In agrochemistry R&D departments, the frustration is real but rarely put into clear words. Teams are drowning in regulatory documentation. AMM (marketing-authorisation) dossiers accumulated over decades. E-PHY, EFSA and ECHA data to be continually cross-checked. Expert knowledge walking out the door into retirement. And yet, when an engineer has to consolidate a pre-submission dossier for ANSES (the French agency for food, environmental and occupational health & safety), they still spend 3 to 6 weeks searching, cross-referencing and verifying. By hand.

Generative AI seemed to promise a step change. Proofs of concept sprang up across 2023 and 2024. The outcome: mostly abandoned. Kept under close watch in a few cases. Genuine deployment: almost non-existent.

Why?

What RAG does well — and where it goes off the rails on regulatory content

RAG (Retrieval-Augmented Generation) is an ingenious architecture. It splits your documents into fragments, indexes them in a vector database, and lets a language model synthesise an answer from the fragments most "similar" to the question asked.

On general-purpose use cases — meeting summaries, assisted drafting, internal FAQs — it works very well. On phytosanitary regulatory corpora, it is another story.

35–45% RAG error rate on phytosanitary regulatory corpora
< 5% KnowWeave error rate on the same corpus
100% of answers traceable to the source document

Three architectural flaws explain this result:

Flaw 1 — Vector similarity does not understand regulatory taxonomy

RAG systems capture general semantic proximity. "Azoxystrobin" and "azoxystrobin-methyl" sit very close together in vector space — to a model trained on general text, these two substances look "similar".

For an AMM dossier, they have completely distinct toxicological profiles, approved uses and regulatory restrictions. Confusing the two in a pre-submission note potentially means a dossier rejected by ANSES.

Flaw 2 — Chunking destroys the structure of regulatory dossiers

An AMM dossier is a structured document. The use restrictions (maximum dose, growth stage, pre-harvest interval) are tied to a specific crop, itself tied to a substance within a precise geographical and temporal context.

When that dossier is split into 500-token fragments, these relationships are broken. A fragment may contain a restriction without its context of application. The model "sees" a fragment, not the structure.

Flaw 3 — The model fills the gaps with its general knowledge

When the RAG system does not find the information in the retrieved fragments, the language model does what it was trained to do: it generates plausible text. It can extrapolate from other European dossiers, from adjacent standards, from its pre-training knowledge.

The problem: this "completion" is indistinguishable from an answer grounded in your documents. There is no signal in a RAG system's output that separates a correct answer from an invented one.

The often deal-breaking obstacle: data confidentiality

Beyond the error rate, generic cloud solutions run into a second obstacle in agrochemistry. Your AMM dossiers under review, your internal studies, your proprietary formulations — these are strategic assets.

Submitting them to a cloud API, even an "Enterprise" one, means routing them through third-party servers, often outside the European Union. In the organisations we meet, this question systematically ends the conversation about cloud solutions: neither the Legal Department, nor the CISO, nor the Regulatory Affairs Department will authorise it.

The architecture that changes the game

The alternative is not "no AI". It is a different kind of AI, resting on a fundamentally different principle:

In a RAG system, the language model generates the answer. In a Knowledge Graph architecture, it generates the query — and it is the graph that produces the answer. The model can no longer hallucinate what it does not know.

Here is how it works:

  1. Structuring the corpus. Your documents (AMM dossiers, studies, E-PHY records, ANSES reports, ECHA data) are analysed and structured into a knowledge network. Each entity — substance, crop, use, dose, restriction, regulatory decision — becomes a node. Each relationship between these entities is typed and traceable back to its source.
  2. The model as translator. When a user asks a question in natural language, the model does not try to answer it directly. It translates the question into a formal query against the network.
  3. Factual extraction. The query is run against the knowledge network. The answer is extracted directly from the graph's nodes and relationships — not generated by a probabilistic model. Each result comes with its source: document, paragraph, version, date of last update.

A concrete example: [Substance: Azoxystrobin] → [APPROVED_ON: Grapevine] → [RESTRICTION: max dose 0.5 L/ha] → [SOURCE: ANSES Decision 2022-031, §4.2]

What this changes in practice for an agrochemistry R&D team

The operational benefits cluster around four recurring use cases we observe with our clients:

  • E-PHY regulatory search — a natural-language question returns, in 3 seconds, the approved substances, with up-to-date restrictions and traceable sources. What used to take 2 hours of manual lookup.
  • Inconsistency detection — the network can automatically flag contradictions between an internal dossier and an ANSES publication update. Without anyone explicitly searching for them.
  • Capturing expert knowledge — past analyses, business decisions and the knowledge of senior experts are built into the network. When the expert retires, their knowledge remains queryable.
  • Dossier preparation — the skeleton of an ANSES summary note can be pre-filled automatically from the network, with every source cited. The expert validates and adjusts; they no longer rebuild from scratch.

On the question of deployment time

One objection comes up frequently: "Building a knowledge network is a 2-to-3-year project."

That was true when each document had to be structured by hand. It is no longer the case. Current approaches make it possible to build a first operational network on a defined business corpus — one type of dossier, one product segment, one regulatory scope — in 8 to 16 weeks.

We prove the value on a concrete case. We then extend to other corpora.

If you are an R&D Director, Head of Regulatory Affairs or CIO in the agrochemistry or phytosanitary sector, we offer a confidential 30-minute demonstration. No generic pitch: you bring one of your real use cases, and we show how our architecture answers it.