Questions we hear most.

Today: Clinical Study Reports (CSRs), Periodic Safety Update Reports (PSUR / PBRER), and Clinical Evaluation Reports (CERs). Each module type ships with its own writing rules — ICH E3 for CSRs, ICH E2C(R2) for safety reports, MEDDEV 2.7/1 Rev 4 for CERs. On the near-term roadmap, built on the same engine: DSUR (ICH E2F) and CTD Module 3 / CMC (ICH M4Q).

No. Asthra removes the mechanical part of writing — assembly, copy-paste, cross-section consistency, table summarization — so writers spend their time on judgment, narrative quality, and review. Every output requires writer review and approval before it ships.

In two places, with one account: inside Microsoft Word as a task-pane add-in, and in the browser as a web app. Writers who live in Word don't switch tools or learn a new authoring environment; teams who review together can co-edit in the web app in real time, with comments and section locks. The drafting agent runs in your data perimeter (managed SaaS or your own VPC); see Trust center for details.

Yes — and you should. Asthra writes into your house templates, preserving section numbering, headings, table-of-contents structure, and styles. Onboarding includes a one-time mapping of your templates to module-specific writing rules.

Both. You can hand Asthra a partial draft and ask it to fill specific sections, refine an existing narrative, or update a section against new data — all with the same closed-system retrieval and ledger that produces a fresh draft.

It is the single surface, in the Word add-in and in the web app, where every step of regulatory drafting happens — source ingestion, plan review, drafting, refinement, citation review, QC review, reviewer handoff — in one continuous conversation tied to one document. It is not a chatbot bolted onto Word; the section overview, approval gates, and audit trail are first-class UI, not buried in a chat transcript. See it on the product page →

A background-execution mode where the writer initiates a run, steps away from the laptop, and returns in 1–2 hours to a first draft with a bibliography and an end-of-run QC report ready. When the writer reopens the document, Asthra loads the full agent history — every retrieval, every tool call, every gap flag, every QC finding — so the review picks up natively where the agent left off. It is not autonomous AI submission; the draft is a starting point for the writer's review, not the finished deliverable. Read about walk-away drafting →

Most of the product is for this stage. Each section arrives already checked against its guidelines and FDA precedent. You can resolve flags one at a time or in a pass over many sections, regenerate a section with steering, rewrite a highlighted passage, ask the draft why it reads the way it does, freeze and compare named versions, and review it with colleagues. In the Word add-in you can also select any text to see the source passage behind it. Every change Asthra makes comes back as a tracked change or a before-and-after diff that you accept or reject. The full flow is on How it works.

Every flag is a tracked to-do: there is an Issues list grouped by section, with a count of what is still open. On a single flag you can ask Asthra to fix it, write the text yourself, or dismiss it, and you can name the source documents Asthra should look in first. The fix comes back as a before-and-after diff to accept or discard. To clear many at once, start a resolve pass: Asthra rewrites only the flagged paragraphs and queues each as a proposal, and nothing is applied until you approve it. Dismissed flags are counted apart from resolved ones, and regenerating a section marks its flags superseded, never resolved. More in Asthra resolves the flags it raises.

A section-by-section FDA/ICH audit that runs right inside the document. On demand, Asthra checks a drafted section against curated ICH E3 expectations (each carrying its own severity — critical, major, or minor), grounds its judgement in real approved-product labeling pulled live from openFDA, and writes specific regulator flags inline as [REGULATOR FLAG] markers — anchored to the exact paragraph they apply to. It recommends, never rewrites: the output is advice with an anchor, and the writer keeps control of the words. The openFDA precedent is used only to judge what regulators expect to see, never to borrow another product's numbers. It is not a chatbot persona role-playing an inspector — it's a deterministic audit pass. (Configurable reviewer-persona scorecards are a separate capability, on our roadmap.) Read Regulatory Review, under the hood →

Yes. The data analysis agent accepts natural-language requests (“Grade 3+ AE table by SOC and treatment”) and generates the analysis as deterministic Python code that runs against your line listing, sales / exposure data, or lab parameter file. The output is an aggregated table or a figure (bar chart, forest plot, time-course) ready to insert into the draft. Every analysis step is logged with the exact code that produced the output. It is not an LLM doing arithmetic — the LLM proposes the analysis plan, Python executes it, so the numbers are reproducible. It does not replace the statistician for pre-specified primary or secondary analyses. Read about data analysis →

From inside the chat, ask the agent to find recent papers on a topic for a specific section. The agent queries PubMed, ClinicalTrials.gov, and bioRxiv in parallel, deduplicates by DOI / PMID, flags preprints, and returns citation-ready results. It is not an autonomous web crawler — every external lookup requires explicit writer approval before fetching, and every approval is recorded in the audit trail. Read the literature search blog →

Reviewer personas are on the roadmap. They are reviewer perspectives a writer will summon over a section to return structured findings — a rubric anchor (e.g. ICH E3 §11.4.1), the snippet that triggered the finding, severity, suggested action. The planned library: FDA Clinical, CMC, CER Device, CSR Writer QC. They will not replace human review; they are tools to pre-empt reviewer feedback. What ships today is the automatic regulatory review, which checks each section against its guidelines and FDA precedent as it finishes drafting.

