When we set out to translate the Telebib2 standard from EDIFACT to JSON, the task looked almost simple. AI can read an EDIFACT message, rewrite it in JSON and explain the result in plain French or Dutch. Surely it could convert the whole standard in a few passes?

Reality turned out to be more nuanced. A standard that has served the sector for three decades carries a great deal of knowledge that lives in practice rather than on paper: business rules that experienced implementers apply naturally, choices that made perfect sense with the tools and constraints of their time, and refinements built up release after release. This is not a shortcoming; it is what a living standard looks like. But it is exactly what an automated translation cannot guess. Many organisations discover the same thing when they hope AI will turn legacy COBOL into modern code overnight: the syntax translates easily, the meaning does not. Converting the format was the easy part. The real work was preserving what the sector had agreed the data means.

That experience shaped our answer to a question many people now ask: if AI can translate anything, do we still need a shared standard? In 2026, the Telebib2 Centre within Assuralia, supported by thewave, used AI on two major modernisation projects. Our conclusion is clear: yes, and more than ever. This article explains why, what we learned along the way, and what we think lies ahead for the standard and for any organisation that maintains a shared reference.

A shared language built by the sector

Sector standardization in Belgian insurance goes back to 1975, when insurers and intermediaries began harmonizing accounting documents to improve productivity. That cooperation gradually extended to data and business processes. In 1994, the TELEBIB protocol formally recognized a standard based on UN/EDIFACT for electronic data exchange. Today, the Telebib2 Centre within Assuralia develops and maintains this shared language.

No single organization owns the standard. The Strategic Committee for Normalization brings together insurers and broker associations to set priorities and direction. Normalization working groups, coordinated by the Telebib2 Center, turn sector needs to agreed messages and data definitions, drawing on insurers, brokers, broker software providers and exchange platforms. The Telebib2 Centre provides continuity: it maintains the reference, publishes releases and documentation, and supports implementation.

In practice, the Telebib2 standard lets insurers and broker management systems exchange structured information for policy administration, claims handling and accounting. It sets shared definitions, message structures and code lists, so that what one system sends, another understands. One distinction matters: Telebib2 is a language, not a platform. It defines what an exchange means, not how it travels. That is what allows different systems and providers to work together without inventing a new language for every connection.

Why 2026 called for two projects

The meaning of the standard was sound; its packaging had aged. EDIFACT is compact and robust, but few developers today read it fluently. The reference information lived in a Microsoft Access database: practical for its maintainers, but hard for modern tools to use and almost opaque to AI. And thirty years of explanations were spread across documents and an ageing website. The knowledge existed, but it was hard to find and sometimes inconsistent.

So the Telebib2 Centre launched two projects with a single goal: bringing the standard into today’s world. Rebuilding the website would make it easier to understand and navigate. Translating it into JSON would give it a modern, widely adopted format: easier to read, supported by the tools developers use every day, simpler to document and validate, and, as a welcome side effect, directly usable by AI. Neither project was meant to change what the standard says. The JSON version is not a new standard: it carries the same definitions and the same business meaning in a form today’s tools can work with.

What AI really changed

Each project took about fifteen weeks. AI clearly sped things up: Claude helped design the new website and migrate its content, and for the JSON translation several AI agents worked in parallel under a coordinating agent. But speed is the least interesting lesson. Three others matter more.

Rules first, production second. The AI produced consistent results because the rules had been written down before generation began: naming conventions, structural patterns, how each EDIFACT construct maps to JSON. Where the rules were explicit, the output was consistent. Where they were implicit, the AI improvised. AI amplifies the clarity of the framework you give it, and its vagueness too. Writing the rules is the real work.

AI as a revealer. Because it can examine the whole standard at once, AI brought to light inconsistencies accumulated over decades: explanations that had drifted apart, terminology that varied, changes not carried through everywhere. Seeing a problem is not solving it, though. A suspected inconsistency is a question for the experts, not proof that a definition is wrong.

Assistance, not authority. An explanation written by AI is not a sector rule. Every output was reviewed, and what gets published or changed remains a decision for the standard's governance. This sounds obvious, but it is worth saying: the faster production becomes, the greater the temptation to skip validation.

