Awareness Is Abundant. Readiness Is Scarce.
Awareness is now the cheapest thing in clinical development. Every board understands that AI will reshape how trials are designed, run, and reviewed. Readiness is the expensive thing, and it is scarce.
Most organizations can describe what AI might do for a clinical trial. Far fewer can show an operating model that reliably carries a decision from one function to the next. The distance between those two states is not a maturity milestone you eventually reach. It is a competitive variable, and it is widening right now. That gap is what I set out to describe on stage at BIO Asia–Taiwan 2026.
What the Room Clarified
Session A-2 of the Innovation Forum was built around a shared ambition: accelerate clinical trial progress and increase the probability of success with AI. The room reflected the whole system, with voices from clinical development, research, technology, hospital networks, and regulators across Taiwan, Japan, Korea, and Singapore.
I opened with two questions. Question one: “What does readiness actually look like?” Question two: “Where is your readiness gap?” The presentation offered a four-layer answer. The discussion clarified something deeper. Readiness becomes visible the moment a protocol changes, a risk signal emerges, data moves across systems, or a decision crosses a functional boundary. This article is my attempt to answer both questions more completely than a stage allows.
Why Awareness Feels Like Readiness
Awareness is legible. It produces artifacts you can see and celebrate. A strategy deck. A named head of AI. A vendor on contract. A pilot with a demo. All of that is visible motion, and visible motion feels like progress. Readiness is the opposite. It lives in workflows, data lineage, shared definitions, escalation pathways, and decision rights. It is buried in the operating model, where no one takes a photo of it.
So three reflexes quietly substitute for readiness. The strategy reflex mistakes a plan for a capability. The pilot reflex mistakes a curated demo for enterprise execution. The procurement reflex mistakes buying a tool for building the foundation it runs on. Each one generates the sensation of movement while the operating model underneath stays exactly where it was. The model may work. The surrounding system may not.
The Readiness Stack Is a Dependency Model
On stage I described AI readiness as a four-layer stack. That framing is correct, but it is easy to misread as a catalogue of four things to mature independently. It is not. It is a dependency model. Each layer carries the one above it, and failure in one layer changes what is possible in the next.
Layer 1: Quality-by-Design Operating Model
Layer 1 defines what matters before the study begins. Critical-to-quality factors, a risk framework, quality tolerance limits, escalation pathways, and decision rights should shape the protocol and the operating plan, not be inspected in afterward. This is now regulatory expectation. ICH E6(R3) Principles and Annex 1 came into effect in the EU on 23 July 2025, the FDA published final guidance on 9 September 2025, and Annex 2 reached Step 4 on 3 June 2026 with an effective date of 15 January 2027. If Layer 1 is weak, everything above it optimizes toward the wrong targets, faster.
The common failure is quality-later. A team completes the design, locks the protocol, then adds Quality-by-Design language at the end. The terminology is present, but it never changes monitoring intensity, signal thresholds, or authority to act. When Layer 1 is weak, structured data only makes the wrong priorities easier to process, and AI detects patterns the organization never defined as meaningful.
Layer 2: Structured Protocol and Standards
Layer 2 converts scientific intent into reusable structure. In an AI-ready model the protocol is not only a document of record. It is the operating asset that feeds quality planning, data capture, oversight, site execution, and submission content. The standards exist now: ICH M11 was adopted at Step 4 on 19 November 2025, with the EMA Step 5 guideline effective 11 June 2026, and CDISC released USDM v4.0 on 3 June 2025, aligned to M11.
The common failure is template-first modernization. Legacy narrative is pasted into an M11-shaped document while downstream functions still interpret and re-key the same criteria, endpoints, visits, and amendments. When the protocol is a structured object, one change can reconfigure downstream systems. When it is a Word file, five functions re-key it five ways, and meaning begins to drift. Each function may perform correctly, yet the study no longer operates from one coherent model.
Layer 3: Governed, Interoperable Data
Layer 3 preserves meaning as information moves. Every material value should have a source, a lineage, an owner, a definition, and an approved use. Data integrity has to be designed at the source, signals pre-specified before you look for them, and human oversight calibrated to the risk of the decision. AI operates on structured data, not on PDFs, but structure alone is not enough.
The common failure is mistaking centralized storage for governed data. A data lake can collect information without establishing common meaning, provenance, or accountability. Faster access to accumulating data raises the governance bar rather than lowering it, because faster data can reach the wrong conclusion faster. When Layer 3 is weak, an AI output may be technically impressive and still be difficult to explain, validate, or place into a clinical decision pathway.
