AI for Pharmaceutical MLR Promotional Review

What MLR promotional review is

Before a pharmaceutical company puts promotional material in front of a physician or a patient — a detail aid, a website, an email, a conference booth panel — it passes through MLR review: a coordinated sign-off by Medical, Legal, and Regulatory reviewers. Their job is to confirm the piece is truthful, not misleading, and fairly balanced: that every claim is substantiated by the label and the data, that risk information is presented alongside benefit, and that nothing implies an unapproved use.

It is careful, high-stakes work, and it is slow. A single asset can loop through reviewers multiple times, and the volume of promotional material a commercial team produces is large and rising. That is the pressure an AI assist is meant to relieve — without taking the judgment away from the people accountable for it.

Where the AI assist fits: it flags, you decide

The assist sits before the human reviewer, not in place of them. It reads the material and raises candidate issues: a claim that may need a citation, a place where fair balance looks thin, language that could read as off-label, a reference that does not match the source. Each flag is a pointer that says "look here," with the reason attached.

What it never does is close the question. It does not approve, reject, or edit. It cannot decide that a claim is compliant or that a piece is cleared. That decision belongs to a qualified Medical, Legal, or Regulatory reviewer, and keeping it there is the entire point — accountability stays where a regulator expects to find it. The assist makes the reviewer faster and more consistent; it does not become the reviewer.

The workflow: Flag → Judge → Prove

1. Flag The assist surfaces candidate issues. It marks the spots in the promotional piece that warrant a reviewer's attention and states why each was raised. This is a shortlist, not a verdict.
2. Judge A human reviewer decides. The Medical, Legal, or Regulatory reviewer accepts, dismisses, or acts on each flag using their own expertise. Every decision is theirs, recorded against the specific item.
3. Prove The record stands as evidence. Each flag, and the human judgment on it, is anchored to a tamper-evident, cryptographically signed audit chain. Later, "why was this flagged and who cleared it" has a verifiable answer, not a reconstruction.

No LLM inside the flag decision plane

This is the design line that matters most, and it is worth stating without hedging: there is no LLM inside the flag decision plane. The logic that decides whether something becomes a flag is deterministic and inspectable, not a language-model guess. A model may help draft context or organize material around a flag, but it never sits in the path that produces the flag itself.

The reason is trust. If a language model decided what to flag, a reviewer would be auditing an opaque output — unable to say precisely why an item was raised or reproduce the same result twice. By keeping the decision plane free of any LLM, every flag stays explainable and reproducible: you can trace the rule that fired and get the same answer again. That is what lets a reviewer, and an inspector, actually trust the flag.

The fuller mechanics of this separation are laid out in the Flag → Judge → Prove workflow.

Built on four verifiable controls

The assist does not ask you to trust the environment it runs in. Hyfstele is a secure-LLM deployment that runs open-weight models inside your own perimeter, air-gapped, with four controls each independently verifiable on your hardware:

1. Weights proven byte-identical to the public artifact — the model you run is the model you audited.
2. Dormant / latent capacity enumerated — unused capability is accounted for, not assumed away.
3. No egress — no route to the internet and no name to resolve. Things go in; nothing comes out. Your promotional material and unreleased claims never leave the perimeter.
4. Every inference call anchored to a tamper-evident audit chain, signed with post-quantum cryptography (ML-DSA), so each flag and each human decision is verifiable evidence.

Together these turn "trust our assist" into "check our assist." A reviewer can confirm the model, confirm nothing left the room, and confirm the record is intact — on their own hardware, without taking anyone's word for it. See the signed, auditable AI control for how the audit chain works.

Relevance to FDA promotional review and Part 11

FDA's Office of Prescription Drug Promotion (OPDP) holds promotional material to a clear bar: truthful, non-misleading, and fairly balanced, with claims consistent with the approved labeling. MLR review exists to meet that bar before anything ships. An AI assist supports that human review — it helps reviewers catch the things OPDP would — but it does not replace the human judgment OPDP expects to see behind the decision.

On the records side, 21 CFR Part 11 requires secure, computer-generated, time-stamped audit trails that are attributable and protected against silent alteration. Because every inference call in Hyfstele is anchored to a cryptographically signed audit chain, the AI portion of the review carries an evidence trail that meets the same standard: each flag and the reviewer's judgment on it can be reconstructed and independently verified, which also supports GxP, data-residency, and model-provenance expectations. See AI in FDA promotional review for the fuller mapping.