- Post History
- Subscribe to RSS Feed
- Mark as New
- Mark as Read
- Bookmark
- Subscribe
- Printer Friendly Page
- Report Inappropriate Content
2 hours ago
No one had to convince finance or CFOs that AI would need its own line item when it hit enterprise at scale. The data bears it out: in 2026, 98% of FinOps teams managed AI spend, up from 63% the year before and 31% the year before that.[1] And yet, none of the public research behind that build-out asks the question by language (at least not that I've found). So, when you break AI spend out by model, by team, by product, does language come up?
The Harness Platform did a recent survey of 700 FinOps and engineering leaders across the US, UK, France, Germany, and India, and found that AI spend has outgrown the systems built to track it. Most organizations can say what they paid a given model provider but can't reliably map that spend back to a specific product, team, or business unit[2]. Separately, Flexera's 2026 asset management research found only 31% of organizations have accurate visibility into their AI software at all[3]. If attribution by team or product is still uneven and spending spikes are difficult to pinpoint, how are FinOps folks addressing the potential disparity of cost for multilingual deployments?
In the last article, Language Governance for Enterprise AI: What CTOs Are Missing, I shared data on how varies across languages and why decision-makers in tech need to include a language dimension in their assessments of AI rollouts and platform deployments. Tokenizers carry a real, measured cost premium running anywhere from roughly two to fifteen times more tokens for the same content depending on the model, language, script and other factors, with one studied example in Mayalam costing thirteen times more than English[4]. That premium doesn't show up as its own line anywhere in the reported FinOps frameworks above. It's sitting inside the aggregate number, invisible.
A language premium that adds a few cents to a single chat exchange is easy to ignore. It's a different story once you're running agentic workflows, which already consume five to thirty times more tokens per task than a standard chatbot exchange, according to Gartner's March 2026 analysis[5]. Take a workload that's already 10x more expensive than a chatbot and multiply that by a language that's 5x more expensive to tokenize than English, and the effects can accumulate. That’s not a fluke; it's potentially a second, compounding cost variable sitting on top of a budget line that's already grown by an order of magnitude.
"Agentic models, for example, require between 5-30 times more tokens per task than a standard GenAI chatbot." —Will Sommer, Sr. Director Analyst, Gartner.
The volatility compounds further once agents start talking to each other instead of just to people. Teams running production multi-agent systems already report actual token usage running three to ten times higher than their prototype estimates, largely from loops, retries, and context getting carried forward step to step[6]. Separate research specifically isolating language as a variable in agent failures found exactly this kind of language-driven retry pattern: a large-scale multilingual benchmark of LLM agents found that failure modes shift by language, with languages like Tamil showing the largest share of errors traced to tool misuse — arguments supplied in the wrong language — and a higher rate of unproductive loops than higher-resource languages[7]. It gets retried. Every retry re-sends the accumulated context, and every one of those context windows costs more if the language it's carrying is one that the tokenizers don’t optimize for.
To be fair, “translation costs” as a line-item now branches into two contributing factors. The first factor is the AI token cost — what it costs a model to process and generate content in a given language, the tokenization tax described above. That's an infrastructure-level cost sitting upstream of any translation workflow, and nobody has solved attribution for it yet, Localization Workspace included. The second factor is the cost of translation itself — what you pay a vendor, a workflow, a reviewer to localize a piece of content. That's a different line, and it's the one that's estimable today.
No one has solved for the consistent attribution for multilingual tokenization factor above, but Localization Workspace – the language translation governance center in the AI Platform – was built in part to help estimate costs within that second translation factor. While this language governance suite helps multilingual enterprise with knowledge base translation, audit trails, and AI-driven glossary creation, it can also estimate translation spend by language and region before the bill arrives.
It’s clear that while FinOps teams are trying to figure out the enterprise AI spending landscape while it is rapidly shifting, language seems to be the tracking dimension that shows up last, if at all. Even then, usually only on the translation-vendor side, not the token side. , when configured to manage translation cost estimates alongside other dimensions, will:
- Estimate word counts and associated potential spend with your translation vendors. Check out our Localization Workspace Platform Academy to learn more.
- Automatically display the estimated financial costs in the specific currency of the requester's user session.
- Run inside the same ServiceNow AI Platform instance already governing everything else — no separate system to procure, secure, or reconcile.
- Configure to the translation vendors and workflows already in place, so the visibility doesn't require unwinding contracts you've already negotiated.
For a multilingual enterprise already running ServiceNow, Localization Workspace can be configured against your AI models and translation vendors to estimate translation costs ahead of time. Token costs may stay a mystery until the invoice but your translation costs don't have to.
- Learn more about Localization Workspace on the app store
- Understand how to configure Localization Workspace
- Email our Globalization@servicenow.com team with questions.
[2] https://www.prnewswire.com/news-releases/new-harness-report-reveals-enterprise-ai-spend-has-outgrown...
[3] https://www.flexera.com/blog/ai/ai-budgets-balloon-enterprise-lessons-flexera-2026/
[4]https://www.researchgate.net/publication/371040637_Language_Model_Tokenizers_Introduce_Unfairness_Be... https://www.researchgate.net/publication/410888242_The_Tokenizer_Tax_Quantifying_and_Explaining_the_...
[5] https://www.gartner.com/en/newsroom/press-releases/2026-03-25-gartner-predicts-that-by-2030-performi...
[6]https://www.mindstudio.ai/blog/ai-token-cost-crisis-enterprise
