Skip to content

How Asthra works: an agent that plans, retrieves, drafts, and hands off

Provenance-backed regulatory drafting inside Microsoft Word — built on Anthropic's Claude.

Request a demoSee supported use cases
Asthra AI Word add-in — Plan, Retrieve, Draft, Hand off
Plan → Retrieve → Draft → Hand off, with chat-mode refinement — all inside Word.
Built on Claude

Agentic retrieval, extended reasoning, and closed-system drafting — purpose-fit for regulated documentation.

Asthra AI is a regulatory writing co-pilot integrated with Microsoft Word. It helps clinical, safety, and regulatory teams generate compliant first drafts from specified source documents, with writer-defined provenance, sentence-level citations, and full human control — backed by an agentic drafting engine built on Claude.

The four-step agent loop

Every run is Plan → Retrieve → Draft → Hand off

1. Plan

The agent produces a retrieval plan — which source files and sections it will use for each part of the document. Writers can approve or adjust the plan before any expensive drafting runs.

2. Retrieve

Semantic retrieval extracts relevant excerpts from your uploaded documents. Nothing from the open internet. Nothing from model training. Excerpts are cached per session for consistency across related sections.

3. Draft

Claude drafts each section against module-specific writing instructions your team validated during onboarding. Missing data is flagged inline — never fabricated. Sections are checked against an optional reviewer pass.

4. Hand off

The draft lands in Word with an immutable transaction ledger embedded in the .docx. Every claim traces back to a source file, page, and passage. Writers edit in place with track changes.

Chat mode

Refine the draft by talking to it

Once the first draft is written, Asthra stays in the document with you. Writers can ask the agent why a passage was cited, what sources were compared, or whether a claim has supporting evidence. They can direct rewrites — “tighten §4.1”, “swap this source for the protocol”, “add a cross-reference to SAP §3.1” — and the agent applies them with the same closed-system retrieval that produced the original draft.

Every chat turn is recorded. Every citation updates automatically. Manual edits are detected and logged too, so the provenance trail stays honest even when a writer types changes directly into Word.

W

Which lines in §10.2 came from the SAP?

A

Two passages. “The primary efficacy analysis was conducted on the ITT population…” — SAP v3.2, §3.1. And “Missing data were handled using…” — SAP v3.2, §3.4. Want me to show the exact page snippets?

Illustrative. Real chats happen inside the Word add-in.

What makes Asthra different

Four principles built into every feature

Writers stay in control

Writers decide which sources to reference, approve the retrieval plan, and direct refinements. Asthra never ships content the writer hasn't seen.

Evidence-backed outputs

Every sentence can be traced to a source file, page, and passage. Citations are generated, not retrofitted.

No configuration at draft time

Templates and prompts are validated during onboarding. At draft time, the writer picks sources and clicks generate.

Built for complex PDFs

The retrieval stack is purpose-built for large, structured regulatory PDFs — tables, figures, cross-references, and mixed content.

From kickoff to first draft

Onboarding happens once per template; drafting happens every day

1

Template onboarding

Your template is onboarded by the Asthra team. Prompts are designed, tested, and validated with your writers. Provenance rules are recorded per section (file, section, table, figure).

2

Template selection in Word

Writer opens the Asthra task pane, picks the right template, and starts a new document. No setup required at draft time.

3

Upload source documents

Upload source files from the local machine or select from an integrated Document Management System. Large PDFs with tables and figures are first-class citizens.

4

Approve the plan, then generate

Review the agent's retrieval plan, adjust sources if needed, then generate. Missing information is flagged inline, never fabricated.

5

Verify provenance and citations

Every section shows the source documents it drew from. Writers can request sentence-level citations on demand with file, page, and exact text snippet.

6

Chat, edit, and QC

Refine sections via chat or edit directly in Word. Track changes is preserved. The transaction ledger travels with the .docx for audit.

