The Price of a Language Empire: What RWS–Acolad Reveals About the “AI Takeover of Translation”

Two acquisitions, two control points

It is tempting to describe TransPerfect–Unbabel and RWS–Acolad as the same playbook. But I’d say they are better understood as two ways of acquiring control over a value chain whose center of gravity is moving.

TransPerfect primarily acquired capability

Unbabel brought TowerLLM, the COMET evaluation family, proprietary translation technology, customer relationships and a research team with a recognized record in multilingual modeling and machine translation evaluation. TransPerfect already possessed enormous global distribution. What it lacked, or could accelerate by acquiring, was a concentrated body of research, models and evaluation capability.

It bought time. It also bought people who knew how to build. Most importantly, it bought a shortcut to a stronger position in the model and measurement /trust layers.

RWS primarily acquired enterprise conversion capacity

RWS already owns translation technology, content platforms, AI data operations and substantial regulated-industry expertise. Acolad gives it deeper access to Western Europe, particularly France and Germany, as well as public-sector and medical-device relationships, interpreting capability and an enterprise base into which RWS can cross-sell current and future platforms.

It bought the right to convert service relationships into platform relationships.

I’m not talking about an absolute distinction because Unbabel had customers and commercial workflows and Acolad has some technology, data services and expertise. But the primary strategic logic between both acquisitions differs. TransPerfect strengthened the machinery. RWS expanded the surface across which the machinery can operate.

Both transactions answer the same question: what has become scarce? Of course the answer is no longer raw translation capacity. It is control over the points at which enterprise demand, proprietary knowledge, automated production, human judgment and measurable quality meet.

What actually destroyed the valuation

The easiest, commonplace conclusion is that AI destroyed the value of language services. That conclusion could be valid as it contains part of the truth indeed, but misses the mechanism.

Acolad was not built organically. Jourik traces the acceleration of the roll-up to the 2019 transaction through which the du Fraysseix family acquired a 65% stake after buying out Naxicap Partners. The subsequent Amplexor and Ubiqus acquisitions produced a group described at its expansion peak as generating approximately €330 million in revenue, employing more than 2,500 people and operating across 25 countries.

Today’s reported picture is approximately 1,200 employees across 22 countries and £182 million in 2025 revenue, with RWS forecasting an annualized contribution of about £155 million. The historical and current figures are not necessarily like-for-like, and they should not be used as a simplistic measure of jobs or value destroyed. But they do show that the assumptions supporting the expansion-era valuation did not survive.

Roll-ups require the acquired businesses to produce enough growth, margin and cash to justify the price and structure used to assemble them. When price-per-word pressure intensifies, translation volumes move into automated workflows and integration costs remain fixed, leverage stops amplifying growth and starts amplifying delay.

That is where AI enters the story. It did not create every weakness: it changed the time available to correct them.

Technology did not necessarily cause the capital structure to fail. It set the clock against it.

This is also the continuity with Unbabel, Lengoo and other heavily financed language-technology ventures. More capital does not necessarily create a durable position. Capital can fund research, acquisitions and rapid expansion, but it cannot suspend a market repricing. When the basis of competition changes, adaptation matters more than the amount previously invested.

The transition is already visible in RWS’s accounts

RWS’s own financial statements provide unusually clear evidence of the transition.

In FY2025, group revenue fell 4% to £690.1 million. Adjusted profit before tax declined 43% to £60.4 million, and RWS reported a statutory loss before tax of £99.7 million. The accounts included an £88 million goodwill impairment relating to Language Services and Regulated Industries. RWS attributed that impairment partly to the transition affecting core localization, weaker linguistic validation, a shift toward a technology-first proposition and macroeconomic discount rates.

The same report states that traditional translation volumes were declining as clients moved toward machine translation and exerted pricing pressure. It also reports increasing demand for language-quality review and testing in AI-enabled workflows.

Six months later, AI-related products and services represented almost one third (32%) of RWS group revenue, up from 26% a year earlier. TrainAI delivered exceptional growth, driven partly by a large program supporting a strategic client and by new client wins. Group revenue increased to £360.3 million, with organic constant-currency growth of approximately 7% and adjusted profit before tax up 33%.

Read together, those accounts do not say that translation disappeared and AI data replaced it. They say something deeper and far more consequential:

Translation remains a large revenue pool, but AI increasingly determines how that revenue is produced, priced, measured and expanded.

The service has not vanished: the operating logic around the service has changed.

