What we’re announcing today, the Phrase Language Intelligence Platform, our four strategic product initiatives, our AI-native operating model, our Forward Deployed Engineering programme and new pricing concepts – is not just a ‘vision’. We deliberately held back to have proof points to share, and are now sharing what is already in motion.

We launched Forward Deployed Engineering several months ago. We have been rebuilding our operating model around AI agents for months. A month ago, we right-sized our organization to reflect our strong conviction. That a smaller, genuinely AI-native team moves faster and produces more than ever before. We’ve been trialling new pricing concepts with customers for months. Our new website launched yesterday.

TL;DR: every company claims to be AI-first. Most are not, most leadership teams aren’t walking the walk. That’s not us, this is our statement of intent. We are all-in, all the way.

The Language Intelligence Platform

Phrase has evolved before. We started as two products, Memsource, a translation management system, and PhraseApp, a developer localization tool. And we unified them into a single platform. We’re evolving again, but now the market is moving faster than at any point in the past decade.

“Localization Platform” no longer captures what we do or where value actually sits. The word “localization” conjures workflows, word counts, and project managers. “Language Intelligence” names something structurally different. An orchestration layer, harnessing AI models and agents to make them actually work for enterprises operating across languages and cultures. “Localization” is too often misunderstood as a solved problem.

Ask any localization professional how often they’ve had to explain why their work matters. “Language Intelligence” names what the best localization teams have always known, and builds it into a platform that delivers it at enterprise scale. “Language” speaks to every department, and every human; it’s also the programming language of both humans and agents.

“Intelligence” speaks to safe use of AI, to action and analytics, to results and business value. Our Intelligence layer applies the context – glossaries, style guides, brand voice, cultural rules, demographic sensitivity. It weighs competing priorities, evaluates quality at every stage, and gets measurably better with every piece of content that runs through the system.

Research from Stanford and MIT (Meta-Harness, Lee et al., 2026) shows that changing the orchestration harness around a fixed language model – while keeping the model itself identical – produces a 6× performance gap on the same benchmark. Critically, as models get better, the harness amplifies rather than diminishes in value.

Customers who use Phrase aren’t just buying a translation tool. They’re accumulating an intelligence infrastructure tuned specifically to their terminology, their markets, and their quality standards – a platform they can build on themselves and with an ecosystem  that gets more effective with every piece of content processed.

Four strategic product initiatives

We are organizing ourselves around four major areas of focus. Two pairs of self-reinforcing development loops. 

Headless and Guided form the first loop: two ways of consuming the platform, both enabling entirely new experiences.

Headless is straightforward to explain and radical in its implications. A business person using any business system (or Claude) might have an agent that speaks to our agent, negotiates an interaction, presents options the person didn’t know to ask for, and the content comes back transformed and adapted – without the user ever knowing they touched Phrase. This isn’t theoretical. A customer found our MCP server before we announced it publicly; their AI champion pulled it from GitHub and described it as “an absolute game changer.” Technical teams are treating Phrase as infrastructure, not a destination. We’re building for both.

Guided is the opposite consumption mode. A visually dynamic interface with chat, context awareness and a coherent design system where users describe what they want and the platform handles the rest. No requirement to learn a translation management system. No complexity visible unless you want it.

This is Phrase Atlas, our guided, agentic experience, and it’s already in early customer preview. One concrete outcome: building the guided experience helped us to discover needs for APIs and endpoints we hadn’t surfaced before. We created 30+ new APIs in three weeks as test users explored needs through the chat layer. We’re reaching down inside the product and turning it inside out in the best possible way. The UX that formed the outer shell is being submerged, and the capabilities that used to be hidden in the core are brought to the surface. The product becomes the roadmap.

Compounding Intelligence and Intent Proximity form the second loop.

Compounding Intelligence is the layer across the entire language stack that learns which combinations of context, and with what weighting, produce the best outcomes for a specific customer. This is customer-specific, not based on legacy linguistics frameworks, and enriches over time. Customers who have run content through Phrase for a year carry accumulated intelligence about their markets and brand that no other platform can replicate.

Intent Proximity moves us definitively beyond translation as the output. The question we’re asking is: what was the intent of the user, and did the content output achieve it? Selling a hotel room to a business traveller requires different copy than selling the same room for an anniversary weekend. Multiply that across 30 cultures and languages and you get a problem no general-purpose AI model solves without specialized orchestration.

Phrase increasingly understands intent on the way in and measures outcome on the way out, then aims to adapt the content accordingly. That’s the shift from “did we translate this?” to “did the content work?”

What AI-first actually looks like inside Phrase

We have fundamentally rewired how we operate as a business, and strive to be a world-leading company in AI ‘ways of working’, across every function.

We’ve been inspired by the model set by Ramp, Y Combinator, Shopify, Coinbase and others. Companies that have embedded AI into the operating system of the business, not just the product. We’ve built our own version of this.