Inputs to outputs

What you provide, what Asthra generates

InputsOutput
Trial protocol, TFL, SAPCSR first draft (ICH E3)
Previous PSUR, CDS, exposure data, safety line-listingsPSUR / PBRER / DSUR first draft (ICH E2C)
Clinical data, literature, post-market surveillanceCER first draft (EU MDR)
Batch records, CoAs, specifications, stability dataCMC Module 3 first draft

Supported workflows

Start generating compliant drafts for your most critical documents

CSR

Generate a full CSR first draft from trial protocol, TFL, and SAP.

Learn more →

PSUR / PBRER / DSUR

Aggregate safety reports from historical PSURs, CDS, exposure and safety data.

Learn more →

CER

Clinical Evaluation Reports for medical devices, aligned to EU MDR expectations.

Learn more →

CMC (Module 3)

Consolidate BMRs, CoAs, specifications, and stability data into structured Module 3.

Learn more →

Trust and compliance built in

Enterprise-grade security and validation readiness

Closed-system data boundary

Asthra uses only the sources you provide. No open internet. No training-knowledge leakage. Your data stays yours.

No fabrication

Missing data is flagged, not invented. Granular context engineering keeps the model on your evidence.

Two-level traceability

Document-level provenance plus on-demand sentence-level citations with file, page, and exact snippet. Ledger embedded in the .docx.

Validation posture

SOC 2 Type 1 ready. SOC 2 Type 2, ISO 27001, GDPR, and HIPAA in progress. Deploy as managed SaaS or in your own VPC.

Frequently asked questions

Asthra is built on Anthropic's Claude. Inside a single drafting run, Asthra selects between Claude models (Opus, Sonnet, Haiku) based on the task — discovery, drafting, or review — to balance quality and cost. Anthropic does not use data submitted through the API to train its models.
Asthra uses only the documents you provide. For CSRs: trial protocols, Statistical Analysis Plans, and Tables/Figures/Listings. For PSURs and related safety reports: historical PSURs, Core Data Sheets, exposure and sales data, safety line-listings. For CERs: clinical data, literature, post-market surveillance, device documentation. For CMC: Batch Manufacturing Records, Certificates of Analysis, specifications, analytical methods, stability data.
Two ways. First, a closed-system architecture: Asthra retrieves only from your uploaded documents — never from the open internet and never from the model's training memory. Second, granular context engineering: the agent is given only the relevant excerpts per section, so it has nothing to hallucinate from. When a source gap exists, Asthra flags it inline instead of fabricating content.
After the first draft lands in Word, chat mode lets a writer refine sections in natural language: "tighten §4.1 to 80 words", "why did you cite this passage?", "compare with the previous PSUR". Every refinement uses the same closed-system retrieval, updates citations automatically, and is recorded in the transaction ledger.
Yes. Before the agent calls the model to draft, it produces a plan — which source files and sections it intends to use for each part of the document. Writers can approve the plan, swap sources, or add constraints before any expensive generation runs.
Yes. Writers edit directly in Microsoft Word, with full track-changes support. They can also modify prompts or change provenance and regenerate a section. Manual edits are detected and logged as direct edits, so provenance stays honest.
Asthra provides two levels of traceability. Document-level: for each section, the source files and sections that were used. Sentence-level: on demand, file name, page number, and exact text snippet behind any claim. An immutable transaction ledger is written into the .docx itself — the audit trail travels with the document.
Asthra is SOC 2 Type 1 ready. SOC 2 Type 2, ISO 27001, GDPR, and HIPAA are in progress. Customer documents are never used to train or fine-tune models. Deploy as managed SaaS or in your own VPC.
CSRs (ICH E3), PSURs / PBRERs / DSURs (ICH E2C), CERs (EU MDR), and CMC modules (eCTD Module 3). Additional regulatory document types can be onboarded typically in a matter of weeks — the underlying doc-type definitions are YAML-driven and extensible.

See Asthra on your documents

Request a personalised demo. We'll run Asthra against a document type you care about — CSR, PSUR, CER, CMC — and walk through the provenance trail end to end.

Request a demoExplore use cases

Last updated: 16 April 2026