Five structural questions: cross-reference integrity, statistical-claim source binding, defined-term ordering, citation density per section type, and gap detection. The result is a structured QC report with pass / flag / gap per check, located down to the section. It is not a copy edit — it does not check prose quality. For prose review, summon a reviewer persona over the section. Read about QC + run bundles →

An audit-ready package shipped with every run: the .docx itself, a per-section provenance projection mapping every claim to its sources and transactions, the QC report, and a run manifest with model, code version, token usage, and reproducibility metadata. It is not tied to Asthra's portal — the bundle is portable, vendor-independent, and survives the vendor; you can hand it to an inspector as-is. Read about run bundles →

Three mechanisms compound. (1) Every draft sentence is bound to a retrieved passage — the model can't generate a factual claim without a citation handle. (2) Retrieval is closed: there's no fallback to pretrained knowledge for facts. (3) Missing data triggers explicit gap flags instead of plausible-sounding text. Hallucination becomes a deterministic ledger entry, not a soft risk.

Document-level citations attach to every section — every section says which source documents it draws from. Sentence-level citations are available on demand: click any sentence and see the file name, page number, and exact text snippet that supports it. The graph also supports blind-review workflows where citations are hidden from QC reviewers until they request them.

Asthra surfaces it explicitly, inline, as a gap flag — and writes around the gap rather than papering over it. The flag is recorded in the ledger so reviewers see exactly which sections need additional source material before submission.

No. Customer documents are never used for model training, by Asthra or by anyone else. Drafting runs through Anthropic and OpenAI under zero-data-retention agreements: nothing is stored after the response, and nothing is used for training. Each project's source set is isolated to that customer.

Models from Anthropic and OpenAI, both under zero-data-retention agreements. In customer-VPC mode, Anthropic Claude can run through your own Bedrock or Vertex deployment. We track frontier model releases and validate Asthra against each one before promoting.

SOC 2 Type 2 attested (report available under NDA). GDPR and HIPAA compliant. ISO 27001 ready; EU AI Act in progress. Full status on the Trust center.

Managed SaaS: isolated project storage on Amazon Web Services in the US-East (Virginia) region. That is the only region for now; if you need EU hosting, email info@asthra-writer.ai and we will arrange it with you. Customer-VPC: data never leaves your AWS / Azure / GCP account — Asthra deploys into your tenant. Encryption: TLS 1.3 in transit, AES-256 at rest, customer-managed KMS keys in VPC mode.

Retention is customer-controlled. Source documents stay in your library until you delete them, and drafts stay with their document until you delete it. Account data is deleted within one month of closing the account. Audit ledgers persist for the lifetime of the document and travel embedded in the .docx.

Yes. Our US deployments are HIPAA-compliant and we'll send our standard BAA alongside the security packet.

Asthra is deployed as a high-risk AI system under the AI Act framework. We maintain technical documentation, conformity assessments, and risk-management documentation aligned to Annex III. Available on request.

Managed SaaS: same day after security review. Customer-VPC: 1–2 weeks for the first deployment, including Terraform install, KMS configuration, and Microsoft 365 add-in publishing. Security review runs in parallel.

We support source-document ingestion from Veeva Vault, SharePoint, S3, and eTMF systems via API and connector. Drafting still happens in Word or the Asthra web app, but source documents stay in their authoritative location.

SAML 2.0 and OIDC for SSO. SCIM for user provisioning. Role-based access control with separate roles for writer, reviewer, QPPV / qualified evaluator, and admin.

Yes. The ledger is embedded in the .docx as a side-channel attachment, and can also be exported as JSON via the API. The format is documented and stable across versions.

30 days, scoped to one document type and one product / trial / device. We do the heavy lift on integration, template mapping, and writing-rule tuning. You bring real source documents and a real deadline. At the end, we benchmark Asthra's draft against your existing process — section by section.

One price per document, no subscription, no seats. A clinical study report or trial protocol first draft is $1,999, or $199 a section, delivered within 15 days of project start. The full list is on the pricing page. Teams that want shared libraries, custom document types and an SLA are billed under a team agreement, which we scope on a call.

You do. Drafts produced with Asthra are your work product, no different from documents written manually. Asthra holds no rights to your content, citations, or audit ledgers.

Yes. Several pilot partners are mid-tier CROs serving sponsor clients. CRO deployments support sponsor-isolated projects so a CRO writer can serve multiple sponsors without data crossover.