While it draftsFree plugin · Available
TRUSTED
19 skills your assistant follows while it works. It uses only what you gave it and shows what is missing.
TRUSTED keeps your assistant to what you gave it. VERIFY checks every sentence against your sources. You decide whether to sign.
Head of Medical Affairs decides · not signed yet
Answers with an invented detail: a standard user, then the same user with TRUSTED. 288 answers per setup, on fictional pharma tasks.
Graded by AI models, not people. We built the tools and ran the test.
A free plugin your assistant follows while it drafts.
A check your assistant calls on the finished draft. 100 checks per key during the beta.
Draft one payer slide on how many patients stay on Velostra at 12 months, from the study summary and the label I attached.
"At 12 months, 86% of patients remained on Velostra."
In the real-world study, 14% of patients stopped treatment within 12 months.
Signed off by: not yet. This is a draft.
TRUSTED works while the draft is written. VERIFY works on the finished draft. Neither tool signs. The last step is always yours.
Notes, data, the label, the papers you supply
Uses only what you gave it and says what is missing
Gaps in plain sight; your decisions left to you
Each source, then each sentence, with a status and a record
Whether you can sign it and stand behind it
19 skills your assistant follows while it works. It uses only what you gave it and shows what is missing.
11 tools your assistant calls to check the finished draft, sentence by sentence, against your sources.
Free plugin. 19 skills that ask what a careful reviewer asks: is it useful, allowed, fair, safe, true, responsibly made, and can you stand behind it. It uses only what you gave it, says what is missing, and points you to your company's procedures.
Works out who will read it, which market, whether the product is approved, what you supplied and what review applies. Asks only what is missing, usually 0 to 3 questions.
A final read against every letter you have on: what to fix, what is missing, what a person must decide, and which internal process it goes to next. It never says a draft passes.
Explains TRUSTED, lets you turn letters on or off, and routes your request to the right skill.
Twelve habits, below, applied at the depth the work needs: a light touch for an internal email, the full set for a doctor, a payer or the board.
Specific instead of generic. Draws the "so what" for this reader and names the decision or next step. Never adds a figure you did not give; it asks or leaves a visible gap.
The right depth, words and format for a board, a payer, a doctor or a patient. Keeps meaning, terms and numbers exact in translation.
Keeps claims inside the authorized label or intended purpose, keeps scientific exchange apart from promotion, and keeps comparisons fair. Asks for the label when it is missing.
Checks which country's rules apply, which statements the market expects, and whether patient or personal details would appear in the output.
Shows risks and limitations with the benefits, in the same piece, and chooses studies, endpoints and time points fairly. Never hides a risk on request.
Certainty that matches the data, charts that look as large as they are, and sponsors and funding named.
Quotes dosing and warnings word for word from the label you supplied. Flags a possible side effect, pregnancy exposure or misuse to your safety process. Gives patients a route to a doctor, nurse or pharmacist.
Nothing invented: no facts, figures, names, dates, quotes or sources you did not give. Numbers that add up. Studies described as they were.
Decides which claims need a source and which kind of source can carry them. Shows a missing source as a gap, never invents a reference.
Before confidential, unpublished, personal or copyrighted material goes in, it asks what your company AI policy allows.
Who must read it before it is used, whether to disclose AI help, whose name it goes out under, and respectful language about patients.
A short record of the materials, the assistant and what the AI did. Assumptions and open items visible. Stays a draft until a named person clears it.
Fact-checks finished work sentence by sentence against the passages you supplied.
Says what a source can carry and what it cannot, before it goes in front of a payer, the board, doctors or the public.
Turns VERIFY's claim checks into the confidence line and the short Verify card that a reviewer reads.
Builds the reference list and the formal record for finished work, one citation per source from the source's own details.
Every fact, figure, name, date, quote and source comes from what you supplied. Anything else shows as a gap.
It pushes a vague draft toward specifics by asking you, never by filling them in itself.
It reports what to fix and what is open. It never calls your work compliant, approved, safe or true.
It points to your company's procedures and reviewers. They always come first.
