e2f and the business of being right

A travel website had a summer problem. Its catalog needed to speak six languages, and the deadline was two weeks away. The assignment contained 1.77 million words. In the project reported by e2f and machine translation company Lilt in 2016, more than 100 language professionals worked with a system that learned from their corrections. They finished in ten days. The revealing detail was how the corrections traveled: back into the software, where they could help the next translator.

The correction was the asset

There is an ordinary version of machine assistance: software produces something, a person fixes it, and the corrected file leaves the building. The travel project made that relationship more useful. Translators accepted or revised suggestions. Approved revisions returned to the system. Learning happened while the job was being delivered.

The plumbing was wonderfully unglamorous. Source Excel files were converted for automated processing, sent through Lilt’s API, then turned back into spreadsheets after translation. Seven project managers and a localization engineer helped coordinate the work. The machine needed an organization around it.

The release said the work arrived ahead of schedule and within budget. That establishes a delivery result, rather than a universal savings rate. Still, the business logic is clear enough: a useful correction can improve both today’s deliverable and tomorrow’s workflow.

A translator reads the room

Michel Lopez founded e2f in 2004. Translation and localization supplied an education in a particular kind of error: language that appears acceptable until someone understands the context. A term can be grammatically correct and inappropriate for the product. A sentence can travel badly between cultures. Fluency alone is an unreliable quality test.

That experience gave e2f a route into collecting, transcribing and annotating data for language systems. Its earlier services included multilingual speech and text datasets, quality evaluation and material for fine-tuning. These jobs require instructions, reviewers and an agreed definition of a good answer.

Its work with Intento shows the evaluation side. The partners collaborated on datasets for translation research; Intento’s 2024 report compared 52 translation engines and large language models across multiple domains. The interesting skill here is building a fair test, then recognizing errors that a smooth sentence can conceal.

Leaving the comfortable work

By 2021, e2f was publicly explaining a deliberate change in its service mix. In company-authored partner coverage, Lopez described reducing exposure to localization staffing and subcontracting while pursuing technology-enabled work. His diagnosis was blunt: “We saw that the localization industry was becoming stale”.

This was a strategic response to automation, rather than a publicly documented rescue after a failed product. e2f wanted more work in which its technology and linguistic knowledge mattered together. It also acquired app localization specialist Applingua, adding capabilities around app-store content and related language work.

The lesson for another services company is practical. Examine the parts of delivery where staff repeatedly exercise valuable judgment. Capture those decisions, standardize the workflow and make the knowledge reusable. Selling another hour of effort becomes less attractive when software can preserve something learned during that hour.

The customer changed the question

In June 2026, CEO Lawrence Snapp announced that e2f was becoming TrustScale. His explanation centered on customer conversations: buyers increasingly wanted help trusting AI outputs. Better model performance still mattered, but organizations also needed reasons to believe a particular answer before using it.

Argus, announced in August, checks generated content for unsupported claims and offers evidence-based corrections. The company describes an approach built around external evidence and deterministic verification. Its launch covered 12 languages, with enterprise deployment and an individual research preview. The pitch fits a familiar frustration: checking a plausible answer can consume the time saved generating it.

September brought ArgusRL, an API-backed service for evaluation and reinforcement feedback. It breaks responses into claims, retrieves supporting or contradictory evidence and returns structured results. Cases needing closer inspection can go to human annotators. TrustScale reported that one unnamed customer accepted more than 95% of automated evaluations without correction. That is a company-reported deployment result, with a specific setting behind it.

“Tell the whole truth, always.”TrustScale’s stated integrity value

Where the money changes hands

e2f’s commercial history spans managed language work, AI data services and software. Transifex still documents ordering e2f translation inside its platform, with minimum fees of $20 or $40 per language. Translators join the relevant projects for an order and leave automatically when it is complete. Access is built around the assignment.

Earlier e2f Connect materials listed human evaluation tasks at $2.99, $5.99 and $8.99, depending on review depth. Those are historical tiers. Current enterprise buyers discuss requirements with TrustScale; ArgusRL is also offered through AWS Marketplace. Review depth, volume and specialist involvement are sensible questions for the buying conversation.

The customer base stretches across technology and digital products. TrustScale displays names including Amazon, Google, Autodesk and Calm. Those logos describe company relationships without establishing which buyer uses which current tool. Its market position sits between language operations, managed AI data and the growing purchase of separate AI assurance.

Give the reviewer something to inspect

The alternative may be an internal review team, a managed annotation provider or another model judging the output. TrustScale’s proposed distinction is evidence that a person can inspect, connected to a workflow that can feed improvements upstream. Its language background matters most where context complicates a simple pass-or-fail decision.

The approach has conditions. Relevant evidence must be accessible; a subjective judgment still needs an appropriate standard. Product teams should test on their own languages, documents and consequential mistakes. The part worth copying is the feedback design: define acceptable work, preserve corrections and route uncertainty to someone equipped to resolve it.