AI is taking over the operating system

The traditional language-services model was organized around a deliverable: a translated document, website, manual, clinical instrument or software release. The commercial unit was commonly the word, page, hour or project. Technology assisted the people producing the deliverable, but the human production chain remained the economic center.

In an AI-mediated model, the deliverable is only one output of a larger system. That system selects or routes content, retrieves terminology and prior translations, chooses models, generates candidate output, estimates quality, applies corrections, sends exceptions to experts, records decisions and learns from the result.

The valuable position is therefore not necessarily the point at which a sentence is translated. It is the position from which the organization can determine:

• which model or engine receives the content;
• which data and terminology condition its behavior;
• which output can move automatically;
• which errors matter in the relevant domain;
• when a human must intervene;
• which decisions become future training or adaptation evidence;
• and who remains accountable when the result causes harm.

This is why RWS can see strategic value in Acolad even while the market assigns a low multiple to the current earnings. Acolad’s installed base can become the distribution and feedback surface for a new operating model. Every enterprise relationship is a potential platform relationship. Every translation memory, terminology decision, correction and quality judgment can become an input into future automated performance, provided the rights, governance and technical architecture allow it.

The per-word service becomes the entry point. The workflow becomes the system of record. The data and decisions produced by the workflow become the compounding asset.

But AI data can become a commodity too

There is an important warning for companies (including us at Pangeanic) that are moving from translation into AI data operations.

RWS’s TrainAI growth was exceptional, but the company also reported that the mix of lower-margin TrainAI activity reduced gross margin. A large part of the first-half outperformance came from one substantial program for a strategic client.

This is not a criticism of TrainAI. It is evidence that “AI data” is not a magic category in which every activity becomes differentiated and highly profitable. At scale, data collection, annotation and human-feedback capacity can face the same pressures that affected general translation: global labor arbitrage, procurement concentration, unit-price compression, episodic programs and increasing automation.

The fastest-growing work can still be commodity work.

The durable layer is not data volume by itself. It is difficult-to-reproduce data with clear rights and provenance; expert decisions in specialized domains; multilingual and regional coverage that models routinely mishandle; evaluation sets designed around real failure modes; repeatable methods; and operational systems that connect evidence to deployment decisions.

This is what AI Data Operations must mean if it is to become a defensible discipline rather than a new label for outsourced labor. It is the continuous work of acquiring, preparing, governing, evaluating and improving the evidence on which production AI depends.

Translation capacity became less scarce. Generic annotation capacity may follow it. The industry should learn the lesson before repeating the cycle under a different name.

Open source does not eliminate the private frontier

The TransPerfect–Unbabel acquisition introduced another structural question: what happens when a major production supplier also acquires one of the industry’s most influential evaluation families?

It is important not to overstate the case. COMET remains an open-source framework that must of us have experimented with. Its code is publicly available, its research is published and widely used checkpoints can be downloaded and reproduced. The COMET family also includes both reference-based evaluation models and reference-free quality-estimation models. TransPerfect stated after the acquisition that it had no present plans to change the operation of the existing open-source versions of COMET or TowerLLM.

That openness (derived from EU grants systems) is important because it permits scrutiny, replication, comparison and continued research. It makes a literal enclosure of the public measurement layer technically inaccurate.

But openness does not eliminate commercial asymmetry: any capable AI company can continue beyond the public release. It can train newer variants on private human judgments, incorporate proprietary client traffic, refine calibration for particular domains, operate larger or more expensive internal models, combine several evaluators, add error-detection layers and improve the orchestration around inference. The public artifact can remain genuinely useful and reproducible while the private performance frontier continues moving.

I am not asserting that TransPerfect has withheld a superior version of COMET or closed the research commons. There is no public evidence on which to make that claim. I am making a more general and durable point: access to a public model does not mean access to the owner’s entire capability.

The governance problem therefore is not simply who owns COMET. It is who controls the entire evaluation regime.

Who marks the machine’s homework?

A risk appears when one supplier selects the production model, operates the workflow, chooses the metric, establishes the threshold and reports that the resulting system has passed. That does not imply misconduct, but it means the customer is receiving a quality claim from the same commercial system whose performance is being judged.

When the system marks its own homework, a quality score becomes a commercial claim rather than independent evidence.

A metric is not an evaluation regime. COMET, CometKiwi, XCOMET and other learned metrics can be extremely useful, but their scores depend on model version, language pair, domain, reference quality, training distribution and the error patterns that matter in the deployment context. A general metric can correlate well with human judgment and still fail to represent the cost of a terminology error in a medical device instruction, a mistranslated obligation in a contract or a culturally inappropriate response in a Gulf-market assistant.

