Deterministic quoting

What is deterministic quoting?

Deterministic Quoting is the mechanism that stops a generative agent from paraphrasing approved content into something unapproved.

Most agents treat approved claims as raw material. They rephrase, compress, or blend them with other text, and the output drifts from what was actually cleared. Deterministic Quoting removes that step. Approved text reaches the patient or HCP exactly as written. Every instance is verified against the source document at the point of output and cited to its page.

Why pharma has never used conversational AI

Conversation is the best engagement format there is. It's the only one built for true personalization. But generative conversation means improvised wording, and no review committee approves improvised wording in advance. So brands ship one static message to everyone. The patient or HCP with a real question goes elsewhere for the answer.

The two choices brand teams have known

Until now, brand teams had two options.

The first is ungoverned generation: A general-purpose model answers freely. Engagement goes up, control goes down. The model can word a claim in ways nobody approved, drop required safety information, or drift off-label. A committee cannot approve an output it cannot predict.

The second is the walled garden: You get control back by confining the AI to a single destination with a fixed content set. It's safer, but the brand loses channel flexibility. Often it can't even show its own ISI or approved claims. The engagement gets rented from someone else's platform.

A conversation that adapts in form while its content stays fixed

Deterministic Quoting pulls every answer word for word from the brand's MLR-approved content library. The model handles understanding and navigation. It works out what the person is asking and locates the approved content that answers it. It writes none of the medical or promotional content itself.

That way, the conversation adapts in form and the content stays fixed. The patient or HCP gets a natural, responsive exchange. The brand gets an output where every claim already has been reviewed.

What we've built is the version engineered for MLR workflows: CMS-loaded content, ISI handling, on-screen marking, adverse-event routing, and a full audit trail.

How it works

Source of truth: Your MLR-approved content lives in your CMS: approved claims, indication statements, ISI, and answers to anticipated questions.

Retrieval and navigation: When a patient or HCP asks a question, the AI reads the intent and picks the approved content that answers it. This is where the conversational intelligence sits.

Deterministic surfacing: The AI shows the picked content and marks it as deterministic, so the reader and any auditor can see exactly which words came from approved content. Required elements like ISI appear where the rules demand.

Content loads straight from your CMS, so the moment you update an approved claim, the conversation shows the new wording. There is no second copy to drift out of sync, and no separate review cycle for the AI.

What MLR can review

• Predictability: Every claim is pre-approved. The AI cannot write a new one, and content that has not cleared review cannot appear.

• Traceability: Each line the AI shows maps back to a specific approved asset in the CMS, so a reviewer can open any conversation and see where every line came from.

• Visible marking: Deterministic content is labeled on screen. There is no ambiguity for the patient, the HCP, or an auditor about what is sourced.

• ISI and fair balance: Required safety information appears according to the rules, because the brand controls both the content set and the logic that shows it.

• One source of truth: When approved content changes, the change carries through on its own. There are no outdated claims and no parallel approval queue.

What's quoted, what's generated

The AI quotes the approved corpus word for word: claims, indication language, dosing, safety information. The framing around it, a greeting, a transition, a clarifying question, is generated within pre-approved bounds. Framing carries no medical or promotional content. It also looks visually distinct from quoted content, so the line between the two is inspectable in every exchange.

When the corpus has no answer, the agent says so. It doesn't guess. This is the behavior clinicians already trust in their daily evidence tools.

One mechanism across every channel

Deterministic Quoting isn't tied to a destination. The same engine, drawing on the same approved corpus, runs across SMS, WhatsApp, email, web, and voice, on the channels you already buy. One content set. Update it once, and the change carries through everywhere.

The examples on this page use US frameworks: MLR, OPDP, 21 CFR 202.1. The mechanism itself doesn't depend on jurisdiction. Entering a new market means loading a different approved corpus under that region's rules.