A cited first draft of every health hazard evaluation, with the gaps declared and the conclusion left to the people who sign
A cited first draft of every health hazard evaluation in half the hours, with every gap declared.
See one case, screen by screen ↓A complaint trend crosses a threshold and the question lands on Medical: how much harm could this cause? Answering it means reading the complaint summary, the risk management file, the design verification, the supplier’s 8D, the clinical literature, the field data by lot and the instructions for use — and writing injuries to date, populations at risk, seriousness, likelihood and consequences into the HHE form, one section at a time.
Every line has to stand up, because Medical, Quality and Regulatory all sign it and the recall decision rests on it. So the hours go on finding the page that supports each sentence, and on the numbers that are not there yet — the distribution figures for the latest lots, the supplier’s root cause. Under time pressure, the temptation is to reason around a missing number instead of saying it is missing.
The working solution’s dashboard figures for hours per evaluation by phase on its sample cases — 26 hours before the agents, 11 with them — shown here in 8-hour working days. Review time goes up on purpose.
Five specialist agents gather, extract, map, draft and check the evidence; Medical, Quality and Regulatory sign.
Complaint trending flags 37 complaints of a loose or detached needle shield on a prefilled syringe. Here is what happens next, screen by screen, in the working solution.
Maya opens the solution to 7 open evaluations, 4 of them due in the next two days, and 3 waiting for her Medical sign-off. Two gaps are declared across the open cases — “missing evidence, never guessed”. At the top of her list: HHE-118, eight documents received, the agents ready to read the set and draft.
Syringe needle shield detachment on the Veltrimab 150 mg/mL prefilled syringe, due in three days. Complaint trending opened it from 37 complaints on lots 26F11 to 26G09, and Dana Okafor attached the field data extract. The document set holds 8 files and 77 pages: the complaint summary, the risk management file (rev F), the design verification of shield pull-off force, the supplier’s 8D on the shield moulding, two literature papers, field data by lot and the instructions for use.
“The agents extract only what the documents say, cite every sentence to a page, and declare what is missing. They do not write the health hazard conclusion or the recall class.”
Each file turns from Waiting to Reading to a count of the facts taken from it. The run is shown step by step: intake reads the case as one set, the evidence extractor finds 4 harms and 20 facts, each with its page, the section mapper lays them onto the 11 sections of the HHE form, version 4, and the grounded drafter writes only from what is cited.
“Draft ready — 4 harms, 20 cited facts, 9 of 11 sections drafted.” Every sentence carries a citation. The completeness checker has declared one gap — the field failure rate cannot be computed — and left section 11 for the signers. Maya’s slot on the sign-off bar now reads “Your signature”.
“100 % of sentences carry a citation · 1 gap declared instead of guessed · conclusion left for the signers”
The evidence matrix lists four harms: needlestick injury (2 – Minor, 3 reported), injection-site infection in patients on immunosuppressants (3 – Serious, none reported), a delayed dose with loss of disease control (2 – Minor, 6 reported), and injection-site pain or bleeding from a bent needle (2 – Minor, 1 reported). Each severity is read from a row of the hazard analysis, and every number opens the exact passage — step 2 of the instructions for use, highlighted, for the infection risk.
“Severity is never estimated by an agent — it is read from the risk file row and cited.”
Product data, problem definition, injuries and events reported to date — each section under its reference, 21 CFR 7.41(a)(1) for injuries. Section 2 states what the supplier found: shields from moulding cavity 7, after a tool change on 2 Jun 2026, pulled off at 3.2 to 7.9 N against a 12 to 35 N acceptance criterion — and cavity 7 makes one in eight shields. Every section shows how many sentences it has and how many are cited, with Edit and Redraft beside it.
The draft gives the historic rate — 0.01 complaints per 1,000 units before June 2026 — and the rate it can compute: 28 complaints over 41,800 units, 0.67 per 1,000, for three of the five lots. Units distributed for lots 26G02 and 26G09 are missing, because the field data extract stops at June production. So the rate across all five affected lots is not written.
“Not written — Field failure rate cannot be computed. … The agents will not estimate it.”
Maya presses “Request from Supply chain”. Dana Okafor is asked for the distribution figures for the two lots, and the request is logged in the case activity. The gap stays declared, and blocks the Quality sign-off, until the figures arrive — then the agents read only the new file and draft section 7.
“Request sent to Dana Okafor — the gap stays declared until the data arrives”
The signing window states what the signature means: the harms, severities, populations at risk and consequences in sections 3–6 and 8 are supported by the cited evidence. Maya writes the medical assessment for section 11 herself — “written by you, never by an agent” — and re-enters her password. A notice reminds her that section 7 has a declared gap: her signature covers the medical sections; Quality signs after the gap is closed.
Her assessment sits in section 11 with her name and time: the evidence supports a remote probability of a serious harm, no serious injury reported to date, signed pending the field failure rate. The recommended class is left for Quality and Regulatory to write when they sign. The sign-off bar moves to Sam Patel for Quality, with Omar Haddad for Regulatory last.
“The recommended class is written by Quality and Regulatory when they sign. Agents do not propose it.”