This is why Pangeanic continues to win work in adaptive Machine Translation Quality Estimation. Customers are not merely purchasing another generic score. They need estimation calibrated to their language pairs, content, terminology, risk thresholds and routing policies. They need to compare internal and third-party engines, determine what can be published, and send only meaningful uncertainty to human experts.

In one published test, Pangeanic MTQE v2 correctly rejected 98.9% of 6,006 deliberately incorrect translations in the ACES benchmark, spanning 16 language pairs and 68 error categories. That result does not establish universal performance, and it should not be detached from the benchmark on which it was measured. Let’s look at it operationally: the test asks whether known errors are stopped before publication, rather than treating a general correlation score as the final definition of quality.

Adaptive MTQE does not replace independent human evaluation, but it operationalizes it. Human judgments establish what quality means; adaptive estimation learns how to identify similar risk at production scale; and continued sampling detects when the relationship between the score and real performance begins to drift. Lastly (experience tells), judgements and preferences evolve so the initial Q1 settings in an MTQE system may be eroded in Q3 after humans have worked with the system. Adaptation is fast, re-tuning or fine-tuning per use case isn’t,

For high-consequence systems, buyers should retain control over their acceptance datasets, error taxonomies, calibration decisions, human adjudication records and the ability to reproduce evaluation outside the production vendor’s platform. They do not always need a different supplier for every test. They do need an evaluation layer whose evidence remains independently inspectable.

What this means for translators and localization professionals

I am not going to tell translators that everything is fine. Nor will I tell them that the profession is over because both statements avoid the real question.

The profession is moving from producing every sentence to determining which machine-produced sentences can be trusted.

Routine, high-volume and low-consequence work has already moved toward automated production with selective or no human review. This affects generalist translation and conventional post-editing most directly, particularly in high-resource language pairs. It reduces the amount of paid human attention required per unit of multilingual output. No amount of professional advocacy will restore artificial scarcity to that work.

At the same time, machine-produced language creates new requirements for judgment. Someone must define terminology, construct evaluation sets, identify error classes, adjudicate ambiguous output, test regional and cultural behavior, detect regressions, establish risk thresholds and decide when automation must stop.

These are not peripheral “language tasks.” They are part of how AI systems are built and governed.

My perspective is that three new markets are emerging.

1. Automated multilingual production

Machines produce most of the initial output. Human intervention is limited, sampled or triggered by risk. Prices will remain under pressure because the principal advantage is throughput.

2. Machine-supervised language operations

Professionals manage terminology, handle exceptions, perform targeted post-editing, investigate low-confidence segments and supervise workflows spanning several models and content types. This work requires stronger technical and operational fluency than traditional project-based translation.

3. High-consequence linguistic assurance

Domain experts construct gold standards, perform error analysis, red-team multilingual behavior, evaluate cultural and regional variation, adjudicate disagreements and accept responsibility for content whose failure carries legal, clinical, regulatory, safety or reputational consequences.

The premium is moving away from capacity and toward judgment combined with responsibility. A model can generate an answer. It cannot assume clinical liability, sign a regulatory declaration or explain to a court why a particular translation decision was acceptable.

But we should be honest about the arithmetic. New evaluation, alignment and assurance roles will not necessarily absorb everyone displaced from routine translation. One specialist supported by automation can supervise far more output than one translator could previously produce. Some new expert roles may command higher rates, while scaled annotation, review and data work may be as commercially pressured as post-editing.

The work is changing shape, but it may also employ fewer people per million words.

The professionals best positioned for the transition will combine language competence with domain depth, error analysis, data literacy, model behavior and the ability to translate an abstract quality requirement into an operational decision. Their value will lie less in being available to process volume and more in knowing where automation fails, why the failure matters and what evidence is required before the system can proceed.

The future belongs neither to those who deny the machines nor to those who trust them indiscriminately. It belongs to the people who can tell organizations when the machines should be trusted.

What this means for buyers of language services

Buyers now face an apparent contradiction: they have more technological options and fewer neutral global integrators.

There are more foundation models, specialized translation systems, APIs and workflow tools than ever. At the same time, the number of suppliers capable of absorbing a complex global program across technology, managed services, interpreting, regulated content and AI data has become more concentrated.

That changes the primary procurement risk. The issue is no longer simply whether another supplier exists. It is whether the buyer can move its accumulated capability from one supplier or platform to another.

