From Fax Machines to AI: When Tools Get Faster but the Work Doesn't Change

01 · The starting point

The fax machine never let me forget it was working.

I remember asking a team member to train me on how to use the fax machine at my first job. Thirteen years ago, I was a Data Tech at Duke University Brain Tumor Center. As the name suggests, our team was all about patient data: filing paper consent forms, requesting records from other hospitals and clinics, extracting data from paper files and writing them onto paper forms. The main fax machine in our Data Tech room sat on the left-hand side of my desk. Loud, constantly spitting out page after page, usually day after day, sometimes overnight.

One of our projects was figuring out how to reorganize all of the patient binders, hefty things filled with dozens of pages that had grown so large we were running out of space to store them. The hallway was lined with Regulatory binders. The cabinets were filled with hundreds of sheets of organized files and templates. Patients were consented via paper informed consent forms, some up to 30+ pages. I remember going up to clinic with a clipboard and a stack of paperclipped forms to get patient sign-offs during their appointments or consults.

In my second job as a Project Specialist at Syneos Health, paper Trial Master Files were still part of the work. I vividly remember volunteering to spend extra time around the holidays completing several paper TMF reviews while a colleague was out of office. Paper was still everywhere, but it wasn't everywhere alone. Electronic systems weren't new. At Duke, I'd been working in Oracle Clinical, which had been a clinical-trial data standard since the mid-1990s. What was new from 2013 onwards was the pace. The shift toward cloud-based SaaS had been underway for years. Medidata had been a public cloud-clinical software company since 2009.² Veeva Technologies' 2013 IPO¹ was a milestone moment for clinical trial management software in particular, and it was during this period that I became increasingly acclimated to electronic Clinical Trial Management Systems (eCTMS), eTMF systems, and cloud-hosted databases. Technology was moving forward.

02 · The forecast gap

eConsent shows that industry adoption forecasts can be wrong by years.

Fast forward to today, and eConsent is supposedly becoming the norm. The forecasts have been around for a while. In 2019, Signant Health's State of eConsent survey found that 85% of sponsor respondents expected to adopt eConsent for some studies within 12 months, and 71% expected adoption to extend to the majority of studies within three years and beyond.³

Four years later, Suvoda's October 2023 eConsent Market Survey told a different story: only one-third of sponsors, and less than one-fifth of sites, were actually using eConsent.⁴ The gap between what the industry said it would do and what it actually did is striking.

eConsent adoption has its own arc worth understanding, and the inflection point was COVID. When the pandemic forced remote-first operations, trial utilization grew roughly 460% by 2021 versus pre-pandemic levels, according to GlobalData's Trials Intelligence Platform. Then it retreated. Adoption slowed by about 31% in 2022 for ePRO/eCOA/eConsent activity, with commercial-sponsor decentralized-trial use down 52% in the same window. By 2024, adoption had stabilized rather than rebounded.⁵

The promise of eConsent is real, and the technology works. In Suvoda's October 2023 survey, 40% of sites named lack of provision by sponsors as the primary barrier to using it.⁴ The bottleneck wasn't the technology or the sites. It was upstream of both: sponsor-level decisions that didn't land on deployment, for reasons the data doesn't fully capture. Some of those decisions were probably active choices about cost, risk, or integration complexity. Some were clarity gaps about what eConsent should look like, study by study. Forecasts that assume technology is the variable consistently miss the real one: the human decisions, conscious or otherwise, that determine whether the tool actually gets deployed.

03 · The numbers themselves are contested

eCOA is becoming a gold standard. What it actually includes is contested.