A check against your sources. It first asks whether each source is fit for your reader, then checks each sentence against a quoted passage, lists what is still open, and leaves the sign-off to a named person. Run by hand without a key, marked "not tool-checked". With a key, they use VERIFY's tools.
Stores short passages from a PDF or page you supply, for your account only.
Grades a source for your question and reader, from its own details.
Says whether a paper was withdrawn, corrected or replaced.
Splits finished text into the sentences a reviewer could ask to see a source for.
Compares one sentence with a quoted passage and returns a status.
Splits and checks a whole text against the sources you name, in one pass.
Formats the reference from the source's own details.
A table of claim, passage, status and source.
Turns finished checks into the Verify card or a brief. It does not invent findings.
Files the record. Refuses without the name of the person who cleared it.
Shows the checks you have used.
A counted check is one call that checks or grades: a sentence against a source, a source for the reader, or a whole text. Usage and reference formatting are free.
Seven while the draft is written, six before it goes anywhere. You choose which TRUSTED letters are on; all seven are on by default.
3 core skills run every time, 12 apply the letters while drafting, 4 check finished work. Four rules hold for every skill: nothing invented, specific and never made up, flag and never grade, your process decides. How TRUSTED is designed to work; not a test result.
VERIFY runs each check itself and shows its work. It never clears its own work. How VERIFY is designed to work; not a test result.
Our own test of 96 fictional pharma tasks, 288 answers per setup. A standard user asks the assistant as most people do. A super user writes a careful expert prompt, the standard a skilled user can reach with effort. Most answers still need your read.
With VERIFY: +16.7 points [+7.6, +25.7] over the super user
With VERIFY: about half the super-user rate
Edits per answerStandard user 1.94 · Super user 1.59 · TRUSTED 1.57 · TRUSTED + VERIFY 1.32
Source: our own test of TRUSTED and VERIFY, October 2026. 95% intervals and all seven exhibits are in the article. A model of the effect, not a guarantee for your work.
You decide. Read what is open, look at the passages, then sign.
What it is, its grade, and whether it was withdrawn or corrected
Backed, partly backed, not backed, contradicted, or could not be checked, with the passage next to it
The share of sentences your sources back, partly backed counting as half. It is not the chance the text is true.
Every sentence it could not back, as an item for a person
Keep, reword, remove, or ask, sentence by sentence
The name of the person who cleared it, or "Not yet signed off. This is a draft."
"Backed" means backed by the source you supplied. It does not mean the sentence is true. An illustration of a session in Claude Cowork with TRUSTED and VERIFY, not a screen recording. The two quoted lines come word for word from our development rounds. Velostra, its figures, and the source scores and confidence figures on the card are fictional and have not yet been tested against a careful reader.
For medical, market access, brand and general management teams who already work in Claude Cowork.
A reply to a doctor that mentions a rash. TRUSTED does not judge it: "Your safety team decides that."
A payer slide that stays strong without a figure no source gave.
Key messages that stay inside the label, with risks next to benefits.
A board update where the gap in your notes stays visible, so you fill it with the real reason.
Your documents matter more than our features. Here is what happens to them.
Only the documents and passages you send it, and the sources you name. It does not search the web for a page that happens to agree.
Short passages and where they sit in the document, never the whole file. Each passage is at most 1,200 characters. Passages go back only to your own account and never into a shared index.
Uploaded documents are kept 90 days by default, and you can change that. They are deleted at the end of that period, or sooner on request. When you upload, you confirm you may process the document.
Check with your company's AI use policy before you send confidential material. TRUSTED will ask you to.
One key per person. If TRUSTED is installed, the same key serves both.
Free, from the Claude plugin marketplace. Then ask: "What can TRUSTED do?"
Paste the request above under your next draft request.
The payer slide, the reply to a doctor.
Installed and left alone, the assistant used TRUSTED in 1 of 546 sessions. Paste this under your next request:
Use trusted-intake before you start, and follow it. Write the draft. Then run trusted-check on your draft and fix what it finds. Then check it with VERIFY against my documents and include the Verify card.
Preparing for an MLR board? MLR Review is a separate service that works with both.