Every Phrase colleague has access to Claude Enterprise, unlimited licences, across the company. We’ve deployed Launchpad, our internal AI platform, where anyone can build and ship internal apps & tools with a secure infrastructure without needing engineering to build it for them. We have a shared skills library where what one person builds, raises the floor for everyone. We’ve rolled out an AI Fluency Framework. Five levels from Beginner to Architect – with a target of every colleague reaching a Builder level. Builder means: has automated a meaningful part of their job and shared it with the team. We’re tracking it.

If you’ve read ‘experts’ on Twitter talking about how they use AI, we’re already doing it and ahead of most. Meta-prompting, LLM Wikis, Agentic Runtimes, Agent Teams, Loops, Goals, Routines, Harnesses, Memory Layers etc.

The internal innovation has been incredible, it’s inspiring. We’re not dabbling with AI as glorified search and a research assistant here. We are all building meaningful new capabilities for our work.

Here are some examples. These are real, named, approved applications, logged in our internal AI registry, security-reviewed, and live:

Engineering. We built an agent-ready design system layer that lets AI coding agents generate production-compliant UI from natural language prompts, without a designer in the loop for routine tasks. A separate agent reads a product spec or RFC in Confluence and automatically generates a full Jira initiative – epics, stories, context – cutting the breakdown work that previously took engineers hours. Another monitors our technical support Slack channel, classifies incoming bug requests in real time, and routes them to the right specialist agent. Engineers are increasingly directing agents; the shift from writing every line to directing what gets built is already happening.

Sales. “SIP” (our Sales Intelligence Platform) researches a prospect across nine internal systems and the web, then synthesises a persona-specific briefing (talk tracks, objection handling, discovery questions) in under five minutes. The call prep that reps never had time to do now happens automatically before every conversation. “Lead Machine” scans approximately 100 enterprise accounts every night for localization buying signals (hiring patterns, technology changes, expansion moves, competitive activity) ranks them by relevance, and surfaces them to BDRs each morning with context for personalised outreach. 

Marketing. Product Marketing is automating the full workflow of working with an agile Product and Engineering team, so that (agent written) PRDs drive automated GTM materials that self-update at the end of every sprint. In ABM, building a target account list for a campaign used to involve seven manual steps across four teams and took six to eight hours. An n8n workflow now automates the data enrichment, Salesforce account creation, and stakeholder notification in a fraction of that time – saving four to six hours per campaign cycle. 

Customer Success. CS is automating meeting agenda prep, writing the follow-up, creating the Slack update, tracking the action items. With six to ten customer meetings per week, that was six to ten hours of administrative overhead per CSM, every week. Now the team can spend more time with customers.  A separate Renewal Conversation Monitor watches every account in the 180-day pre-renewal window, analyzes call transcripts to assess whether renewal conversations have happened, and escalates overdue cases automatically – so nothing slips through.

Finance and Legal. The Finance team built a hub that replaces manual Google Sheets and a separate reporting tool for monthly P&L work, automating imports from NetSuite and using AI to detect patterns and flag anomalies across reporting periods. The OTC Cockpit gives every finance team member a single-screen view of any customer’s commercial state across six systems simultaneously. Legal built a bot to handle the 20–30 NDA reviews and up to 40 contract reviews the team processes every month, alongside the daily volume of FAQ questions that previously occupied significant counsel time – freeing lawyers for the complex, strategic work that genuinely requires their judgment.

People: The People Team is building a Learning Lab that gives every employee access to personalized, on-demand training built around their individual learning objective, in their preferred format, in minutes rather than weeks. The system produces a tailored module in 5–10 minutes. This replaces a traditional content development cycle that typically takes 2–6 weeks, without compromising quality or relevance.

We’re not at the end of this journey. We’re at the beginning. But the infrastructure is in place, and the results are visible.

The momentum is already there

None of this is happening in a vacuum. While we’ve been rebuilding internally, customers have been showing us, in real production results, what’s possible when the platform is deployed properly.

A global health business opened new markets they could not previously afford to open with human effort, that now account for ⅓ of global revenue. One of their product lines grew revenue by 15% in a year, based on new market entry.

An LSP partner running fully automated translations for one of the world’s largest enterprise software companies put it directly: Phrase’s Translation Agent is “the best LLM-based feature we’ve seen in any language platform to date.” Zero human post-editing. Faster turnaround. A repeatable framework they’re now extending to other clients.

A global health company running multilingual customer support in Zendesk was copying and pasting content into external tools for translation – manually, ticket by ticket. Phrase’s integration removed that entirely. The result: an estimated 30–50% reduction in operational effort and 20–30% improvement in turnaround times, with the team’s attention now on the conversations rather than the logistics.

