AI has transformed translation at extraordinary speed. Content that once took weeks now moves in days, sometimes hours. Costs have fallen and volumes have surged.
As translation accelerates, visibility is fading, and enterprises are starting to feel the consequences.
Across global organisations, multilingual content is increasingly produced through AI tools that few people can fully explain, audit or defend. In consumer contexts, that may be acceptable. In enterprise environments governed by ISO 17100 and beyond, regulatory scrutiny or reputational risk, it is not.
Undoubtably speed is valuable, but control is non-negotiable.
The Governance Gap in AI Translation
AI translation is scaling faster than the controls designed to manage it.
In most organisations, source content enters a Translation Management System (TMS). AI and Machine translation (MT) engines generate output. Translation Memories (TMs) are applied. Human post-editing workflows still exist. None of this is new.
What is new is the volume and velocity at which AI-generated content is now produced and reused across large language model (LLM)–driven workflows, often without equivalent advances in quality governance.
Decisions are made continuously across these workflows, yet the rationale behind them is rarely visible or recoverable. Terminology choices, stylistic decisions, and quality thresholds are applied, but accountability is diffuse and difficult to trace.
This is not a failure of AI technology; it is a failure of governance.
When localisation workflows outpace oversight, organisations absorb risk long before they recognise it. By the time issues surface through audit, regulation or reputational impact, the damage is done, the money spent and the time lost.
If You Can’t Audit It, You Can’t Trust It
This is where we’ve seen buyer behaviour change fastest.
Procurement teams are no longer impressed by “AI-enhanced” claims. Compliance teams are asking harder questions. Localisation leaders are being asked to defend decisions they cannot always evidence.
The questions are direct:
- How was this content validated?
- What MQM thresholds were applied?
- Which terminology and style rules were enforced, and where is the proof?
If an organisation cannot trace how TMs were applied, how AI-generated output was post-edited, or where quality gates triggered human intervention, translation is not being actively managed. It is being assumed.
From Black Box to Glass Box
This is where the industry has to change course.
The next phase of AI-powered localisation will not be shaped by smarter models, but by transparent, auditable systems that can be opened up, inspected and defended.
In practice, this means moving beyond standalone MT toward LLM–driven workflows that are context-primed, governed, and orchestrated across the localisation lifecycle.
As AI becomes embedded in translation workflows, organisations must know where their data is processed and whether it ever leaves controlled environments. If those answers are unclear, the model is wrong, no matter how fluent the output looks.
At enterprise level, the conversation moves beyond language quality into data governance and AI provenance. These are no longer secondary considerations, but foundational.
At THG Fluently, our AI operates within ISO-accredited and Cyber Essentials Plus–certified environments. Advanced capability is developed through controlled architectures, including Gemini-based models delivered via THG Ingenuity’s partnership with Google.
It is this architectural distinction that now defines how enterprise localisation is built.
This represents a move toward Glass Box Translation, built on four non-negotiable principles:
- Traceability means every step in the workflow is visible. AI-generated output, TM leverage, fuzzy-match thresholds, post-editing actions and quality interventions are logged and recoverable within the TMS.
- Measurability replaces subjective judgement with Multidimensional Quality Metrics (MQM) applied at multiple stages. Quality is scored, tracked and compared over time, not inferred.
- Governance ensures AI operates within defined constraints. Terminology, style guides and TM function as active linguistic assets that guide both AI-generated content and human decision-making.
- Human accountability ensures linguists are deployed deliberately. Not blanket post-editing, but targeted intervention based on content type, language pair, risk profile and MQM thresholds. The decision is driven by evidence, not habit.
Crucially, a governed localisation model does not apply the same level of intervention everywhere. Some content can be delivered through AI-only workflows validated by MQM. Other content requires targeted human review. The difference is intent, risk and measurable outcomes.
This approach does not slow AI adoption; it makes AI usable at enterprise scale.
MQM Changes the Conversation
When applied properly, MQM is not a scorecard; it is a control framework. Embedded within the Translation Management System, MQM functions as a rule-based quality gate, validating AI-only output where thresholds are met and triggering review, rework or escalation when they are not.
Procurement teams gain benchmarks.
Compliance teams gain audit trails.
Content owners gain confidence.
Meaning localisation stops feeling risky and becomes operational.
Tools Translate. Systems Deliver Trust
AI tools generate translations.
AI systems deliver trusted content at scale.
Enterprise localisation is no longer linear. It is an orchestrated system of Translation Management Systems, LLMs, TMs, linguistic assets and human expertise, governed by measurable quality frameworks.
That is why human-in-the-loop still matters. Not as a default, but as a control point applied where risk, nuance and brand impact demand it.
When applied deliberately, the impact is clear. MQM analysis across THG Fluently’s workflows shows AI-only output closing much of the quality gap, while targeted human-in-the-loop judgement can be applied selectively to push results beyond traditional human benchmarks.
Used properly, AI does not replace linguistic expertise; it amplifies it.
The Question That Makes or Breaks Growth
When translation decisions are challenged by regulators, auditors, partners or customers, can you open the AI box and explain exactly how those decisions were made?
Organisations that can do this do more than reduce risk.
They move faster with confidence, launch globally without hesitation, and scale content without friction.
In the age of AI, trust is not a safeguard. It is the price of admission for scale.