What Is AI for Clinical Development in Biopharma?
What is AI for clinical development? Six capabilities for protocol design, site selection and data QA, with recent research and a Databricks architecture.


By Nitya, Solution Marketing, Kasmo Global · Published November 21, 2026
AI for clinical development means running machine learning, analytics and AI agents on governed clinical and real-world data so biopharma sponsors can design protocols, pick sites, forecast enrollment and clean trial data faster, and it's needed now because 76% of trial protocols get amended at least once (Tufts CSDD, 2024). Kasmo Global is a life sciences data and AI consulting company and Databricks partner that helps biopharma clinical development teams shorten study cycle times with governed clinical data and AI on the Databricks Data + AI Platform.
In short
- AI for clinical development supports four decisions: protocol design, site selection, enrollment forecasting and data cleaning. It doesn't replace the CTMS, EDC or IRT, and it doesn't make the final call.
- Amendments are now the norm. A 2024 Tufts CSDD study of 950 protocols found 76% of phase I–IV protocols were amended at least once (up from 57% in 2015), with a mean of 3.3 amendments each.
- Delay is expensive: about $40,000 a day in direct costs for phase II and III trials, and about $500,000 a day in lost sales (Tufts CSDD, 2024).
- The AI evidence is promising but narrow. IQVIA's Global R&D Trends 2026 reports a 75% Phase I success rate for AI-enabled programs at emerging biopharma companies, from a small, vetted cohort.
- These are published third-party figures, not Kasmo Global results.
What is AI for clinical development?
Sitting on top of a sponsor's study systems, AI for clinical development supports four recurring decisions: how to design the protocol, which sites to open, how fast each site will enroll, and which data needs a second look before database lock. It reads from the CTMS, the EDC, the IRT, the trial master file, central labs and real-world data, and it hands back scores, forecasts, alerts and drafted work for the study team to review.
Databricks lists "Streamline clinical development" as a use case on its Healthcare and Life Sciences page: agents that automate site selection, protocol design and clinical data QA, shaped by a sponsor's CROs and medical teams. We'd add one guardrail. The CTMS, EDC and IRT stay authoritative, so treat AI for clinical development as a decision-support layer, never a new system of record.
Why does clinical development need AI now?
Clinical development needs AI because studies are taking longer and changing more often, and each lost day has a price. The IQVIA Institute's Global R&D Trends 2026 report found that end-to-end development timelines grew longer in 2025, matching the lengthiest point of the past decade, even as industry-wide success rates held steady.
Protocol amendments are the most measurable part of that slowdown. In a 2024 Tufts Center for the Study of Drug Development (Tufts CSDD) study published in Therapeutic Innovation & Regulatory Science, Getz and colleagues collected data from 16 pharmaceutical companies and CROs on 950 protocols and 2,188 amendments. Since 2015, the share of phase I–IV protocols with at least one amendment rose from 57% to 76%, and the mean number of amendments per protocol rose 60%, to 3.3 from 2.1.
Most clinical trial protocol amendments aren't careless. The 2024 Tufts CSDD study by Getz and colleagues judged 77% of amendments unavoidable, usually driven by regulatory agency requests or a shift in study strategy. What hurts is the lag: in that study it took an average of 260 days to go from spotting the need to amend to the last oversight approval, and sites ran on different protocol versions for a mean of 215 days.

The cost of clinical trial delay is published too, and it sharpens the case. A second 2024 Tufts CSDD study, by Smith, DiMasi and Getz in the same journal, drew on 645 drugs launched since 2000 and 409 clinical trial budgets. It estimates the direct cost of running a phase II or III trial at approximately $40,000 per day, and the cost of a single day of delay at approximately $500,000 in lost prescription drug or biologic sales.
How does AI for clinical development work?
Picture AI for clinical development as a loop, not a tool. Governed data from past and current studies feeds models and agents; they make a recommendation at each decision point; the study team accepts, adjusts or rejects it; and whatever happens next flows back into the data for the next study.