These aren’t edge cases or bespoke implementations. They’re examples of what happens when enterprises stop treating localization as a workflow and start treating it as an intelligence layer. The platform was ready. The customers proved it.

Forward Deployed Engineering: why AI needs it

There’s a reason the most sophisticated AI companies (OpenAI, Salesforce, Ramp) have all moved toward forward-deployed engineering as the dominant enterprise AI deployment model. FDE job postings grew 800% in 2025 alone.

The reason is this: AI is not deterministic software. It drifts. A pipeline that performs well in evaluation, or the exec demo, can behave differently in production, at scale, across languages, with your specific content and brand constraints. Traditional software can be installed and left alone. AI will drift – what works at launch will not work weeks later. It needs to be tuned, monitored, and adjusted as it runs.

This is why continuous engineering effort alongside AI investment matters. Not just as a delivery model but as a structural necessity for AI services. Every enterprise that deploys AI seriously will need engineers embedded throughout the value creation, not just at the point of sale. We saw this coming, built the model, and have already put it to work.

Our Forward Deployed Engineers are not a professional services team. They write production code. They deploy AI pipelines in your environment. They guarantee outcomes, not slide decks. The difference from traditional consulting is the difference between an architect who draws a house and a builder who constructs one. We build the house, you move in. As we build it, we can teach you to maintain it, introduce partners, or do it for you.

Every production deployment our FDEs run also teaches us something the product doesn’t yet know. Which context layers matter most for which industries, which quality interventions actually change outcomes, which orchestration configurations unlock adoption at scale. That knowledge flows back into the platform, benefiting every customer.

The results are already concrete.

For a global company operating sensitive health content across seven languages, we built a self-improving AI orchestration pipeline. The challenge was precision: profanity filtering, medical terminology adherence, and brand style compliance across every locale, with no automated QA pipeline previously in place.

We built a LearningsDB; a system that feeds QA outcomes back into the pipeline after every cycle, raising quality autonomously. The outcome: three automated quality checks per segment, surgical AI fixes across all seven languages, and approximately 50% reduction in the human review time needed. The pipeline improves without anyone having to intervene.

For a SaaS company localizing product pages across 22 locales, the goal was to move away from full human post-editing – a process that was expensive and slow. We deployed Phrase Auto Adapt across all 22 locales with two rounds of AI tuning for brand voice, tonality, and style. The outcome: 100% replacement of human post-editing, 70–80% cost savings per content stream, and a delivery timeline cut from approximately one month to one to two business days. Not in a lab. In production.

These aren’t projections. They’re done.

Evolving how customers pay: introducing new pricing concepts

As the platform evolves, so does how customers consume it, and we’re evolving pricing to match.

For a long time, the cornerstone of localization technology has been the Processed Word – a word that has entered any part of the workflow. This makes less sense in a world with looping workflows, where quality isn’t binary but is scoped to meet a business objective, where AI handles some content fully autonomously and human expertise is applied selectively to add the most value. 

So, we’re introducing two additional concepts as alternatives to the Processed Word:

Downloaded Word – a word that is delivered to the customer, regardless of the level of human involvement. We don’t count the words ‘in’ to the Platform, we count the words ‘transformed’ by the platform. This reflects the growing volume of content that enterprises localize entirely through AI-driven pipelines.  ‘Processed Words’ are unlimited, so you can add all the context you need to, to drive the best fit-for-purpose AI outcome. 

Human-touched Word – a word where a human has intervened in the localization process, whether through post-editing, review, or quality evaluation. This recognizes that human expertise is now a scarce and premium input, deployed where it matters most rather than applied uniformly across all content.

Alongside these new alternatives to Processed Words, we’re introducing Credit-Based Pricing in our Business and Enterprise tiers as an alternative to our capacity-based models.

Rather than paying for a fixed-volume commitment tied to a specific capability, customers purchase credits and use them flexibly across the platform, products, use cases, and content types. Credits scale and get cheaper with actual usage, not with licence seats or module boundaries. For enterprises with multiple content types and ambitions to scale AI volumes, this provides significantly more flexibility and better alignment between spend and value received.

These changes reflect a simple conviction. Pricing should follow value, not drive infrastructure decisions. As the way our customers use language intelligence evolves, the commercial model should evolve with it.

What this adds up to

The more content AI generates, the more the world needs a system to make that content actually work – in the right language, for the right audience, in the right market, with measurable outcomes. That system is Phrase. That’s what a Language Intelligence Platform does.

We’ve deployed our AI infrastructure and led cultural change from the top. We’ve launched our FDE programme. We’re modernizing our pricing. We’ve launched four product initiatives that change how Phrase will be consumed and how intelligence compounds inside it. We’re deeply committed to delivering value for customers. 

I am super excited about how fast things are moving, and to explore what we can build together with you – our customers and our partners.

Georg