Lokalise

Ask a team that has shipped a multilingual product where localisation actually goes wrong, and nobody says the translations were bad. They say a string was added on Tuesday and nobody noticed until a German user reported English text in production.

Localisation fails at the seams: between the developer adding a key, the designer changing a label, the translator working in a spreadsheet dated three weeks ago, and the release that shipped before the file came back.

Lokalise has spent years building for that problem rather than the translation problem, and it shows in the shape of the product. The AI capabilities layered on top are substantial and current, but the reason to choose it is the pipeline underneath. This review covers both, plus the compliance detail that matters if you handle regulated content and the point at which the platform is simply too much.

Lokalise: The Platform Facts

Continuous Localisation, Not Bulk Translation

• It is a pipeline, not a translator — strings move from GitHub, Figma, Jira and your CMS into translation and back, without anyone handling a spreadsheet.
• It routes between multiple language models rather than committing to one, selecting by language and content type.
• Brand voice is enforced structurally through custom AI profiles applying your approved terminology and tone, not left to a translator's judgement each time.
• Humans stay in the approval path with role-based sign-off, which is what separates a translation management system from a machine translation endpoint.
• Compliance is unusually deep: SOC 2 Type II, ISO 27001 and ISO 27017, EU-region AWS hosting, annual penetration testing and SAML SSO.

Sixty Integrations Is The Moat

Lokalise states more than sixty integrations, thirty-plus file formats, thirty-three webhooks and ten pre-built SDKs, alongside a REST API with ninety-five endpoints, a CLI and MCP server support.

That list is the product's actual competitive position, and it is worth understanding why rather than treating it as a feature count. Localisation is not a task you perform; it is a state you maintain. Every string added, every label changed in a design file, every article published in a CMS creates translation debt the moment it lands.

A platform that connects to GitHub and GitLab, Figma, Jira, Contentful, Webflow and WordPress can detect that debt automatically, which is the ground Crowdin competes on too. One that does not requires a person to notice, export, send, receive and re-import — and that person will eventually be on holiday during a release.

The SDKs matter for a related reason. Over-the-air delivery of translations means a copy fix does not require an app store submission, which for mobile teams is the difference between correcting a mistranslation in an hour and correcting it in a fortnight.

MCP server support is a genuinely forward-looking inclusion. It means a coding agent can query and update localisation state directly, which is where a meaningful share of this work is heading.

Localisation is not a task you complete. It is a state you maintain, and every untranslated string is debt accruing from the moment it is written.

Routing Between Models Instead Of Betting On One

Lokalise describes AI orchestration across multiple large language models, selected dynamically by language and content context.

This is a better architecture than it may appear, and the reason is empirical. Model performance in translation is not uniform across languages. A model that handles French and German excellently may be noticeably weaker in Japanese, Turkish or Finnish, where morphology and word order differ more sharply from English.

Committing a platform to one provider means inheriting that provider's weak languages permanently. Routing means a bad language pair can be reassigned when a better model appears, without the customer changing anything.

The practical implication for a buyer is to test in your weakest language rather than your strongest. Everything handles Spanish well, including a raw engine such as DeepL with no pipeline attached. What you need to know is what happens in Korean, Hungarian or Arabic, where the differences between models are largest and where your reviewers will spend their time.

The 95 Percent Figure Needs Unpacking

The site states 95 per cent AI accuracy, alongside claims of ten times faster releases, up to 90 per cent lower costs and fully on-brand output.

Translation accuracy is not a single measurable quantity, which is the first problem with the number. Automated metrics compare output against a reference translation, but a good translation that differs from the reference scores badly, and a fluent but wrong translation can score well. Human evaluation is more meaningful and produces different results depending on who evaluates and against which brief.

The number also cannot be uniform across languages. Any accuracy figure is an average, and the average conceals exactly the information you need — which of your target languages fall below it.

None of this means the figure is dishonest. It means it is not usable as a purchasing input, and no vendor in this category publishes one that is.

What is usable is the structure around the number. Custom AI profiles applying your approved terminology, AI scoring for quality evaluation, and role-based human approval before publication together make quality a managed process rather than a hoped-for property. That architecture is worth more than any percentage, because it is what catches the five per cent.