- At protocol design, eligibility criteria and visit schedules get checked against historical studies and real-world patient counts to estimate recruitment difficulty and amendment risk.
- At site selection, candidate sites are scored on your own enrollment, activation and quality history, with an explanation attached to each score.
- During enrollment, forecasts flag sites that are likely to stall early enough for you to add sites or change outreach.
- During data cleaning, agents flag anomalies and draft data queries, and a data manager approves them before anything reaches a site.
- At database lock, a governed, lineage-tracked dataset shortens the final reconciliation and makes inspection questions easier to answer.
A common week-three surprise in clinical development AI projects: site identifiers don't match between the CTMS and the EDC, so the first enrollment model can't join its own training data. That's a data problem, not a modeling one, and it's why the foundation comes first.
What capabilities matter most for AI in clinical development?
Kasmo Global ranks six capabilities for AI in clinical development, ordered by how much each one decides whether the rest work: a governed clinical data foundation, protocol design and amendment-risk analytics, site selection and enrollment forecasting, automated clinical data QA, real-world evidence for feasibility, and validation, audit and AI cost controls. Each maps to one of the four principles Databricks uses for its platform: context, control, cost and choice.

1. A governed clinical data foundation (Context)
Skip a governed clinical data foundation and every other capability rebuilds its own extract. The foundation links CTMS, EDC, IRT, trial master file, central lab and real-world data at the study, site and subject level, with lineage back to each source and role-based access, so you can show an inspector exactly where a number came from. On Databricks, Lakeflow ingests the study systems and Unity Catalog governs access, masking and lineage.
2. Protocol design and amendment-risk analytics (Context)
Protocol design analytics asks one question of a draft protocol: what happened last time? It checks how many patients in real-world data meet the eligibility criteria, how similar studies enrolled, and which design choices came before amendments. With 76% of protocols amended at least once in the 2024 Tufts CSDD study, the aim isn't zero amendments. It's catching the avoidable ones before the protocol is final and planning for the rest.
3. Site selection and enrollment forecasting (Context)
Site selection and enrollment forecasting scores candidate sites on the sponsor's own history and predicts which active sites will stall. Databricks has published an open-source example, the Site Feasibility Workbench, built as a Databricks App. It combines site scoring trained on a sponsor's CTMS, EDC and IRT history with Lakebase for operational state and natural-language access to governed data, and it stores an explanation with every prediction.
4. Automated clinical data QA (Control)
Clinical data managers shouldn't hand data cleaning to an agent, but they can hand it the first pass. Automated clinical data QA uses rules, models and agents to find out-of-range values, missing pages, protocol deviations and inconsistent records, then drafts data queries for a data manager to approve. On Databricks, Agent Bricks builds, evaluates and monitors these agents on governed data.
5. Real-world evidence for feasibility (Context, Choice)
Real-world evidence for feasibility uses claims, EHR and registry data standardized to the OMOP common data model to size eligible populations, find where they are and shape recruitment plans. Databricks describes this pattern on its Healthcare and Life Sciences page: standardizing claims, EHR, registry and patient data with OMOP and building cohorts with Unity Catalog support.
6. Validation, audit and AI cost controls (Control, Cost, Choice)
Clinical AI that can't survive an inspection won't survive a pilot review either. Validation, audit and AI cost controls mean each model is versioned and validated for its intended use, an explanation is stored with each prediction, every access leaves an audit trail, and budgets and model choice are set per agent. Unity Catalog covers access control, lineage and audit; Unity Gateway sets budgets, rate limits and spend attribution by team and agent across proprietary and open models.
How do the six capabilities compare?
