The email came with an attachment: a contract worth $1.93 million, spread across three years. For Harbor, a San Diego startup barely a year old, the deal represented something more consequential than revenue. A publicly traded medical device company was entrusting its entire clinical trial—end to end—to a team that had never run one before. Not as a consultant. Not as a software vendor. As the prime contractor, replacing the kind of legacy CRO that typically handles such work.
Perhaps the medical device company had run the numbers. Perhaps it had grown tired of waiting. Because the math, when you looked at it, was difficult to ignore.
A genomics study involving 1,600 participants—the kind Harbor documented in a case writeup earlier this year—went from contract signature to active patient enrollment in nine days using Harbor's AI-native approach. Nine. Competing bids from established electronic data capture vendors promised the same milestone in 10 to 12 weeks, minimum. The difference wasn't clever project management or weekend sprints. It was a fundamentally different infrastructure, one where AI generates case report forms from clinical protocols in minutes instead of months, and where source document extraction happens automatically rather than through the manual data entry that devours most of a site coordinator's day.
If this sounds like vendor hype, consider the industry Harbor is disrupting. The contract research organization market—companies that sponsor pharmaceutical and device trials hire to manage the operational chaos—is substantial and growing, with estimates varying by source and methodology. Grand View Research estimates $59.62 billion in 2025, climbing to $105.73 billion by 2033. Precedence Research counters with $69.56 billion for 2025 and $133.75 billion a decade out. The spread reflects more than methodology; it highlights how slippery the boundaries are between pure CRO work and the adjacent sprawl of software, site networks, and consulting that surrounds modern trials.
What no one disputes is the dysfunction.
A study published in JAMA Network Open examined trials initiated between 2010 and 2014 and found that only around 20 percent finished on schedule. The median delay? 12.2 months. More recent industry surveys suggest the problem has worsened rather than improved. A market report from PPD in 2025 noted that 45 percent of respondents saw trial timelines lengthen compared to the prior two years. Meanwhile, 55 percent of trials shut down early due to recruitment failures—an expensive way to learn nothing.
The inefficiency has a face, or at least a line item. Medidata, one of the dominant EDC platforms, published an analysis last fall estimating that source data verification—the painstaking process of manually cross-checking every data point in the trial database against original medical records—eats up 46 to 50 percent of on-site monitoring time and represents 25 to 40 percent of total trial costs. A 2013 study by TransCelerate, a pharma consortium—still widely referenced despite being over a decade old—found that only 2.4 percent of data queries in critical fields were actually triggered by SDV, which suggests the industry has been pouring resources into theater while real risks slip through.
So when a startup claims it can collapse a three-month database build into nine days, the question isn't whether the industry needs disruption. It's whether this particular disruption will stick.
When Regulation Becomes Tailwind
Three dynamics are aligning in ways that favor platforms built from scratch rather than retrofitted from legacy architecture.
The first is regulatory evolution, which isn't always an oxymoron. ICH E6(R3), the updated Good Clinical Practice guideline finalized late in 2024 and posted by the FDA last September, shifts emphasis from exhaustive documentation toward risk-based quality management. Translation: sponsors no longer need to verify 100 percent of source data if they can demonstrate that their monitoring strategy targets actual risk. The FDA's guidance on decentralized trials, also published last September, provides a formal framework for remote visits, local healthcare providers, and digital health tools—legitimizing practices that were pandemic-era improvisations.
Europe has moved in parallel, though with its characteristic regulatory flair. The European Medicines Agency published a reflection paper on AI in drug development last September, establishing expectations around transparency, human oversight, and risk management. The EU AI Act, approved by the Council in May 2024, labels AI systems used in clinical contexts as "high-risk," triggering compliance obligations that smaller vendors may find onerous. Together, these frameworks nudge the industry toward centralized monitoring, validated electronic systems, and explainable AI—requirements that favor platforms designed with those principles baked in rather than bolted on.
The second force is the slow-motion collapse of source data verification as orthodoxy. ACRO, an industry group representing CROs, surveyed its members last June and documented steady adoption of risk-based quality management since 2019, though uptake varies wildly by sponsor size. A peer-reviewed survey in Therapeutic Innovation & Regulatory Science found adoption hovering around 50 to 60 percent for core RBQM components, with smaller sponsors lagging. The regulatory blessing from E6(R3) has effectively given the industry permission to abandon a paradigm it never particularly liked but felt compelled to follow.
The third dynamic—and perhaps the most consequential—is that AI-native infrastructure can finally deliver automation that doesn't require sponsors to rewrite protocols or force site coordinators to learn new workflows. AI extraction from scanned source documents, for instance, enables remote monitoring without upending how sites operate. Protocol-to-CRF generation eliminates weeks of negotiation between sponsors, CROs, and EDC vendors over dropdown menus and field validation logic. Confidence scoring for data quality replaces manual query generation with risk-stratified review.
These aren't vaporware features. Medidata launched Health Record Connect and Rave Companion in April last year, claiming up to 90 percent faster form completion through EHR-suggested data. IgniteData's Archer platform, tested with Memorial Sloan Kettering and AstraZeneca, progressed from pilot studies to Phase 3 trials by early this year. And IQVIA's collaboration with NVIDIA, unveiled in January, uses orchestrated AI agents to handle trial start-up tasks—transcription, medical coding, data extraction, document summarization—coordinated by a higher-level "orchestrator" that routes work and assembles outputs.
The tools exist. What's changing is the regulatory and operational permission to use them at scale.
