ICH M11 structured protocol: how AI changes the trial design layer.
The clinical-trial protocol has been a Word document for forty years. ICH M11 reframes it as a structured, machine-readable artefact. The shift looks editorial — it is not. It is the layer at which AI-assisted trial design and regulator review begin to align, and it changes what a protocol is.
ICH M11 — the Clinical Electronic Structured Harmonised Protocol (CeSHarP) — reached Step 4 endorsement in early 2025 and is being implemented across ICH regions through 2026–2028. The deliverable is two-sided: a Technical Specification that defines the data model and exchange formats, and a Template plus Guideline that defines the structured protocol content. The companion eXtensible Markup Language (XML) and JSON schemas are published. The protocol becomes an exchange artefact between sponsor, regulator, ethics committee, site, and electronic Trial Master File — not a free-text PDF that each downstream system re-parses.
/ 01What changes at the protocol layer.
A traditional protocol contains structured information in unstructured form. Inclusion criteria are bullet points in a Section 4. Endpoints are paragraph text in Section 8. Statistical methods are an appendix that may or may not match the SAP. The information is there; the structure is implicit. M11 makes the structure explicit. Inclusion criteria become coded entries with attribute fields. Endpoints carry typed metadata — primary, secondary, exploratory; binary, continuous, time-to-event; analysis population. Statistical methods reference defined sections of the SAP rather than restating them.
The change in how the document is authored is small. The change in what can be done with the document downstream is large.
/ 02What AI can now do.
Assisted protocol drafting.
Large language models trained or augmented on the M11 schema can scaffold a protocol from a target product profile, a synopsis, and a list of decisions. The structured fields constrain the output in a way that free-text drafting did not — the model is filling in slots, not generating prose, which makes the output reviewable rather than rewritten. The 2025–2026 sponsor experience is that the assisted draft cuts authoring time in half on routine sections (population, schedule of assessments, safety reporting) while saving none on novel sections (rationale, risk-benefit, statistical innovation). The structure is the constraint that makes the assistance useful.
Pre-submission validation.
A structured protocol can be machine-validated against schema completeness, internal consistency, and cross-reference integrity before submission. Population referenced in endpoint definition matches population in inclusion criteria. Visit windows in schedule of assessments match visit references in safety reporting. Statistical population in primary endpoint matches population in sample-size calculation. The sponsor catches the inconsistency before the regulator does. The structured-protocol validators that vendors are building in 2026 will replace a substantial fraction of the internal medical-writing review cycle.
Regulator-side triage.
FDA, EMA, PMDA, and Health Canada are in various stages of operationalising M11 ingestion. The regulator side gets structured fields to prioritise review, flag inconsistencies, and compare across submissions in a way that PDF-based review made impractical. EMA's CTIS (Clinical Trial Information System) under Regulation (EU) 536/2014 is the channel for EU CTR submissions, and the M11 data model is what flows through it. The reviewer's job is not eliminated; the reviewer's preparation is accelerated.
The structured protocol is not faster paperwork. It is the same paperwork made addressable. Everything that becomes addressable becomes automatable in some part. The discipline is in what part.
/ 03What AI cannot do.
The protocol's substantive design decisions remain human. The choice of primary endpoint, the rationale for the comparator, the population justification, the statistical approach, the risk-benefit framing — all are decisions where the model can scaffold but cannot decide. The 2026 reality across major sponsors is a clear pattern: AI handles the structural and consistency layer; humans handle the design layer. The dividing line is not arbitrary. The structural layer is where errors are easy to detect and easy to correct. The design layer is where errors compound through the entire trial.
/ 04The eTMF consequence.
The DIA TMF Reference Model 3.3 (the de facto eTMF standard) does not yet pre-suppose M11-structured protocols. The intersection is being worked through. A structured protocol's elements can map directly to specific TMF zones — schedule of assessments to Zone 5, monitoring plan to Zone 4, statistical analysis plan reference to Zone 8 — in a way that the free-text protocol could not. The first wave of integrations is appearing in 2026 from Veeva, IQVIA, and the larger CRO eTMF stacks. The structured protocol becomes the spine that the TMF zones hang off, rather than a document that is itself one TMF artefact among many.
/ 05The implementation asymmetry.
FDA's adoption is being phased — the agency accepts M11-structured protocols today on an optional basis, with mandatory adoption signalled for 2027–2028 depending on protocol type. EMA is moving on a similar trajectory through CTIS. PMDA is implementing in lockstep with ICH timing. The asymmetry is in sponsor preparedness. Large pharma sponsors are building M11-native authoring into their internal systems through 2026. Mid-cap and small biotech are largely waiting. CROs that serve both populations are bridging — providing M11 conversion at submission time, which preserves the legacy authoring flow but loses the upstream benefits of structured drafting.
Sponsors that author in M11 from day one capture the validation benefit and the regulator-triage benefit. Sponsors that convert at submission time capture only the regulator-triage benefit, and even that is partial because conversion introduces re-parsing errors. The cheap path is to wait. The cheap path costs in the next inspection cycle, when consistency between protocol, SAP, CRF, and CSR becomes harder to demonstrate without the structured spine connecting them.
M11 is the first ICH guideline that treats a regulated document as data. It will not be the last. The lesson worth carrying forward is that structure changes what is possible, and what is possible eventually becomes what is required.