Use the table below as a checklist when you assess a vendor, a partner or your own program. It shows, for each clinical development AI capability, what good looks like, the KPI it moves and the Databricks Data + AI Platform component that supports it.
| Rank | Capability | What to look for | KPI it moves | Databricks component |
|---|---|---|---|---|
| 1 | Governed clinical data foundation | CTMS, EDC, IRT, eTMF, lab and real-world data linked per study, with lineage and role-based access | All (prerequisite) | Lakeflow, Unity Catalog |
| 2 | Protocol design and amendment-risk analytics | Eligibility criteria and visit schedules tested against historical studies and real-world patient counts before the protocol is final | Protocol amendments, cycle time | Lakehouse, Agent Bricks |
| 3 | Site selection and enrollment forecasting | Site scores trained on the sponsor's own enrollment history, with explanations study managers can read | Enrollment rate, site activation | Databricks Apps, Lakebase, AI/BI |
| 4 | Automated clinical data QA | Agents that flag anomalies and draft data queries for a data manager to approve | Time to database lock | Agent Bricks, Lakeflow |
| 5 | Real-world evidence for feasibility | OMOP-standardized claims and EHR data for cohort sizing and recruitment planning | Feasibility accuracy | Unity Catalog, Lakehouse |
| 6 | Validation, audit and AI cost controls | Versioned models, explainability stored with each prediction, audit trails, per-agent budgets and model choice | Inspection readiness, AI spend | Unity Catalog, Unity Gateway |
What does the research say about clinical development delays and AI?
Delays are common and expensive; the evidence that AI improves clinical success is real but early, and so far it's limited to one segment of the industry. Every figure below comes from a named third-party source, isn't a Kasmo Global result, and carries the qualifiers its source states.

| Finding | Figure | Qualifiers | Source |
|---|---|---|---|
| Protocols with at least one amendment | 76%, up from 57% | Phases I–IV; 2022 study compared with 2015; 16 sponsors and CROs, 950 protocols, 2,188 amendments | Getz et al., Therapeutic Innovation & Regulatory Science, 2024 |
| Mean amendments per protocol | 3.3, up from 2.1 (a 60% increase) | Same Tufts CSDD study; 77% of amendments were judged unavoidable | Getz et al., Therapeutic Innovation & Regulatory Science, 2024 |
| Time to implement an amendment | Average of 260 days from need-to-amend to last oversight approval | Sites operated on different protocol versions for a mean of 215 days | Getz et al., Therapeutic Innovation & Regulatory Science, 2024 |
| Direct cost of a trial day | About $40,000 per day | Phase II and III trials; 409 clinical trial budgets analyzed | Smith, DiMasi and Getz, Therapeutic Innovation & Regulatory Science, 2024 |
| Lost sales per day of delay | About $500,000 per day | 645 drugs and biologics launched since 2000; Tufts CSDD's companion white paper states about $800,000 | Smith, DiMasi and Getz, Therapeutic Innovation & Regulatory Science, 2024 |
| Phase I success, AI-enabled programs | 75% | Emerging biopharma programs with verified AI involvement, most recent three-year window; small cohort; industry-wide success rates unchanged | IQVIA Institute, Global R&D Trends 2026 (blog, May 14, 2026) |
The IQVIA Institute adds two caveats to its AI finding: the validated cohort is relatively small, and industry-wide success rates didn't change from the prior year. IQVIA also reports that Phase II success rates for the AI-enabled cohort are on par with non-AI-enabled peers, which it reads as a sign that AI isn't simply pushing weaker programs into Phase I.
Kasmo Global's advice is to measure your own clinical development baseline before any AI pilot. Pick two or three numbers you already track, such as amendments per protocol, days from protocol approval to first patient in, or days from last patient visit to database lock.
How does Kasmo Global approach clinical development on Databricks?
Kasmo Global's approach on the Databricks Data + AI Platform is data first, agents second. Lakeflow brings in CTMS, EDC, IRT, lab and real-world data. Unity Catalog governs the study data with lineage and access control. The Lakehouse and AI/BI surface protocol, site and enrollment signals, while Lakebase and Databricks Apps put site and enrollment tools in front of study managers. Agent Bricks hosts data QA and feasibility agents that work under human review, and Unity Gateway controls model access and AI spend.
