Will AI replace CRAs? What the evidence actually says

Type the question into a search engine and look at who answers it. The scariest result, an article titled "Meet Your New Boss: How AI Will Replace Clinical Research Jobs by 2028," is published by CCRPS, a company that sells certification courses. The boldest automation number, agents absorbing "up to 90% of the tactical and administrative work" of clinical monitoring, comes from Medable, the company that sells the agent. The reassuring "AI will never replace the human touch" pieces come from CRA staffing firms whose business depends on placing CRAs.

Nobody on page one is disinterested. So set the takes aside and look at what can actually be verified: what the tools do, what happened to CRA jobs the last time technology came for monitoring, and where the 2025-26 job losses actually came from.

The short answer

Tasks are compressing. The job is not disappearing. The 2025-26 pain in clinical research employment traces to the biopharma funding downturn, not to automation. And the industry has already run a decade-long experiment on exactly this question, with a clear result. The rest of this article is the evidence for those four sentences.

What the tools can actually do

Draw the line by task, because every serious source, vendors included, draws it in the same place.

On the automatable side: source data verification assisted by OCR, auto-generated queries, site correspondence drafts, CTMS and eTMF updates, visit-report first drafts, protocol summarization, and central statistical monitoring that flags outliers across sites. Medable's CRA Agent, launched September 2025, targets precisely this list, drafting queries, sending emails, and updating systems with a human approving each step. Its headline numbers, eight hours saved per CRA per week and $40 million in productivity for a large team, appear in an article written by a Medable employee and have no independent audit. Treat them the way you would treat any sales figure.

On the other side sits the work that makes monitoring a profession rather than a checklist. The clearest inventory comes from a CRA-community analysis of AI's limits: a model cannot interpret a protocol deviation in context, cannot decide whether and how to intervene when it flags an anomaly, and cannot build the trust with a site that gets a struggling coordinator to tell you what is actually going wrong. Even the alarmist CCRPS piece concedes the point in its own body copy: when a regulator asks why an imputation was chosen, a human owns the rationale, and only humans can pull the investigator, sponsor, statisticians, and safety together to agree on a fix.

The disagreement between the sellers and the skeptics is not about where the line sits. It is about how much of your week sits on the automatable side of it. That is worth knowing your own answer to, honestly, because it varies enormously by role and sponsor.

We already ran this experiment

Here is the part the 2026 think pieces skip. A decade ago, a regulator-endorsed technology arrived that was supposed to shrink the monitoring workforce: risk-based monitoring. FDA issued guidance in 2013, EMA the same year, and ICH E6(R2) wrote it into GCP in 2016.

The evidence behind it was devastating to the traditional CRA task. The TransCelerate analysis by Sheetz and colleagues, published in 2014 across 1,168 studies, found that 100% source data verification corrected about 1.1% of all site-entered data, and drove only about 2.4% of queries on critical data. The activity that defined on-site monitoring, checking every value against source, caught almost nothing worth catching. There has never been a stronger case that a CRA's core manual task was expendable.

What happened next: not much, slowly. A 2021 review found only 22% of trials had adopted even one of five risk-based monitoring components, with individual components ranging from 8% to 19%. And through that entire decade, the industry kept reporting CRA shortages and attrition rates of 22 to 30%. The task shrank in importance; demand for the people did not.

That is the base rate to reason from. A proven, regulator-backed reason to need fewer monitors took ten years to partially adopt and never collapsed the job market. AI is a genuinely more capable technology, but it lands in the same industry, with the same validation burdens, the same site relationships, and the same regulatory accountability that slowed the last transition. Betting on wholesale replacement by 2028 means betting that this industry suddenly moves ten times faster than it ever has.

The layoffs are real. The attribution is wrong.

If you lost a role or watched your bench shrink in 2025, nothing above makes that less real. But look at where the cuts actually happened: Novo Nordisk about 9,000 roles, Bayer about 12,000, Moderna roughly a tenth of its workforce, in restructurings that trade coverage attributes to patent cliffs, competition, and the venture funding drought, with biotech M&A in Massachusetts down 74% in one quarter. The Fierce Biotech layoff tracker tells the same pipeline-and-funding story. Meanwhile the large CROs were reporting strong earnings through the same period, and Parexel was hiring thousands in India. Across the whole US economy in 2025, one aggregation counted about 55,000 layoffs directly attributed to AI out of roughly 1.17 million total, under 5%.

A note on two numbers you may have seen, including on this site: clinical research job openings down 32% year over year and more than 26,000 professionals displaced. Both come from the CCRPS 2025 workforce report, which publishes no methodology and cites no primary sources, and whose own text elsewhere says employers plan to maintain or grow CRA headcount through 2026. A course seller's unsourced downturn statistics, in a report that also sells the cure, deserve the same skepticism as a vendor's ROI slide.

The mechanism to actually watch is the one CCRPS names in passing and then buries under its headline: workload compression. Not pink slips attributed to a model, but one CRA carrying more sites because the admin got faster, and backfills quietly not opening. That shows up in requisition counts and site loads before it ever shows up in a layoff notice, which makes it worth tracking at your own employer.

Where the role is actually going

The regulator has already told you. ICH E6(R3), finalized in January 2025 and adopted in FDA final guidance in September 2025, formally recognizes centralized monitoring as a core oversight approach and no longer expects traditional SDV as a default. That is a regulator-driven shift toward central and remote monitoring, quality tolerance limits, and critical-to-quality thinking, and it would be happening with or without large language models.

When working CRAs and clinical ops people think through the replacement question, they tend to land in the same place: the transcription and box-checking layers go, the GCP-compliance judgment stays, and central monitoring reads as a growth area because someone has to design and interpret the risk indicators the algorithms watch. Job postings already reflect it, with monitoring roles drifting toward study-integrity and quality profiles: less verification, more investigation.

The practical moves follow directly. Learn risk-based quality management and central-monitoring concepts, because E6(R3) makes them the grammar of the job. Get fluent with the AI tools your organization actually deploys, inside its data policies, so the efficiency accrues to you. And treat any percentage anyone quotes about your job's automatability the way this article treats the 90% and the 32%: ask who is selling what before you believe it.

For the wider picture across every clinical research function, see AI in clinical research: what's real. For what the credibility rules mean when a model's output touches a submission, see FDA's AI guidance, explained. And if you want to know what CRAs are actually paid while all this shifts, the salary data is the reason this site exists.