Enterprise adoption of AI translation has accelerated sharply. Recent industry reporting documents millions saved in translation costs, turnaround times halved, and on-time delivery rates approaching 99.7%. For organisations translating hotel listings, product descriptions, help centre articles and interface text across dozens of languages, these results are credible and reproducible.
They are also domain-specific. And that distinction is consequential.
The Training Data Problem
Large language models learn from published text. Performance scales with the volume and quality of domain-specific material available during training.
Hotel listings, e-commerce copy and marketing descriptions exist in enormous quantities online across multiple languages. AI systems trained on this material perform accordingly.
The picture changes in specialist domains. Shipbuilding specifications. SOLAS compliance documentation. Cross-border acquisition covenants. Pharmaceutical regulatory submissions to the EMA or FDA. IFRS disclosures in consolidated financial statements.
Publicly available corpora in these fields are comparatively small, highly technical and frequently jurisdiction-specific. Model competence is proportional to training exposure — and models do not signal when that exposure is insufficient. Fluent output is produced regardless, carrying the same surface confidence as a hotel description.
There is a second issue. Specialists write for other specialists, compressing years of contextual knowledge into shorthand that is transparent within a discipline and opaque outside it. A domain-experienced translator recognises that compression and resolves it correctly. An AI system resolves the ambiguity statistically — in whichever direction its training data suggests. That direction may be wrong.
Squat is a well-documented hydrodynamic phenomenon whereby a vessel's effective draught increases with forward speed, particularly in shallow water. It has established Spanish terminology in naval architecture and maritime operations. When queried, Perplexity produced calado de agachada — a term that does not exist in the technical literature. Presented with authoritative references, the system argued its position rather than revising it.
The error is not the primary concern. Confident defence of an incorrect technical term in a compliance or specification document is.
When 0.18% Is Not Good Enough
Recent quality benchmarking of AI translation across close to half a million words recorded an error rate of 0.18%. Measured against the MQM (Multidimensional Quality Metrics) framework — the industry standard scoring methodology — this represents strong performance for high-volume localisation work.
Context determines whether that figure is acceptable.
Apply 0.18% to a pharmaceutical regulatory submission, and a single terminological inconsistency may trigger a request for clarification from a regulatory authority. Apply it to covenant definitions in an acquisition agreement, and one ambiguous rendering can alter liability exposure. Apply it to a vessel's structural specification or a safety compliance document, and the consequences move beyond the contractual.
An error rate that is statistically negligible in marketing localisation becomes legally and operationally material in technical, financial and regulatory translation.
Espejo de popa rendered literally as "mirror" rather than transom is not a stylistic issue. It is a substantive one.
Accuracy in specialist translation is not a percentage. It is a function of contextual risk.
Regulated Sectors Operate Under Different Rules
Pharmaceutical documentation is produced within a regulatory framework where terminology, phrasing and cross-document consistency carry direct approval implications. It is not promotional copy.
Financial reporting is governed by IFRS and jurisdiction-specific disclosure requirements where terminology must align precisely with established legal and accounting standards. It is not lifestyle content.
Legal agreements do not accommodate probabilistic interpretation. A single ambiguous clause can redefine liability.
These are not edge cases in specialist translation practice. They are routine.
What AI Contributes — and What It Cannot Replace
AI is now integrated into professional translation workflows. It accelerates first drafts, improves terminology consistency across long documents, and increases overall throughput. These productivity gains are real and are being realised across the industry.
What AI cannot replicate is domain knowledge accumulated through years of specialist practice. It cannot identify when a source text is internally inconsistent. It cannot recognise when a term appears terminologically correct but conflicts with regulatory usage in a specific jurisdiction. It cannot flag the assumptions embedded in source text that an experienced practitioner would query before proceeding.
It does not pause before delivery to ask a clarifying question.
In high-stakes translation environments, that pause is frequently the difference between efficiency and exposure.
Two Distinct Markets
Enterprise AI translation has defined the volume tier with considerable precision: automation, scale, speed, and cost-per-word optimisation. That segment will consolidate further and prices will continue to fall.
What this clarity at the volume end also produces is clarity at the specialist end.
Organisations in maritime, pharmaceutical, legal and financial sectors are not optimising for the lowest words-per-euro ratio. They are seeking translation accuracy aligned with regulatory, contractual and technical reality — produced by practitioners who understand the domain, not only the language.
In specialist translation, the operative question is not speed.
It is risk.