• Test in your hardest language pair, not your easiest — the average hides your worst case.
• Load your real glossary before evaluating, since terminology enforcement is most of what separates platforms.
• Include your most awkward content type: legal copy, error messages, or anything with strict length limits.
• Check how it handles placeholders and variables, which is where machine translation still breaks user interfaces.
• Have a native reviewer score output blind against your incumbent process rather than reading vendor benchmarks.

Compliance Depth Is Genuinely Unusual Here

Lokalise publishes SOC 2 Type II with annual audits, ISO 27001 certification held since November 2020, ISO 27017, GDPR compliance, HIPAA alignment through its SOC 2 programme, AWS hosting in EU regions, SAML SSO, role-based access control, exportable audit logs on the enterprise tier, and annual penetration testing conducted through YesWeHack.

That is a deeper set than most tools in any category publish, and considerably deeper than typical for localisation. It matters because of what passes through a translation platform: unreleased product copy, marketing plans before launch, support content containing customer scenarios, and in regulated industries material with genuine legal weight.

ISO 27001 since 2020 is the detail worth noticing. A certification held for years, with the surveillance audits that implies, is stronger evidence of a maintained security programme than a recently obtained one.

EU-region hosting answers the data residency question directly for European buyers, which for public-sector and regulated procurement is frequently a threshold requirement rather than a preference.

The gap to note is placement rather than absence: SSO and exportable audit logs sit at the enterprise tier, with audit log retention stated at six months. If your security policy requires SSO — and most enterprise policies now do — that decides your tier before any feature comparison does.

Running A Localisation Pilot That Proves Something

Localisation platforms demo beautifully and integrate painfully, and the gap between those two experiences is where most disappointment lives. This sequence surfaces it during a trial rather than after signature.

The bar is step four. If a string added in your repository does not appear in the platform without human intervention, you are buying a translation editor rather than a pipeline.

1. Connect one real integration — your actual repository or CMS, not a sample project. Integration friction is the thing you are testing.
2. Import your real glossary and style guide, then translate a genuinely awkward content type: error messages, legal copy, or strings with hard length limits.
3. Translate into your most difficult language pair and have a native speaker score it blind against your current process.
4. Add a new string in the source system and watch whether it appears in the platform for translation without anyone doing anything.
5. Push a translation back and confirm it lands correctly, including placeholders and variables, which is where machine translation most often breaks an interface.
6. Time the whole loop. That figure, not the accuracy percentage, is what tells you whether releases actually get faster.

How Lokalise Is Licensed

Access is a tiered subscription with a trial available, rising to an enterprise tier quoted through sales.

Two things drive placement rather than one. Volume matters — how many keys or words you manage — but so does capability gating, and the second is what usually decides. SAML SSO, exportable audit logs and the deeper administrative controls sit at the enterprise level.

For a small team the practical consequence is straightforward: the lower tiers are genuinely usable, though a lighter platform such as Ailaysa is worth weighing before committing. For a larger organisation, the security requirements in your own policy will typically place you at the enterprise tier regardless of your word count, and it is more honest to plan for that from the start than to price the middle tier and be moved later.

We do not publish amounts, since tiers change. What to establish before committing is which specific tier carries SSO and audit logs, and how the volume meter counts — keys, words or languages — because that determines how cost scales as you add markets.

Localisation Terms That Matter

Five terms that decide whether a platform will fit how your team actually works.

Translation management system (TMS)

The pipeline that moves content between source systems, translators and production. The thing you are actually buying — the translation itself is a component.

Continuous localisation

Translation triggered automatically by source changes rather than run as a periodic batch. The difference between shipping multilingual and shipping English then catching up.

Translation memory

A store of previously approved translations, reused when the same string reappears. Improves consistency and reduces cost the longer you run it — and it is yours, so ask about export.

Glossary and term base

Your approved terminology, enforced automatically. The main defence against a product feature acquiring three different names across five languages.

Over-the-air delivery

Shipping updated translations to a live app without an app store submission. Turns a mistranslation from a fortnight's problem into an hour's.