Can’t ChatGPT just do that?

Question 1

“Can’t ChatGPT just do that?”

It can, and it did, in twenty-three minutes. What it returned, and what happened to the form on the way, is the material for the rest of this page.

ChatGPT Pro, the newest model, with both files attached and a plain request: keep the template, put each answer in its field, cite the CSR page, and say so where the CSR has nothing. Twenty-three minutes later it returned a Word file that reads as complete. Two things about that file matter for a submission. Every field in it was written by the model and reviewed by nobody, so there is no record of who stands behind any figure. And the agency’s template was altered on the way: table captions became headings, and the file grew from 50 pages to 71.

Question 2

“We would review it anyway. We always do.”

Reviewing afterwards means reading the CSR again to check page numbers that were typed for you. The part of the work worth automating comes before the decision: putting the right passage in front of the person who makes it.

Palmera does not produce answers to be checked afterwards. It offers suggestions with the source passage, page and section beside them, and a person accepts each one or leaves it. What is accepted keeps its evidence, and the fact of its acceptance, inside the file. The review is the output.

What makes this checkable is how the answer is produced. Autofill does not write. Its answer is a passage quoted verbatim from the filed source, and each passage is looked up in the source before it is offered; the proof page counts the exceptions. The model’s part is limited to choosing which passages fit the field. It cannot introduce a figure or a claim the source does not contain, and two runs on the same document differ only in which passages they choose (that too is counted on the proof page), never in inventing one.

Question 3

“You prompted it badly. A better prompt, or Copilot in Word, would keep the form.”

The prompt that would keep the form already exists. The agency wrote it, field by field, into the form itself. ChatGPT treated that text as content to replace; Palmera treats it as the question.

We tried a better prompt: a follow-up in the same chat naming exactly which fields had been missed and what was wrong. It changed 43 fields and added 162 of its own to the agency’s form. A model given a document edits the document; it has no notion of what a field is for unless something tells it. Feeding it the form a section at a time would protect the layout, but would not add a page number that can be checked, or a record of who accepted what, which is what the rest of this page is about.

In Palmera the instruction in each field is kept as the question put to the sources for that field. The people who set up the template extend those instructions with what the form does not spell out: the current methodological standards and agency requirements, including equity and the other elements of the ISPOR value flower, and the conventions of economic evaluation. That is written once, for every submission on the form, not typed on the day.

Question 4

“It only knows what I put in the prompt.”

ChatGPT works from the sentence typed into the chat. Palmera holds the PICO, the filed sources and each field’s question before it reads a page.

The PICO is entered once and belongs to the project; every field is answered in the light of it. Each source, a CSR here, a product monograph, a global value dossier, a systematic review or an economic report elsewhere, is uploaded, classified, prepared into quotable passages that carry their page number, and stays in the project. In the chat, the context was one sentence and two attached files.

Question 5

“Quotes are not a submission. I need prose, and tables.”

A verbatim quote is the one thing an agency cannot say was made up. Prose is made from it, and Palmera checks its own prose against the quote and shows the differences before it is kept. ChatGPT returned prose with page numbers and no quotes.

Tables come back row by row: the source’s rows, figures intact, offered for the agency’s table. Asked for prose, Palmera rewrites the field and then checks its draft against the original, figure by figure and protected term by protected term, before the text goes in. A changed number shows as a warning at that moment rather than as a clarification question months later.

The rewrite is contained by construction. Summarize is given the field’s own accepted content and nothing else: not the source, not the internet, not the rest of the form. It cannot bring in a fact the field does not hold, because it never sees one. What it can do is change wording, and that is what the check after it is for: every figure, every protected term and every reference in the draft is compared with the field, and the writer sees the differences (5b) before deciding.

Question 6

“What if I need to fill from multiple documents?”

A submission draws on several documents, and a CSR cannot answer half of the form. Each source is filed once, the fill is pointed at one source at a time, and what a source cannot answer is left as the agency wrote it.

Submission date, Health Canada NOC and indication, list price, provincial prevalence, the systematic review, the second pivotal trial: none of these is in a CSR. Palmera says so on the field, “No supported answer found in this document”, and leaves the agency’s guidance in place, so what remains to do, and from which source, stays visible. ChatGPT wrote “not reported” over them.

The next source is then filed: the product monograph, the global value dossier (GVD), the second trial’s CSR, the systematic review, the economic report. The fill is pointed at it and run again. Answers already accepted keep their evidence, and the PICO and the template are not entered again.

Beta note: on this run, suggestions were offered for second-study columns from the pivotal trial, and in a sample read by hand more than half of the suggestions for this trial’s own fields quoted the right CSR but the wrong table. The passage and its section sit beside each suggestion, so this is visible and the suggestion is left out in a click. Scope and table matching are being tightened; the counts are on the proof page.

Question 7

“And when INESSS wants the same evidence next month?”

Another form on the same project: the same PICO, the same sources, every answer traceable the same way.

A second deliverable is created from the next agency’s template inside the same project. The sources and the PICO are already there, and each answer is traced to its passage the same way. In a chat, next month is a new conversation and a new file to reconcile with the last one by hand.

Question 8

“Is your model as good as GPT?”

Palmera has no model of its own. It runs on the commercial models, and the answers on this page came from Google’s.

This case ran on Google’s models: Gemini 2.5 Flash for Autofill, Gemini 2.5 Pro for Summarize, which the dialog in 5b names. Earlier runs used OpenAI’s models, and Anthropic’s are supported. The model is asked to do narrow things. In Autofill it helps phrase the search for a field and chooses which of the retrieved passages answer it; the passages it chooses are quoted verbatim and looked up in the source before they are offered. In Summarize it rewrites the accepted passages of one field into prose, and the draft is checked against those passages, figure by figure and term by term.

The rest is fixed logic that does not change with the model: reading the form’s questions, preparing the source into passages that carry their page, the verification of each quote and each figure, and the record of who accepted what. So the choice of model affects how often a good passage is found for a field, which the counts on the proof page measure. It does not affect whether an answer can be checked, which is what a submission needs.

Question 9

“A submission is a team. Medical, market access, the agency, an outside writer. How does that work?”

A shared workspace shares files and conversations. A submission also needs its decisions kept at the level of the field, and its outside contributors let in for one project at a time.

A ChatGPT Business workspace with SharePoint connected gives a team shared files with version history, and a shared chat that can search them. What it does not hold is the decision on a field: which passage a reviewer accepted for Recommended dosage, who accepted it, on what page, and what medical said about it afterwards. That is what an agency asks for months later. In Palmera the writer accepts a suggestion with its passage beside it (9a); a reviewer who disagrees says so on the field (9b); and the passage, the page, the account that accepted it and the time are written into the exported file itself (9e).

The second difference is who can join, and for how long. Submissions are done with outside help: a global team hires a consultancy for one product launch; a regional team hires one firm for the first draft of a dossier and another for its final revision. In Palmera each of them is an editor on one project and nothing else (9c), added and removed without touching the company’s own document store. Regional teams can share a project, so evidence prepared for one agency is not rebuilt from scratch for the next.

Within a project everyone works in the one form, with the access their role allows, and each save, export and comment is logged by name (9d). The model has no part in any of this.

Same two files, both sides: the CDA-AMC template as issued and the STEP 1 CSR, one source on purpose so that every answer could be checked. Palmera was driven through its normal browser screens, click by click, nothing scripted. Every field in every returned file was compared with the blank template, and every quoted passage was looked for in the source word for word. Nothing here is a mock-up; 9e is drawn from the exported file’s own record. The files and the checks are on the proof page.