The future of the standard: opportunities and their conditions

AI will change how Telebib2 is used far more than how it was built. We see five opportunities, and each one comes with a condition. Ignoring the conditions would turn the opportunities into risks.

1. A reference for machines, not only for people

The Centre now provides JSON schemas covering the full standard. They can be loaded into a development environment, giving developers and AI coding assistants an explicit reference to build from, with no need to retrain a model. Tomorrow, AI agents could consult the standard directly instead of relying on what they think they know.

The condition is our opening paradox. If AI can translate anything, why keep a standard? 
Because translating on the fly is an educated guess. Each translation differs slightly, and none of them is binding. A standard is a contract: agreed meanings, versioned and governed by the sector. The better AI gets at translating, the more it needs a stable target to translate to.

2. Lower barriers to entry

AI is bringing new players into insurance: services that process incoming emails or streamline claims handling for brokers and insurers. To be useful, they need data held in existing systems. If every provider arrives with its own definitions, each connection becomes another bespoke integration to build and maintain. Telebib2 gives newcomers a documented language to plug into, and AI can help them understand how their data relates to it. The same goes for organisations still working in EDIFACT.

Test data is a concrete example. Writing test messages in EDIFACT has always been tedious. With the JSON schemas, AI can generate realistic synthetic messages, which are then checked against the schemas and reviewed for business rules rather than assumed to be correct.

The condition: avoiding dialects. If every AI tool interprets the standard in its own way, fragmentation returns, only faster than before. The answer is shared reference material, such as official test sets, conformance checks and reference examples, so that "Telebib2-compliant" means the same thing for everyone.

3. A standard that evolves with better insight

AI can show how a proposed change would ripple through every message before the working groups decide. It can also help a member turn a business need into a well-formed change request.

The condition: stability. Once a release is frozen, it no longer changes, because partners build on it. An error that slips through stays for a whole cycle. Faster drafting must therefore be matched by stronger review before each freeze.

4. Wider access, without lowering our guard

Looking ahead, one could imagine a website that answers questions asked in plain language and points to the exact definition. This feature does not exist today; it is one possible next step. In the same way, French, Dutch and English documentation could one day be maintained together from a single source. And as more people get their answers through AI assistants rather than by browsing, documentation will need to serve both audiences.

The condition: harmonization is not security. Insurance data is personal and sometimes sensitive. A common language does not authorize access or make an exchange secure. Data sharing still requires permissions, confidentiality safeguards, GDPR compliance and, for certain insurance uses of AI, the obligations of the European AI Act. Telebib2 provides the shared meaning; it does not replace the responsibilities of the parties exchanging data.

5. More capacity, same expertise

AI lets a small team do far more. The condition is to keep the knowledge in human hands. If fewer and fewer people understand why the standard is built the way it is, the sector ends up depending on tools it can no longer check. The same applies to depending on a single AI provider. The answer: document the business reasoning, not just the structures, and keep AI as a replaceable assistant. Maintaining the standard must still be possible without it.

Questions worth asking in your organisation

These lessons are not specific to insurance. Any organisation that maintains a shared reference, whether a standard, a data model or a product catalogue, can ask itself:

• Can a machine read our reference data, or only a person?
• Are our rules written down, or do they live in the heads of a few experts?
• Who validates what AI produces, and at which point in the process?
• If two partners use AI to interpret our standard, will they reach the same result? How would we know?
• Could our process keep running tomorrow without our current AI tool?

Conclusion

So, if AI can translate anything, do we still need a standard? Yes, because translation is not agreement. AI can bridge two formats, but only a sector can agree on what the data means.

The more AI transforms insurance, the less the sector can afford fragmented data exchanges. If every new provider brings its own language, innovation multiplies interfaces instead of connecting services. Telebib2 offers a strong foundation: shared definitions and documented structures that insurers, brokers, software providers and AI newcomers can all build on.

The opportunity is twofold: use AI to strengthen the standard and use the standard to connect AI-driven innovation. AI can help every player move faster. A shared standard helps the whole sector move forward together.

Written by Benjamin Theunen and Olivier Grignard