Harbor's Hypothesis: Own the System of Record, Provide the Service
Harbor's founding thesis was narrow at first. The company launched in 2025 with an AI-native electronic data capture system: Magic Build for protocol-to-CRF generation, Magic Capture for AI extraction from source documents, Magic Monitor for centralized risk-based monitoring. Pricing started at $2,000 per month for commercial studies, with a free tier for academic trials. The target was a specific pain point—the three-month slog between protocol finalization and database readiness.
But CEO Albert Cai and CTO Nathan Leung spotted a larger opportunity, one they articulated in their Y Combinator launch post with notable bluntness. Most of the money in clinical trials, they noted, flows to services rather than software. If you control the system of record and can automate the workflows around it—study build, data capture, query generation, monitoring, database lock—you can deliver the same regulated services that legacy CROs provide, only with fewer handoffs and compressed timelines. By early this year, Harbor had pivoted into full-service CRO work. The $1.93 million contract validated the bet.
The MotilityCount study provides texture. NOVA Genomics and MotilityCount ApS, a Danish firm, needed to launch a 1,600-participant genomics study. Traditional EDC vendors quoted 10 to 12 weeks for database build. Harbor delivered the study—protocol to live enrollment—in nine days, charging a flat $2,000 build fee. The trial, registered as NCT07369362, was enrolling patients as of March.
Nine days. It bears repeating because it represents an order-of-magnitude improvement, not an incremental one.
The Incumbents Aren't Sleeping
Legacy CROs are responding, though their strategies diverge in revealing ways.
Parexel, one of the larger players, has opted for partnerships rather than internal development. Last September, the company announced a collaboration with Paradigm Health, a site-network platform that raised $78 million in December and counts Flatiron Health among its strategic partners. Parexel CEO Peyton Howell described the model as "AI-native trial operations that compress timelines and lower operational costs while making trials accessible to more patients"—a formulation that nods toward both efficiency and patient-centricity without committing to specifics. A second partnership, with Weave Bio in December, focuses on accelerating regulatory submissions. The subtext is clear: Parexel believes the future involves orchestrating best-of-breed platforms rather than building everything in-house.
IQVIA has taken the opposite approach, betting on agentic AI and proprietary development. The company's NVIDIA collaboration, demonstrated at last year's GPU Technology Conference in Paris, showcased orchestrated agents handling trial start-up workflows under the coordination of a higher-level orchestrator. In September, IQVIA launched an AI-enabled Clinical Trial Financial Suite, automating budget negotiations and contract management—workflows that traditionally consume weeks of back-and-forth between sponsors and CROs. The message: IQVIA intends to automate the full stack, not just discrete pieces.
Smaller platforms are staking out niches. Grove, an AI agent platform for site operations, has signed CROs including Celerion, IMA Clinical Research, and K2 Medical Research. Client testimonials cite faster prescreening and higher screening pass rates, though published data remains scarce. Pi Health positions itself as a unified operating system combining EDC, clinical trial management systems, trial master files, and investigator site files, all with AI-driven builds and EHR-to-EDC automation. The company has announced collaborations with GSK and Filamon, though deployment scale is unclear.
Medidata, the market-leading EDC platform owned by Dassault Systèmes, reported in February that it had supported over 500 clinical studies with AI over the past decade, with more than 120 AI-supported studies launching in 2025 alone. The company's strategy emphasizes augmentation—EHR-assisted data entry, centralized monitoring dashboards, risk scoring—layered into an ecosystem that already serves most large pharma trials. It's a bet on stickiness: if you're already running trials on Medidata, incremental AI features are easier to adopt than migrating to a startup with no track record.
The Next Eighteen Months
Whether AI-native CROs become a distinct category or an interim configuration will clarify soon. Harbor's trajectory—EDC to full-service CRO in under a year—argues for the former. So does Paradigm Health's $78 million raise, a signal that investors believe site-network platforms with AI-enabled execution layers can compete on complex studies. But the incumbents retain structural advantages that startups dismiss at their peril: regulatory track records, global site relationships, therapeutic-area depth, and installed bases that generate formidable switching costs.
Regulatory evolution will shape the battlefield. The FDA's draft guidance on Bayesian methodology, released in January, could reduce timelines for adaptive and pediatric trials if finalized as proposed—benefiting platforms with integrated data and analytics. The EU AI Act's high-risk classification for clinical AI systems will raise compliance costs, potentially favoring larger players with established quality management infrastructures. And the EMA's emphasis on explainability in its September 2024 reflection paper sets a higher bar for "black box" automation than some early AI vendors anticipated.
Data unification is emerging as the decisive capability, the one that determines who captures value and who gets squeezed. Platforms that control the system of record—whether that's an EDC, a unified clinical data repository, or something else—and automate the workflows feeding into it can credibly collapse study start-up from months to weeks. Harbor's nine-day case study, Medidata's claims of 90 percent faster EHR-assisted completion, and IgniteData's progression from pilot to Phase 3 at Memorial Sloan Kettering all point in the same direction: the winners will be the companies that eliminate handoffs, not the ones that optimize them.
The efficiency crisis in clinical trials has persisted for decades partly because incentives never aligned with solving it. CROs billed by the hour. EDC vendors charged per subject. Monitoring teams justified budgets with exhaustive source data verification that added cost without demonstrably improving quality. AI-native platforms invert those incentives. They profit from compressing timelines, not extending them. And they do it with infrastructure regulators increasingly prefer—centralized, risk-based, auditable, transparent.
Whether this shift produces a handful of dominant platforms or a fragmented ecosystem of specialized tools remains uncertain. But the direction is clear enough. The era of 12-week database builds and 12-month trial delays isn't finished yet—legacy infrastructure has inertia, and inertia is a powerful force. But it may be closer to over than the incumbents would like to admit. And sometimes, in industries as ossified as clinical research, nine days is all the proof of concept the market needs.
