Questions we hear most.

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

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.

Inside Microsoft Word, as a task-pane add-in. Writers don't switch tools or learn a new authoring environment. 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, swap a source, 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 Word add-in surface where every step of regulatory drafting happens — source ingestion, plan review, drafting, refinement, citation review, persona invocation, 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 bibliography, hyperlinks, and an end-of-run QC report ready. When the writer reopens the document, the studio 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 →

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 →

Personas are reviewer perspectives that can be summoned over any section of the draft. Today: FDA Clinical Reviewer, CMC Reviewer, CER Device Reviewer, CSR Writer QC. Each persona returns structured findings — a rubric anchor (e.g. ICH E3 §11.4.1), the snippet that triggered the finding, severity, suggested action. Every invocation is logged. They are not a replacement for human review; they are tools writers use to pre-empt reviewer feedback. SME teams can author additional personas as markdown skills, without engineering involvement. Read about the skill framework →

Every cross-reference (“see Table 14.3.1.2”, “as discussed in §11.4.1”, “per Figure 1”) is bound to a stable structural anchor at the moment the section is written. When the add-in writes the draft, it creates real Word bookmarks and resolved hyperlinks in the .docx. Sections can be reordered without breaking links; numbers re-derive from current document state. It is not a publishing-phase fixup pass — the hyperlinks ship in the draft, ready for eCTD B21 validation. Read the hyperlinks blog →

Six structural questions: cross-reference integrity, statistical-claim source binding, defined-term ordering, citation density per section type, gap detection, hyperlink resolution. 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 — neither by Asthra nor by Anthropic. Our enterprise contract excludes prompts and content from training and fine-tuning. Each project's source set is isolated to that customer.

Anthropic Claude, via Anthropic's enterprise API or, in customer-VPC mode, via your own Bedrock or Vertex deployment. We track frontier model releases and validate Asthra against each one before promoting.

SOC 2 Type 1 and ISO 27001 ready. SOC 2 Type 2, GDPR, HIPAA, and EU AI Act in progress; EU data residency in flight. Full status on the Trust center.

Managed SaaS: single-tenant project storage in EU or US regions. 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 and project data live in your tenant for as long as you need them; Asthra doesn't retain copies outside your perimeter. Audit ledgers persist for the lifetime of the document and travel embedded in the .docx.

Yes, on HIPAA-bound engagements. Reach out 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, 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.

Annual subscription, tiered by document volume and team size. Pilots are fixed-fee. We share full pricing once we understand your team's mix of document types and current cycle times — happy to send a quote after a 30-minute scoping 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.