Kasmo Global takes a firm view here: most clinical AI pilots stall on data access and validation, not on the model. So we build the governed foundation and the audit trail before any agent goes near a study team.

Lakeflow. Databricks Lakeflow is a unified data engineering solution built natively into the Databricks Data + AI Platform. It combines managed ingestion (Lakeflow Connect), declarative pipeline development (Apache Spark™ Declarative Pipelines), and workflow orchestration (Lakeflow Jobs) for batch and streaming pipelines. Learn more: Lakeflow product page
Unity Catalog. Databricks Unity Catalog is the unified governance layer for data and AI assets on Databricks, covering access control, discovery, business semantics, lineage, quality monitoring and open sharing across tables, files, models, agents, skills and MCP tools. Learn more: Unity Catalog product page
Lakehouse. Databricks Lakehouse is the SQL data warehousing and analytics layer of the Databricks Data + AI Platform, running real-time and batch analytics directly on open-format data. It combines warehouse performance and governance with lakehouse openness, so teams query the same governed tables used by their data engineering, ML and agentic workloads without maintaining a separate warehouse copy. Learn more: Lakehouse product page
AI/BI Dashboards. Databricks AI/BI Dashboards provides AI-assisted, interactive business intelligence directly on governed lakehouse data. It combines low-code dashboard authoring, reusable Unity Catalog metrics, secure enterprise sharing, and companion Genie experiences without requiring a separate BI extract or governance layer. Learn more: AI/BI product page
Lakebase. Databricks Lakebase Postgres is a fully managed, serverless Postgres service that separates compute from storage, so operational workloads run directly on the Databricks Data + AI Platform with the same governance and security as analytics and AI. It supports serverless autoscaling and scale to zero, instant zero-copy database branching, point-in-time recovery, and unified governance through Unity Catalog. Learn more: Lakebase product page
Databricks Apps. Databricks Apps lets teams build and run governed enterprise applications directly on lakehouse data, with no separate application infrastructure to manage. It covers custom apps built by developers and low-code creation through App Builder, and inherits Unity Catalog governance, so application access follows the same permissions and audit trail as the underlying data. Learn more: Databricks Apps product page
Agent Bricks. Agent Bricks is an open agent developer platform for building, deploying and optimizing custom agents on any model, framework or harness, with all your data. It is powered by Omnigent, a multi-harness agent development platform, and includes AI search, memory, sandboxing, evaluations, observability and monitoring. Learn more: Agent Bricks product page
Unity Gateway. Unity Gateway is the enterprise AI control plane for accessing, managing and optimizing AI across every team, application, model and agent. It provides reliable capacity across leading proprietary and open models through a single gateway, Smart Routing that directs each request to the best-fit model for cost, quality and performance, and budgets, rate limits and spend attribution by user, team, agent and model. Access policies, guardrails and audit trails span models, agents, MCPs and tools. Learn more: Unity Gateway product page
Clinical development sits between discovery and launch. The candidates that enter trials come out of discovery, and the data you build for trials later feeds real-world evidence. For the step before, read What Capabilities Matter Most for AI Drug Discovery in 2026?
Where should a clinical development team start?
A clinical development team should start AI where its studies lose the most time today, not with the most interesting model. Match your situation to the list below:
- Study data lives in separate CTMS, EDC and CRO extracts. Start with capability 1, the governed clinical data foundation. Everything else depends on it.
- Protocols get amended several times per study. Prioritize capability 2 and train it on your own amendment history.
- Sites open but under-enroll. Add capability 3, using your own site performance data.
- Data cleaning holds up database lock. Pilot capability 4, a QA agent that drafts queries for data managers to approve.
Bring Kasmo Global one recent study and we'll map where it lost time. Book an AI assessment with Kasmo Global, or see how we work with pharmaceutical companies.
