AssuranceLens | Every requirement accounted for before you submit

Why now

Regulators are starting to read submissions with AI. Read yours the same way first.

The first pass over your file is increasingly a machine's. Amendments remain the most expensive way to find out what it would have flagged.

76%

of protocols are amended at least once

Tufts Center for the Study of Drug Development, protocol amendment research.

$141k to $535k

direct cost of one substantial amendment

Phase II to Phase III median, before the schedule slip. Tufts CSDD, Getz et al.

Clause by clause

is how assessor tooling now works

FDA has deployed agency-wide generative AI for reviewers, and the MHRA is building AI-assisted assessment tools. Files are parsed against criteria, one at a time.

The platform

Three engines, one content-provenance graph

Compliance, consistency and provenance are three views of the same documents. AssuranceLens keeps them in one place so a change in one document is visible everywhere it matters.

Live

Requirement Coverage Engine

Every applicable requirement gets a verdict. Nothing is skipped silently.

866 verbatim-indexed clauses from ICH E6(R3), E8(R1), E9, EU CTR 536/2014 and 21 CFR 50, 54, 56 and 312.
Verdicts of met, partial, missing or not applicable per clause, with scope exclusions listed rather than hidden.
Findings placed in a closed taxonomy of eight categories, from missing content to administrative gaps.
Related requirements across guidance documents are clustered so a reviewer handles each obligation once.
Gap status tracking, severity, and CSV, PDF or API export after human review is confirmed.

Live

Content Alignment Map

See where a protocol, consent form and analysis plan disagree before a regulator does.

Cross-document map of identical, equivalent and divergent statements, computed deterministically.
Numeric divergence flagged explicitly: a dose, window or sample size changed in one document and not the others.
Blast radius for an amendment: which documents share the paragraph you are about to change.
Version comparison for v1 versus v2 drift, contradictions, and content added or removed.
No model calls and no cost to recompute. Re-upload the corrected document and the map converges.

Pilot

Evidence-to-Summary Engine

Summaries built from an evidence ledger, so a failed test cannot disappear on the way to the summary.

Extraction ledger first: every result row is captured with its source location before any prose is written.
Each summary claim cites ledger rows. Uncited claims and ungrounded numbers are rejected by a deterministic check.
Completeness accounting: an unaccounted failed row is a blocking error, not a footnote.
Adversarial challenge pass attaches doubts for the reviewer; the deterministic layer remains the gate.
Vertical-agnostic: clinical study report sections today, device submission summaries as an output adapter.

How it works

From upload to a defensible ledger in four steps

The engine is deliberately not an agent. It sweeps a fixed clause index so that coverage is enumerable and every omission is visible.

Read the full technical explainer

1
Declare the scope
Create a project, choose the clinical or device track, and declare the study profile or device profile. Applicability by phase, region and design is decided from what you declare, not guessed.
2
Upload the document set
PDF, Word or text. Each file is tagged with its document type. Pages without a readable text layer are detected and reported instead of being silently read as empty.
3
The engine reads clause by clause
Requirements are judged in batches against your cached documents. Each cited passage is checked verbatim against the source, then a second pass tries to refute every finding.
4
Review, adjudicate, export
Work the coverage ledger by category and severity, resolve disputed findings, confirm human review, and export the report for your submission file.

What makes it different

Designed so that an omission is impossible to miss

Most document AI optimises for a fluent answer. AssuranceLens optimises for an auditable one.

Fail-visible coverage ledger

The report accounts for every indexed requirement: evaluated, excluded by scope, or routed to manual review. A gap the engine could not judge is shown as such, never dropped.

Verbatim grounding

Every cited requirement is matched against the source text. Citations that cannot be located are downgraded and flagged, so hallucinated references do not reach your reviewers.

Adversarial refute pass

After the first verdicts, a second pass searches your documents for counter-evidence. Findings it can challenge are marked disputed with the evidence attached for adjudication.

Scan-aware extraction

Signed consent forms and scanned appendices have no text layer. The engine detects unreadable pages and routes affected verdicts to manual review instead of reporting false gaps.

Closed finding taxonomy

Eight fixed categories, chosen from a constrained vocabulary rather than free text, so findings group consistently across runs, documents and models.

Built for re-runs

Documents are cached, requirements are batched, and verdicts for unchanged documents are reused. Re-analysing after an amendment does not mean paying for the whole study again.

Your model, your data

Cloud model by default, with a provider seam for self-hosted open-weight models where data-residency rules require it. Available to pilot partners on request.

Human oversight by design

AI transparency disclosure on every result, a human-review gate before export, model version recorded per analysis, and an audit log of who did what.

Who it is for

Built for the teams who sign the submission

Two tracks, one engine: clinical trial documentation and medical device technical documentation, with a dedicated expectations pack for AI-enabled software.

Clinical trial sponsors and CROs

Protocols, amendments, informed consent forms, investigator brochures, statistical analysis plans and study reports checked against ICH GCP, the EU Clinical Trials Regulation and FDA IND rules.

Protocol · ICF · IB · SAP · CSR · Synopsis

Radiology AI and SaMD manufacturers

An AI and software-as-a-medical-device expectations pack drawn from published MHRA AI Airlock reports: intended-purpose scope, validation evidence, change control, post-market monitoring, explainability.

