Our AI approach

AI agents harnessed for label compliance

A regulatory decision has to be defensible months after it was made, by someone who wasn’t there. That rules out anything that cannot show its reasoning. Our agents are constrained on purpose: deterministic where determinism matters, domain-grounded where judgment is needed, and logged everywhere.

Six conditions

What an agent has to satisfy before we let it touch a label

Technical skills

Validated tools, frameworks and workflows

Domain credibility

Relevant, representative and fit-for-purpose data

Contextual judgment

A domain mind map behind every informed decision

Reproducibility

Consistent, testable and validatable results

Traceability

Full output logs and audit-ready lineage

Audit trails

Documented evidence of conformity

Designed against

EU AI Act21 CFR Part 11GAMP 5EU Annex 22

The what, how and where

Zero black box, in a place you control

Three architectural choices do most of the work.

What

Deterministic orchestration

The sequence of checks is fixed and inspectable. Models are asked narrow questions inside a defined pipeline, never handed the whole decision.

How

Domain mind map

Regulations, claim types, ingredients, markets and precedent held as an explicit map of the domain — so context comes from structure, not from a prompt.

Where

Small language models, on device

Small enough to run inside your environment. Unreleased labels and claims never have to leave a place you control.

The result

An output a regulatory reviewer can sign their name to

Predictable, repeatable and explainable outputs

Documented decision logic

Source-traceable data lineage

Decision maps

Every verdict is a path you can walk back

A claim is not judged in one step. Each branch is recorded with the clause it turned on, so a reviewer can see exactly where the verdict was decided — and disagree with that step specifically.

Claim under assessment

“Clinically proven to support joint mobility”

Conditional — evidence required

Claim type identified — function claim, not disease

Language refers to maintaining a normal bodily function. No reference to treating or preventing a condition.

Permitted list checked per market

Equivalent wording appears on the permitted register for two of four target markets.

“Clinically proven” triggers a substantiation requirement

This is the branch that decides the verdict. The qualifier raises the evidence standard above the claim itself.

Verdict, with two routes out

Attach the study that meets the standard, or drop the qualifier and the claim clears three markets outright.

Every node carries its clause, its source and the date it was read.

Illustrative decision map. The same structure applies to a pharma promotional claim and a food front-of-pack claim.

100%

Findings carry a citation

A regulatory decision has to be defensible months after it was made, by someone who wasn’t there.

Our differentiator

Why a general model is the wrong tool for this

A general-purpose model

Reads the clause as prose and paraphrases its gist

Gives a different answer to the same question next week

Cannot say which source produced the verdict

Sees a phrase, not the impression the whole label leaves

Labelex^® agents

Rules extracted atomically — the modal verb, the conjunction, the exception

Deterministic runs — the same inputs give the same verdict, and it is testable

Clause-level citation and data lineage on every finding

Overall perception judged as a whole, market by market

Eighteen months of regulatory research

The rule extraction methodology came first; the agents were built to run it.

The human keeps the decision

Agents prepare and evidence the assessment. Accepting or overruling it stays a person's job, and that choice is logged too.

Runs where your data lives

SaaS or agents over MCP inside your own environment, alongside your DMS, RIM and QMS.

Ask us the hard questions about the model

Talk to us