Patients complete assessments from their own mobile device, or one supplied by the sponsor or CRO supporting the trial. Adoption now spans the majority of sponsor trials: used in 53% of trials over the past two years, projected to reach 64% in the next two.⁶ MarketsandMarkets' May 2025 update pegged the market at $2.27 billion in 2025, growing to $4.78 billion by 2030, a 16.1% CAGR.⁷

Medable's own vendor reporting cited 80% revenue growth in 2024 across enterprise customers adopting portfolio-level eCOA versus study-by-study contracts, suggesting that deployment strategies may be maturing.⁸ But the spread tells its own story: Data Bridge sizes the 2024 market at $1.70B, MarketsandMarkets at $1.94B, Cognitive at $6.5B (Cognitive uses a broader scope that includes adjacent eClinical categories, so the 4x headline overstates the gap). Even among comparable-scope analysts, the disagreement is real, and it's about scope rather than math: what does "eCOA" actually include? Patient-reported outcomes captured on an app (ePRO)? Clinician-rated assessments (eClinRO)? Observer-reported items (eObsRO)? Performance outcomes (ePerfO)? Wearable and sensor data? The associated services and integrations, or just the software? Different analysts draw the line in different places. That ambiguity carries downstream into sponsor procurement decisions, site deployment configurations, and integration choices. That's a clarity deficit hiding inside the growth chart.

04 · Crisis as catalyst

COVID proved that crisis can drive real change. But only some of it sticks.

2020 changed everything for decentralized trials. Registered DCTs grew from 102 in 2019 to 189 in 2020, an 85% surge, coinciding with the COVID-19 pandemic.⁹ In the first year of the pandemic, 76% of pharmaceutical companies, device manufacturers, and CROs adopted decentralized techniques. Of those, only 7% implemented fully decentralized methods. The majority folded select remote components into otherwise conventional trial designs.¹⁰

It wasn't that the technology hadn't been advancing before 2020. COVID introduced a genuine operational requirement for sites and patients to adapt to digital tools faster than most would have otherwise. The FDA followed up with draft guidance in 2023, then finalized its DCT guidance in September 2024. The guidance, titled "Conducting Clinical Trials With Decentralized Elements: Guidance for Industry," provided the first formal regulatory framework for what the industry had spent four years improvising.¹¹

But the gap between "some DCT" and "fully decentralized" is where the story gets interesting. What stuck was the easier work: telehealth visits, eConsent for select trial phases, remote monitoring for stable patients. What largely didn't stick was the harder work. Fully remote trials, decentralized investigational product delivery, in-home clinical assessments. Each of those required decisions that hadn't been made. Which protocol elements actually need a site visit? What data-integrity standards apply to samples collected at home? Which therapeutic areas tolerate decentralization without compromising signal? The pandemic forced the adoption. The unanswered design questions slowed the retention.

05 · A new arrival

When enterprise AI fails, the failure is almost always a clarity problem, not a technology problem.

Each of those technologies (eConsent, eCOA, eTMF, DCT) was about digitizing an existing workflow. Faster, more remote, more measurable. But the interpretive work, what counts as informed consent, how an adverse event gets graded, what counts as a complete file, was still defined by humans. AI is the first technology in this arc that doesn't just speed up the work. It does the interpretation. That changes the clarity question from important to non-negotiable.

The next four sections look at where enterprise AI is struggling and what the failures share in common. The argument isn't that AI doesn't work. Plenty of evidence by 2026 shows it does, in the right places. The argument is that the failures are remarkably consistent, and they point upstream of the technology.

MIT's NANDA Initiative published The GenAI Divide: State of AI in Business 2025, a study drawn from executive interviews, leader surveys, and analysis of public AI deployments. The headline finding: 95% of pilots delivered no measurable P&L impact, despite an estimated $30-40 billion spent on AI systems by American enterprises in 2024.¹² The report's own language: "The 95% failure rate for enterprise AI solutions represents the clearest manifestation of the GenAI Divide."

The headline finding is dramatic, but the more interesting part is what's underneath. NANDA's taxonomy is part real technology limitations, part organizational design failures. The pattern across both: failures show up where teams deployed AI without defining what it needed to do.

Agentic AI architectures address some of these limitations directly: persistent memory, learning from feedback, deeper workflow integration. But the clarity question doesn't get easier with more capable tools. It gets sharper, because an autonomous system needs more explicit upfront definitions, not fewer.

