judgment. The question is no longer just what a partner can translate, but what they can verify and stand behind when content is subject to scrutiny. A per-word rate is a clean number, but it hides the reality of the work. It won’t tell you about the diligence of the review, the health of your language assets, or how a provider manages the AI used on your behalf.
These ideas won’t make every purchasing choice simple. But they’ll help you ask the right questions and know what a good answer looks like.
If you’re evaluating providers or reassessing how your current language program is structured, contact us to start the conversation with Argos.
Argos Multilingual
5 min. read
EUDAMED, the European Database on Medical Devices, is a central part of the EU’s medical device regulatory landscape. This centralized hub houses critical information about medical devices circulating in the EU market, including details about their makers, safety assessments, and compliance records. For medical device manufacturers (MDMs), understanding EUDAMED is more than a regulatory obligation. […]
Argos Multilingual
13 min. read
For two years the industry has obsessed over models, prompts, and fine-tuning, and that work has paid off: content generation is faster, more fluent, and more scalable than ever. But teams keep hitting a ceiling. The output passes every linguistic quality check, yet how it actually performs across markets stays unpredictable, sometimes better than expected, […]