The same evaluation, seen by Medical, Quality, Regulatory and Supply chain — what their week looked like, and what it looks like now.
Five specialist agents gather, extract, map, draft and check the evidence; Medical, Quality and Regulatory sign.
Assembles the case as one document set — complaint, risk file, verification, supplier investigation, literature, field data and instructions for use — lists what is missing and requests it from its owner.
Extracts each harm, hazardous situation, population at risk and severity, with page-level citations.
Maps cited facts to the 11 sections of the HHE form, following 21 CFR 7.41, and reports sections with no facts.
Writes sections 1 to 10 only from their mapped evidence, in plain, precise language — every sentence carries a citation.
Checks citation coverage and template completeness, and declares what is missing — naming the data needed and who owns it — instead of reasoning around it.
Stay accountable, all three sign. The agents propose; a named person decides.
Signers and contributors can ask about the evidence in any case, request missing data, or add a drafting rule. It answers from the cited evidence and does not write medical conclusions or recall classes.
What evidence supports the severity of the needlestick harm?
Needlestick injury (H1) is rated 2 – Minor. The rating is copied from the risk management file, row HZ-14 — not estimated. It is consistent with the complaints: three finger punctures, first aid only. Patients with reduced grip have a 2.4-fold higher rate of needle contact, which raises likelihood, not severity.
Which sections of HHE-118 are still missing evidence?
One gap is declared: section 7, likelihood. Units distributed for lots 26G02 and 26G09 are not in the field data extract, so the complaint rate for the affected lots cannot be computed. Section 11 is reserved for the signers. Everything else is drafted and cited.
What recall class should HHE-118 get?
That decision belongs to the signers — the agents do not propose a recall class or a medical conclusion. What it can give you is the evidence they sign against: highest severity 3 – Serious (infection, none reported), three minor needlesticks, and a likelihood that waits for the distribution figures.
Always flag cases where distribution data is missing
Added a drafting rule: “Flag missing distribution data at intake”. From the next case, the intake agent checks units distributed for every lot in scope and declares the gap before drafting starts. You can switch it off in Settings.
13 screens from the working solution, on its sample data. Pick one to see it large.
How the agents draft, who signs and in what order, and the template every evaluation follows are settings.
| Setting | Default | Choose from |
|---|---|---|
| Draft only from cited evidence | Always on | Locked on — a sentence without a page citation is not written |
| Leave the conclusion and the class to the signers | Always on | Locked on |
| Copy severity from the risk management file | On | On · Off |
| Block signing while a sentence is uncited | On | On · Off |
| Search the literature when a population has no source | Off | On (adds up to 5 papers for review) · Off |
| Signing order | Regulatory last | Regulatory last · Strict order · All at once |
| Evaluation template | HHE form · version 4 (11 sections) | Version 4 (11 sections) · Version 3 (9 sections) |
| Evaluation due | 5 working days | Working days from case open to the last signature |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample cases · “target” = the design goal, measured in the live solution · The working days by phase are the working solution’s dashboard hours on its sample cases, at 8 hours a day · Sources: 21 CFR 7.41, health hazard evaluation. People and products named on this page are characters in the working solution.
It is the assessment of how much harm a product issue could cause, used to decide whether and how to recall. FDA’s recall rules (21 CFR 7.41) list what it weighs: injuries already reported, conditions that could lead to a hazard, the populations at risk, the seriousness and likelihood of harm, and the immediate and long-range consequences. The solution drafts each of these as a section of your HHE form.
Five agents work in order. The intake agent gathers the document set, the evidence extractor pulls out each harm with its page, the section mapper lays the facts onto the form, the grounded drafter writes sections 1 to 10 only from cited evidence, and the completeness checker confirms coverage and declares what is missing.
No. Section 11 — the health hazard conclusion and recommended class — is reserved for the signers, and that rule cannot be switched off. Asked for a class, the assistant gives the evidence the signers sign against instead.
The section is not written. The completeness checker declares a gap, names the data needed and who owns it, and the request goes to that person. When the file arrives, the agents read only the new file and draft the section; Quality signs once each gap is closed or accepted.
From your risk management file. The evidence extractor copies each harm’s severity from the cited hazard analysis row, on the ISO 14971 five-level scale, and never estimates one. The citation opens that row.
Yes. Medical, Quality and Regulatory each sign electronically, with the meaning of the signature stated and the password re-entered. Signers can edit or redraft any section, and every action by people and agents is kept in the case activity.
The complaint or issue description, the risk management file, design verification and supplier investigations, clinical literature, field data by lot and the instructions for use — whatever is linked to the case or its product. Add a file and the agents re-read only what changed.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your HHE form, your risk management files and a set of past cases — and it goes live once every quality gate has passed. We will walk you through it on one of your own past cases first.
Any batch-record question answered in seconds, cited to the exact page and step.
1-min filmValidation DocumentsValidation protocols and reports drafted in minutes, every value traced to its source page.
1-min filmAsk the PlantBatch analysis that took three months now takes hours, with the query behind every answer.
1-min filmWe’ll run Health Hazard Evaluation on the document set of one of your own past cases.