Layer 4: AI-Enabled Execution
Layer 4 is where copilots, matching tools, RBQM triage, signal-support models, and submission assembly become visible. It earns its place only when the three layers underneath support repeatable use. Its only inputs are the outputs of those layers. That is the whole argument. Layer 4 cannot manufacture a foundation it was handed incomplete.
The common failure is pilot exceptionalism. A use case works with curated data, a small expert team, and manual support behind the scenes, then becomes evidence that the organization is ready to scale. The more demanding question is whether the next study can use the capability across validated systems, defined governance, changing data, multiple functions, and inspection-ready workflows without the pilot team holding it together.
For any proposed AI use case, readiness is constrained by the weakest layer on which that use case depends.
The Real Readiness Gap Appears at the Handoffs
Most clinical development organizations do not lack capable functions. They have experienced medical writers, statisticians, data managers, clinical operations teams, quality professionals, regulatory strategists, and technology specialists. The readiness gap appears when the work has to move between them.
Consider a protocol amendment. In a document-driven model, the amendment changes the protocol, but its operational meaning must then be rediscovered by each downstream function. Data management determines what must change in the EDC. Clinical operations assesses the effect on sites. Biostatistics evaluates the analysis implications. Quality revisits the risk controls. Training and submission content are updated through separate processes. The organization may complete every task correctly and still not be AI-ready, because the change did not propagate through a shared model. It was repeatedly interpreted.
The same pattern appears with clinical signals. A model may surface an unusual pattern, but the value of that output depends on questions that sit outside the algorithm. Was the signal defined prospectively? Is the underlying data complete? Who reviews it? What threshold requires escalation? Which decisions may be AI-supported, and which stay under direct human command?
This is why readiness is more than the sum of four mature layers. The protocol must carry scientific intent into execution. Quality priorities must shape data and monitoring. Data must preserve lineage into review. AI outputs must enter a governed decision pathway. The real test of an AI-ready operating model is not whether each function has modern tools. It is whether meaning survives the handoffs.
Readiness does not live only inside the layers. It lives in the connections between them.
The Hidden Cost of Readiness Debt
There is a reason weak operating models can look healthy for years. Experienced people quietly repair the gaps. A data manager reconciles two definitions that should have been one. A medical writer re-enters an endpoint definition into a downstream document because the protocol did not carry it in reusable form. A statistician reconstructs the logic behind a decision no system recorded. Each repair is invisible, and together they create the illusion of a functioning model.
I call the accumulated cost of those repairs readiness debt. It builds up through repeated re-keying, inconsistent definitions, local workarounds, parallel versions of the truth, governance bolted on after deployment, decisions whose logic cannot be traced, and pilots that only keep working because an expert is holding them together by hand.
The accumulated cost of the manual repairs that keep a weak operating model functioning. It remains hidden while skilled people absorb the burden and is exposed when automation must operate without those workarounds.
Organizations can service readiness debt for a long time, because people absorb it. AI is what calls the loan. Automation removes exactly the human interpretation and reconciliation that kept the old model standing. This is the sharper form of a point I made on stage: the algorithm amplifies the operating model rather than creating it. Legacy workflows often appear more reliable than they are because skilled people continuously patch them. AI does not eliminate those gaps. It exposes them, and then it scales them. AI readiness work is therefore more than tool implementation. It is the work of paying down operating-model debt.
Skilled people can make a weak operating model look reliable. Automation must carry what they have been absorbing.
Readiness Has to Be Owned
If readiness lives in the connections between functions, then ownership is not a detail. It is the mechanism. The common answer, that readiness should be owned cross-functionally, is correct and insufficient, because “everyone owns it” usually means no one has the authority to resolve a conflict across functions.
Three kinds of ownership have to coexist. Functional owners maintain the disciplines: quality owns Quality-by-Design, data management owns standards, biostatistics owns estimands, regulatory owns compliance. A use-case owner is one accountable executive for a specific clinical or operational outcome, not a committee. A system owner is accountable for whether protocol, quality, data, systems, and decision rights actually operate as one model rather than five.
The practical position is this. AI readiness should be governed cross-functionally, but each use case still needs one accountable business owner. Technology can enable readiness. A committee can govern it. Neither can own the clinical outcome. When something breaks at a handoff, there must be a named person whose job is to see the whole model, not just their function of it.
What BIO Asia–Taiwan Added
The panel widened the frame in a way worth stating plainly: AI-ready clinical development is an ecosystem property as much as an organizational one. A sponsor can build all four layers internally and still be limited by the environment it operates in. Inconsistent site infrastructure, disconnected vendors, incompatible data definitions, and unclear cross-border data governance are handoff failures too, only now they occur between organizations rather than between functions.