Own the evidence, not just the output

Translation memories, glossaries, terminology databases and style guides remain important. But an AI-mediated program produces additional assets:

• evaluation sets and gold references;
• error taxonomies and adjudication guidelines;
• model-comparison results by language and domain;
• prompts, routing policies and workflow configurations;
• human preference and correction data;
• approved terminology and exception decisions;
• quality thresholds and calibration histories;
• and evidence of model drift across versions.

These assets determine how well the system performs for the organization. Contracts written for a per-word service may not define who owns them, how they can be exported or whether they may be used to improve systems serving other customers.

Those questions must now be explicit.

Separate production from acceptance

For high-consequence content, the party producing the output should not have unilateral authority to define success. Buyers should establish their own representative test data, error tolerances, human adjudication method and escalation rules. Automated metrics should support that regime, not replace it.

Ask about topology and exit before price

Where is the content processed? Which models receive it? Is customer material retained or used for training? Can the system operate in private cloud, on-premises or air-gapped infrastructure? Can the organization export its linguistic and evaluation assets in usable formats? What happens if the supplier changes ownership, discontinues a model or alters commercial terms?

If those answers are unclear, the buyer is not pricing a translation service. It is pricing a dependency.

Do not confuse consolidation with capability

A larger organization can fund platforms, offer wider coverage and assume complex global programs. It can also carry integration risk, organizational layers and incentives to move customers toward its own stack. During a long acquisition and integration period, account teams, platforms, contracts and operating models may change.

Acolad customers should therefore map critical dependencies before completion, not after it. That does not mean abandoning the supplier. It means knowing which assets, processes and decisions must remain portable if the relationship changes.

The European control question

Acolad was the largest major language-solutions group headquartered in continental Europe. It is set to become part of a UK-listed company. Unbabel, one of Portugal’s most visible language-AI companies, became part of a US group in 2025.

The United Kingdom remains European, but it is outside the European Union. A US acquisition creates a different jurisdictional relationship again. Corporate domicile alone does not determine data residency, GDPR compliance or technical security, and it would be simplistic to claim otherwise.

The strategic issue is control.

European enterprises and public institutions increasingly depend on multilingual AI to communicate, retrieve knowledge, operate public services and make information accessible across languages. When major suppliers or critical technologies change ownership, customers can face new product roadmaps, infrastructure choices, contractual terms and commercial priorities even if the service initially continues unchanged.

Sovereign AI should not be understood as autarky or as an insistence that every component be built domestically. It is the capacity to retain meaningful choice when ownership, models, regulation or commercial conditions change.

For multilingual AI, that means control over:

• the data used to adapt and evaluate the system;
• the infrastructure on which sensitive content is processed;
• the models and versions permitted in production;
• the quality and behavioral evidence required for acceptance;
• and the ability to move without losing accumulated organizational knowledge.

Sovereignty is not ownership of every model. It is the preservation of decision-making capacity when the rules of the market change.

The market has changed twice in twelve months. Portability is no longer a theoretical requirement.

What the transaction means for the industry

Five conclusions follow from the RWS–Acolad agreement.

1. Translation is not disappearing, but translation capacity is no longer scarce

The world will produce and consume more multilingual content, not less. But a larger share of that content will be generated, translated and adapted by machines. Volume growth will not automatically restore human production margins.

2. Scale has moved from production to access and orchestration

The decisive scale advantage is not simply the number of linguists available. It is the ability to embed technology across enterprise relationships, integrate with customer systems, accumulate feedback and operate across languages, domains and regulatory environments.

3. AI data is a growth market, but not every AI data activity is defensible

Large, episodic collection or annotation programs can produce rapid revenue and still face low margins and concentration risk. The moat lies in provenance, rights, domain difficulty, multilingual authenticity, evaluation methodology and repeatable operating infrastructure.

4. Evaluation is becoming a control point

As production becomes cheaper and more automated, deciding what is acceptable becomes more valuable. Metrics alone are insufficient. Organizations need independent evidence, domain-specific thresholds and adaptive workflows that connect scores to operational decisions.

5. Human expertise is moving closer to consequence

Machines will perform more of the routine production. Humans will increasingly define quality, handle ambiguity, investigate failure and carry responsibility where errors matter. The transition is real, valuable and unlikely to be painless.

What I take from it

Jourik is right that RWS executed an opportunistic acquisition and that the pattern resembles TransPerfect’s purchase of Unbabel. The buyers had the balance sheets, patience and strategic clarity to wait for valuable assets to become available on favorable terms.