06 · The human cost

Two independent 2026 studies converged on the same finding: AI intensifies work rather than reducing it.

The early data on what AI does to the people doing the work is worth taking seriously. The same tools that compress timelines can, over time, expand the cognitive load on the people operating them: more output to review, more edge cases to catch, more decisions to make faster.

These findings should make us more considerate of what we're deploying, and at what pace.

07 · The deal wave

The pharma AI deal wave keeps escalating, even as some of its earliest poster children quietly disappear.

While that human cost surfaces, pharma-AI dealmaking has only accelerated. Novartis closed up to $2.3B with Schrödinger in late 2024, one of the largest publicly disclosed AI-pharma deals at the time.¹⁷ The CRO side moved in parallel: in January 2025, IQVIA and NVIDIA announced an agentic AI collaboration at JPM, which shipped as IQVIA.ai in March 2026.¹⁵¹⁶

By JPM 2026, the pace had accelerated, with deals stacking through Q1. In a single week in January, Pfizer announced its Boltz collaboration, Eli Lilly disclosed a Chai Discovery partnership, Lilly + NVIDIA committed $1B over 5 years to a co-innovation lab, and J&J struck Isomorphic Labs' third pharma partnership.^15181920 By the end of Q1, Lilly had also expanded its Insilico collaboration into a deal worth up to $2.75B.

But not every AI-pharma story is going forward. BenevolentAI, one of the highest-profile pure-play AI drug-discovery companies of the previous decade and AstraZeneca's partner since 2019 on chronic kidney disease and pulmonary fibrosis,²¹ announced its proposed delisting via merger into Osaka Holdings in February 2025, with the extraordinary general meeting completing in March.²² One highly publicized AI-generated CKD target, followed by a quiet exit.

08 · The shape of the failure

Unlike the fax machine, AI failures are silent and invisible. In clinical trials, that matters more than almost anywhere else.

I have to ask myself: even with all of these tools at our fingertips, is the work really becoming easier? Are answers actually becoming clearer? In the days of working next to a fax machine, the technology was loud, slow, and visible. AI is the opposite. It is silent, fast, and invisible.

The MIT NANDA report includes one stat that captures this perfectly. 90% of workers report daily use of personal AI tools like ChatGPT and Claude. Only 40% of companies have an official LLM subscription.¹² Work product is being shaped, every day, by tools that don't appear in any system of record.

In clinical trials, that invisibility matters more than it does almost anywhere else. The failures I worry about aren't the obvious ones, like an AI that hallucinates a citation or produces an obviously wrong output. Those get caught. The ones that worry me are the quiet kind. An adverse event miscategorized at scale because the model confidently mapped it to the wrong MedDRA term. A TMF assembled to look complete while quietly missing a protocol deviation. eCOA data smoothed cleanly enough that a real patient signal gets flagged as noise. None of these announce themselves the way a fax machine running overnight does. They land in the dataset, in the file, in the analysis. They only surface, if they surface at all, when somebody asks the right question after the fact.

09 · The honest counterweight

Where AI is earning its keep, the clarity was already there.

To be clear: AI is genuinely earning its keep in places. Literature reviews can be compressed substantially. Data review cycles can be cut significantly. Protocol parsing, which pulls inclusion and exclusion criteria from dense documents, is the kind of high-volume, well-defined work AI handles beautifully. I've seen this in my own day-to-day. Recently I had AI merge 20 reports together, run integrated tools across the contents, and extract the data into the formats I needed. That kind of work, high-volume with clean inputs and a clear endpoint, is where AI is genuinely good. I've also used it to build small personal tools, scripts and workflows that automate tasks which used to take me hours. It can be useful, and honestly, fun, when the work is well-defined and I stay in the loop on the output. Wherever it succeeds, the pattern is the same: someone defined the data model and the workflow before AI began running over top of it. Clarity work was done first.