Intended purpose · Validation · PCCP · PMS · Explainability

Medical device manufacturers

Technical documentation reviewed against a library that spans EU MDR and IVDR, FDA 21 CFR Parts 803 to 820, UK MDR, and the Japanese, Chinese and Australian frameworks.

Technical file · 510(k) · Design inputs and outputs · Risk file

Regulatory consultants

Run the first pass in minutes, spend your hours on judgement. Organisation workspaces, role-based access, and an audit trail you can show a client.

Multi-client workspaces · Viewer seats · Export bundles

Regulatory coverage

Source texts, stored with provenance

The clinical index is verbatim: each clause is the regulation's own wording, extracted once and reviewed. The device library holds the full texts used as reference during analysis.

Clinical clause index

866 live

ICH E6(R3) Good Clinical Practice162
ICH E8(R1) General Considerations for Clinical Studies83
ICH E9 Statistical Principles for Clinical Trials217
EU Clinical Trials Regulation 536/2014151
21 CFR Part 312 Investigational New Drug Application188
21 CFR Part 50 Protection of Human Subjects30
21 CFR Part 54 Financial Disclosure by Clinical Investigators20
21 CFR Part 56 Institutional Review Boards15

ICH texts are reproduced under the ICH public licence with copyright acknowledged. EU texts are reused under Commission Decision 2011/833/EU. US federal regulations are public domain.

Device and SaMD library

45 texts

European Union

MDR 2017/745 and IVDR 2017/746
Regulations 2022/2346 and 2022/2347
EU AI Act 2024/1689, medical-device extract
MDCG 2019-11 software qualification and classification
Directives 90/385/EEC, 93/42/EEC, 98/79/EC

United States

21 CFR Parts 803, 806, 807, 812, 814 and 820
FDA AI/ML-based SaMD Action Plan
FDA Predetermined Change Control Plan guidance and guiding principles
Good Machine Learning Practice guiding principles

United Kingdom

UK Medical Devices Regulations, as amended by SI 2023/627
MHRA Software and AI as a Medical Device change programme roadmap
MHRA AI Airlock pilot and Phase 2 reports

Japan, China, Australia

Japan PMD Act (Japanese and English) and QMS comparison
China Order 739 and SAMR decrees on registration, manufacturing, distribution, UDI and adverse events
Australia Therapeutic Goods Act 1989

The AI and SaMD expectations pack is derived from published MHRA AI Airlock reports under the Open Government Licence. It describes sandbox expectations, not statute, and implies no regulator endorsement.

Trust and governance

Security, AI governance, and honest limitations

Security

Documents encrypted at rest with AES-256-GCM and in transit with TLS.
Role-based access: platform staff are excluded from customer document content by design.
Multi-factor authentication, session controls, and a full audit log of security events.
Controls mapped to ISO 27001 and SOC 2 criteria; security headers and CSP enforced at the edge.

AI governance

Transparency disclosure on every AI-generated result, including the model version used.
Human-review gate: export is locked until a qualified person confirms the review.
Expert feedback loop: reviewers can mark findings sensible or not, and add what the engine missed.
Practices aligned with EU AI Act Articles 13 and 14 and the GMLP guiding principles.

Limitations, stated plainly

Verdicts are probabilistic. False positives and false negatives occur, and the results are not exhaustive.
Pages without a text layer cannot be judged; they are routed to manual review.
The alignment map catches near-verbatim overlap; semantic restatements need a human eye.
Decision support only. Never the sole basis for a submission decision.

Questions

Frequently asked

Is this a replacement for a regulatory affairs professional?+

No. AssuranceLens is decision support. It performs the exhaustive first pass, accounts for every requirement, and shows its evidence. A qualified person reviews the output, adjudicates disputed findings and confirms the review before anything can be exported.

How do you stop the model from inventing requirements?+

Requirements are not generated at run time. They come from a curated clause index built from the source texts, and every citation the engine produces is checked verbatim against that index. Citations that cannot be located are downgraded and flagged for review.

What happens to our documents?+

Documents are encrypted at rest with AES-256-GCM and transmitted over TLS. Platform staff cannot view customer document content; that permission is restricted to your organisation. Model providers process documents under API terms that exclude training on your data, and a self-hosted model option exists for pilots with stricter residency requirements.

Which regulations and guidance are covered?+

The clinical clause index covers ICH E6(R3), E8(R1) and E9, EU CTR 536/2014 and 21 CFR Parts 50, 54, 56 and 312. The device and SaMD library spans EU, US, UK, Japanese, Chinese and Australian texts. New texts are added with provenance and version tracking.

Does it work on scanned PDFs?+

Scanned pages have no text layer, and the engine detects this per page. Verdicts that depend on unreadable pages are routed to manual review rather than reported as missing content. Optical character recognition and image-level review are on the roadmap.

Is the MHRA AI Airlock pack an MHRA endorsement?+

No. The pack is derived from the published Airlock reports and reflects expectations the reports describe. It is not statute, and AssuranceLens is not affiliated with or endorsed by the MHRA or any other regulator.

Every requirement accounted for.Before you submit.