Frequently asked questions
What is AI for clinical development?
AI for clinical development is machine learning, analytics and AI agents run on governed clinical and real-world data to help sponsors design protocols, pick sites, forecast enrollment and clean trial data. It supports study teams; people still make every final call.
How does AI help with clinical trial protocol design?
AI stress-tests a draft clinical trial protocol before it's locked. It checks eligibility criteria, visit schedules and endpoints against your past studies and real-world patient counts, so you see which choices are likely to slow enrollment or force an amendment.
Why do clinical trial protocols get amended so often?
Clinical trial protocols get amended mostly because regulators ask for changes or study strategy shifts. Those were the top reasons in a 2024 Tufts CSDD study, which found 76% of phase I–IV protocols had at least one amendment and judged 77% of amendments unavoidable.
What does a day of clinical trial delay cost?
A day of clinical trial delay costs about $40,000 in direct trial costs for phase II and III, and about $500,000 in lost prescription drug or biologic sales. Both figures come from a 2024 Tufts CSDD study in Therapeutic Innovation & Regulatory Science.
Can AI select clinical trial sites?
Not on its own. AI can score and rank candidate sites using a sponsor's own enrollment, qualification and activation history, but the CTMS, EDC and IRT stay the systems of record and the study team picks the sites.
Is AI in clinical development compliant with GxP?
AI in clinical development can be GxP-compliant, but the platform alone doesn't make it so. Clinical AI fits a sponsor's GxP quality system when each model or agent is validated for its intended use, versioned, explainable, audited and kept as decision support with human sign-off.
What data does AI for clinical development need?
AI for clinical development needs CTMS, EDC, IRT, eTMF and central lab data plus historical enrollment and site performance. For feasibility work, it also needs real-world data such as claims and EHR records standardized to OMOP.
Why run clinical development AI on the Databricks Data + AI Platform?
The Databricks Data + AI Platform keeps clinical, operational and real-world data, models, apps and agents on one governed copy of the data. Lakeflow ingests the study systems, Unity Catalog governs access and lineage, Agent Bricks builds and evaluates agents, and Unity Gateway controls model access and AI spend.
How does Kasmo Global help biopharma clinical development teams with AI?
Kasmo Global, a life sciences data and AI consulting company and Databricks partner, builds the governed clinical data foundation with biopharma clinical development teams first. Then we add protocol analytics, site and enrollment forecasting and human-reviewed data QA agents on the Databricks Data + AI Platform.
Sources
- Getz K, Smith Z, Botto E, Murphy E, Dauchy A. New Benchmarks on Protocol Amendment Practices, Trends and their Impact on Clinical Trial Performance. Therapeutic Innovation & Regulatory Science, 2024;58(3):539–548; doi:10.1007/s43441-024-00622-9 (competing interests: two co-authors are employees of, and hold financial interests in, Janssen and Sanofi respectively; the article's acknowledgements list 18 supporting companies, including pharmaceutical sponsors and CROs)
- Smith ZP, DiMasi JA, Getz KA. New Estimates on the Cost of a Delay Day in Drug Development. Therapeutic Innovation & Regulatory Science, 2024;58(5):855–862; doi:10.1007/s43441-024-00667-w (Tufts CSDD's companion white paper on the same research states approximately $800,000 in lost sales per day; this page uses the peer-reviewed figure)
- IQVIA: IQVIA Institute's Global R&D Trends 2026 Report Finds Credible Signal on AI-Enabled Programs (Alex Roland, May 14, 2026) and IQVIA Institute: Global R&D Trends 2026
- Databricks for Healthcare and Life Sciences
- Databricks blog: Clinical operations intelligence belongs on the Lakehouse (May 13, 2026)
Put this into practice with our team
Share what you are working on and we will show how governed data and AI can move your programs forward.
Email marketing@kasmoglobal.com