Here's the strongest case against my own argument. NANDA found that 5% of integrated AI pilots are extracting millions in value, and some startups have gone from zero to $20M in revenue inside a year.¹² Mid-market organizations move pilot-to-production in roughly 90 days versus 9+ months for large enterprises, meaning speed of iteration can partially substitute for the upfront clarity work I keep arguing for. AI can also be the forcing function that surfaces a clarity gap, not just consume one. Some teams do the alignment work because the AI deployment finally made the cost of not doing it visible.

Where the task is clear, the data is clean, and a human reviewer stays in the loop, AI is delivering real time back to teams. And that's the point that's getting missed in the debate: AI isn't failing because the technology is broken. It's failing where we've asked it to substitute for clarity we never had to begin with.

10 · The thing AI cannot do

The thing AI cannot do is define what good looks like. That's still on us.

As leaders, we need to reframe the question. Call it clarity-first: instead of asking "where can we use AI," we should be asking: "where do we have enough clarity already that AI can actually help us, and where do we need to do the human work first?" AI may be here to stay, and only time will tell how the technology continues to advance. Each new model seems stronger and more capable than the last. But we need to make sure we are not cutting corners. We are not outsourcing the most important element of what makes our work meaningful, in any industry: the human element.

Of all the questions I keep coming back to, the one I'm most willing to stake a position on is data definitions. AI cleans messy data faster than any human can. But it will also confidently produce clean-looking output from inputs that are conceptually broken. If your team hasn't aligned on how adverse events get categorized at the edges, or what "protocol deviation" means in your specific study, AI will not save you. It will accelerate your existing disagreement and make it harder to detect after the fact. The work of getting humans to align on what we mean is not optional, and it cannot be outsourced. Every hour spent fighting about definitions before the trial starts is worth ten hours of AI-aided cleanup you won't have to do later.

The FDA is making the same argument, in regulatory language. In January 2025, CDER issued its first AI-specific draft guidance, Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products. It proposes a seven-step credibility assessment framework for any AI model whose output supports a regulatory decision about safety, efficacy, or quality.²³ The order of the steps is the thing worth reading carefully. The technical work (building, validating, and documenting the model) doesn't begin until step four. The first three are something else entirely:

AI puts the onus more firmly back in our lap. The foundation we build on becomes even more critical the more we trust automation to take over.

11 · The clarity-first checklist

For clinical trials, the clarity-first questions are concrete.

• Protocol and Endpoints: Are we clear on what we're actually trying to measure before we let AI optimize the trial design?
• Data Definitions and Integrity: Do we know what good data looks like before we ask AI to clean it or pull insights from it?
• Patient Identification and Recruitment: Have we defined who the right patient is before we let AI find them at scale?
• Site Performance and Risk: Have we agreed on what "risk" means at a site before we let AI flag it?
• TMF and Regulatory: Do we have clear standards for what should be in the file before we use AI to assemble it?
• eCOA and Connected Devices: Do we know what data actually matters clinically before we collect everything we can?
• Project and Study Management: Are our processes clear, or are we about to automate broken ones?
• Pharmacovigilance and Safety Reporting: Do we have clear case definitions and coding standards before we let AI categorize adverse events?

Each of those questions has the same answer underneath it. The human work doesn't get skipped. Align on what we're measuring before any AI tool optimizes. Define what good data looks like before AI cleans it. Decide what "the right patient" means before AI finds them at scale. The work gets done up front, or it gets exposed once AI is operating on top of it. There's no third option.

None of this is new. The tools change every few years. The work underneath them doesn't. The teams that put the clarity work in first build trials that hold up. The teams that buy the tools first end up building clarity later, expensively, and usually under pressure.

12 · Where we land

The fax machine made noise. AI doesn't. That's our job now.

Thirteen years later, the fax machine is gone. The binders are gone. Most of the paper is gone. What hasn't gone away is the question I had to learn back then: what is this data actually for? Who needs it? What decision will it shape? Those questions don't get easier when the technology gets faster. They just get easier to skip.

The fax machine made noise. AI doesn't. That's our job now.