This is what gives integrated clinical networks their real strategic value, and it is where Taiwan and the wider Asia-Pacific region have a genuine opportunity. The advantage is not more sites or more participants or more data. Many regions have volume. Networks matter when they establish common operating structure across institutions: shared definitions, connected expertise, coordinated activation, and consistent data governance. The future advantage will belong to the ecosystem that can turn trusted data into reliable evidence, not the one that simply holds the most of it.
Three Tests for the Next Trial
Enterprise readiness can feel abstract. The next trial is a more useful unit of action. You do not need a transformation to close a readiness gap. You need your next trial to survive three tests, each of which targets a handoff rather than a tool.
1. The change-propagation test
Change one protocol element, an endpoint, a visit, an eligibility criterion, and ask whether the organization can identify and execute the consequences across data capture, randomization, ePRO, monitoring, site training, analysis, and submission content without rediscovering the meaning in each function. If the change has to be rediscovered five times, Layer 2 has not carried Layer 1 forward.
2. The signal-traceability test
Take one material signal and try to trace it end to end: from source data, to its prospective definition, to its threshold, to its named reviewer, to the escalation decision, to the action, to the documented rationale. If the chain breaks anywhere, you have data access, not data readiness.
3. The production test
Ask whether the next trial can use the capability under normal conditions, with governed data, validated systems, defined human oversight, and measurable outcomes, without the original pilot team holding it together. If it cannot run without its inventors in the room, you built a demonstration, not a capability.
These tests do not require an overnight transformation. They require one study to operate more coherently than the last. Govern before you automate. Make the next trial more AI-ready than the last.
Companion Resource
The 4-Layer AI-Readiness Scorecard
A free, five-page self-assessment that turns the stack into thirteen questions across the four layers, with the anatomy of each.
- ✔Thirteen questions, mapped to each layer's anatomy
- ✔Aligned with ICH E6(R3), ICH M11, and CDISC USDM
- ✔Print-ready · No form, no gate
Self-Assessment Scorecards
Closing
AI readiness is not demonstrated by the sophistication of the model. It is demonstrated by whether scientific intent, data, risk, and decisions can move through the organization without losing meaning. AI is not something you buy. It is a capability you build, and it is built at the seams between functions, not inside any one of them.
The answers to “What does readiness actually look like?” and “Where is your readiness gap?” will differ by organization, program, trial, and use case. The discipline is the same: find the weakest dependency, strengthen the handoffs, assign ownership, and retire your readiness debt before automation exposes it. Awareness is a starting point. Readiness is an operating capability. Let’s build it.
Explore further on kushdhody.com
- The full BIO Asia–Taiwan 2026 presentation deck — also listed in the Conference Presentations section.
- The 4-Layer AI-Readiness Scorecard and related diagnostics for ICH M11, ICH E6(R3), and FDA real-time clinical trials.
- AI-Enabled Clinical Research Operations: From Concept to Implementation (June 2025).
Selected sources
- EMA — ICH E6(R3) Good Clinical Practice scientific guideline (Principles + Annex 1 effective 23 July 2025; Annex 2 Step 4 3 June 2026, effective 15 January 2027)
- FDA — E6(R3) Good Clinical Practice final guidance (9 September 2025)
- ICH — M11 CeSHarP Step 4 final guideline (19 November 2025)
- EMA — ICH M11 Step 5 guideline (effective 11 June 2026)
- CDISC — USDM v4.0 / Digital Data Flow (released 3 June 2025)
About the author
Kush Dhody, M.D., M.S. is a physician-scientist and clinical development executive with more than 20 years of experience leading global clinical programs, protocol design, regulatory strategy, and clinical operations across multiple therapeutic areas. He currently serves as President of Amarex Clinical Research, LLC, An NSF Company, and is involved in AI-enabled regulatory and quality workflow innovation, including the NSF/Microsoft Azure initiative featured as a Microsoft customer story.
DISCLAIMER: The views expressed in this blog are those of the author and do not necessarily represent the official position of Amarex Clinical Research, LLC (An NSF Company), NSF, any sponsor or partner, or any regulatory authority. This post reflects the author's interpretation of publicly available information and emerging developments in AI-enabled clinical development, protocol standardization, and regulatory modernization. Adoption of any approach discussed here should be evaluated in the context of the specific product, study design, therapeutic area, regulatory jurisdiction, organizational capabilities, and applicable health authority expectations. It is intended for informational and educational purposes only and should not be construed as regulatory, legal, compliance, or medical advice.