But the deeper lesson is not merely that two giants made precision strikes.

TransPerfect strengthened its control over models, research and evaluation. RWS is expanding the enterprise surface across which its AI platforms can be deployed. One primarily acquired machinery; the other primarily acquired the installed base. Both transactions show that value is moving away from the undifferentiated production of translated content and toward the systems that determine how multilingual AI learns, performs and is trusted.

That does not make language services worthless. It makes their former unit of scarcity obsolete.

For companies in this industry, the warning is that neither accumulated revenue nor accumulated capital guarantees survival when the basis of competition changes. Nor is it enough simply to rename translation operations as AI data. The next commoditization cycle has already begun.

For language professionals, the honest message is that the work is moving from production toward judgment and responsibility, but not on a one-for-one basis.

For buyers, the message is more urgent. They must control the assets and evidence that allow them to evaluate, adapt and move their systems. They have more technological choice, but they will exercise that choice only if their data, evaluation methods and organizational knowledge remain portable.

Translation will not disappear. But the translation industry, understood as a market that rewards the production of translated words, is ending.

The next industry will reward those who control multilingual workflows, data, evaluation and accountability. Linguists will remain essential, but fewer will be paid simply to produce text; more will be needed to determine when machine-produced language can be trusted. Buyers, meanwhile, must decide whether they want cheaper output or control over the system that creates it.

The RWS–Acolad transaction is what that choice looks like when written into a balance sheet.

Frequently asked questions

What did RWS agree to pay for Acolad?

RWS agreed to total consideration of £40.2 million. That figure includes approximately £17.8 million of cash expected to be present at completion, producing an enterprise value of £22.4 million. The enterprise value represents approximately two times Acolad’s expected adjusted EBITDA for the year ending September 2027.

Does the low valuation mean language services have become worthless?

No. Acolad generated £182 million in 2025 revenue, and RWS is acquiring substantial enterprise relationships, regulated-industry access, interpreting capacity and operational expertise. The valuation indicates that the market assigns a lower premium to traditional, low-margin translation capacity while valuing the ability to automate, orchestrate and expand those relationships through AI.

How is AI changing the translation industry?

AI is automating more first-pass production and moving value toward workflow control, adaptation data, terminology, evaluation, exception handling and domain accountability. Translation remains necessary, but the economic center is shifting from producing every sentence to governing the system that produces and validates multilingual content.

Does TransPerfect’s ownership of Unbabel make COMET closed?

No. COMET remains an open-source framework with published research and publicly available models. However, public access does not guarantee access to every private model variant, dataset, calibration layer or production capability developed by a commercial owner. Buyers should therefore maintain independent acceptance data and evaluation processes even when using open metrics.

What should localization professionals do?

Develop domain expertise and learn to evaluate model behavior, analyze errors, manage terminology, construct test data and connect quality decisions to operational risk. Routine production will remain under pressure, while value will increasingly attach to judgment, accountability and knowledge of where automation fails.

What should buyers do before the acquisition completes?

Map critical supplier dependencies, confirm ownership and export rights for linguistic and evaluation assets, establish independent acceptance criteria, review data-processing and model-training terms, and ensure that workflows can be transferred if technology, ownership or commercial conditions change.

Sources and further reading

• RWS: RWS reaches agreement to acquire Acolad , 3 August 2026.
• RWS Holdings RNS: RWS agrees to acquire Acogroup , 3 August 2026.
• RWS Holdings: Half Year Report for the six months ended 31 March 2026 , 11 June 2026.
• RWS Holdings: Results for the year ended 30 September 2025 , 11 December 2025.
• Acolad: Acolad enters an exciting new chapter with proposed acquisition by RWS , 3 August 2026.
• TransPerfect: TransPerfect acquires Unbabel , 22 August 2025.
• Slator: TransPerfect acquires Unbabel and outlines integration plans , 26 August 2025.
• Unbabel COMET: open-source framework and model documentation .
• Pangeanic: Adaptive Machine Translation Quality Estimation and ACES benchmark results .
• Jourik Ciesielski: “Another masterclass in opportunistic M&A?” , August 2026.
• Manuel Herranz: Visit by Jourik Ciesielski to Pangeanic’s Valencia offices , July 2026.
• Manuel Herranz: TransPerfect’s Unbabel acquisition is a bellwether for a new era in AI language technology , 24 August 2025.