Ask any quality director to walk you through their last significant CAPA and you will usually hear a story about a person. A coordinator who missed something. A site that did not follow the current version. A new hire who had not been fully trained. The corrective action, predictably, targets the person: retraining, re-acknowledgment, a memo, a meeting.

Six months later, the same category of deviation happens again, often at a different site, involving a different person. Because the person was never the root cause. The information delivery system was.

The retrieval problem nobody budgets for

Consider what we actually ask of a clinical research coordinator running several concurrent studies, which is to say, a typical coordinator. A typical oncology study can add as many as seventeen additional systems on top of what the site already runs, each with its own login (Applied Clinical Trials, 2025). Nearly seventy percent of sites report six or more separate logins per study, and in Advarra’s 2023 Study Activation Survey of more than five hundred sites, fifty-five percent ranked sponsor technology setup and training as their single greatest startup burden, ahead of budgeting and contracting. Layer on staffing: more than sixty percent of site professionals report shortages, and new research staff can take up to a year to fully ramp (2024 SCRS Site Landscape Survey; WCG).

Inside that maze lives every piece of information a coordinator needs to execute a protocol correctly: the current visit schedule, the amended lab kit instructions, the eligibility clarification from last month’s memo, the sample handling requirements that differ between the two studies she is running this morning.

Here is the point that reframes much of it: the answer almost always exists. Sponsors and CROs are diligent about producing and distributing information. The failure is often not production. It is retrieval, at the moment of need, with a patient in the chair.

Anatomy of a deviation

Walk through how a deviation actually happens, and the pattern is remarkably consistent.

An amendment is issued. It goes out by email and gets posted to a portal. The coordinator is mid-visit on another study when it arrives, and the days that follow are full. Formal retraining is scheduled, but it happens weeks before the affected procedure actually occurs. By the time that procedure is in front of her, at Visit 3, month four, the training is a faded memory competing with the procedures of two other protocols.

In the moment, she has a vague sense that something changed. Verifying means hunting: the inbox, the binder, the portal, a call to the CRA who may or may not pick up. The patient is waiting. So she makes a judgment call, and her judgment is good, because she is good. Most of the time the call is right.

Some of the time it is not, and the industry’s published analyses of deviation root causes read exactly like this story: inadequate or faded training, missed or misapplied amendments, delayed communication of changes, and inconsistent updates across sites that let some locations revert to outdated procedures. Add coordinator turnover, which runs high across the industry, and the institutional knowledge that would have caught the error keeps walking out the door mid-study.

Much of this is not negligence. It is a delivery architecture that relies on email, memory, and goodwill to move safety-critical information across the last mile, into the hands of the person performing the procedure, at the moment they perform it. People stay in the loop, so deviations will still happen; the opportunity is the preventable, information-driven subset.

Why the standard fixes underperform

The industry’s reflexive answers are more training and more communication. Both matter, and both hit the same wall: timing. Training delivered at site initiation decays before it is used. Communication pushed through email competes with everything else in the inbox. Read-and-acknowledge tracking proves a document was opened, not that its content will be recalled at the right moment three months later.

Portals were supposed to help and mostly added another login. Single sign-on initiatives help at the margin but do not change the fundamental shape of the problem, which is that the knowledge lives in one set of places and the work happens in another.

The fix has to invert the model. Instead of pushing information out and hoping it survives until the moment of use, the current, approved answer has to be available at the moment of use, inside the system where the work is already happening, specific to the trial, the site, and the role of the person asking.

What just-in-time guidance has to get right

This is where AI enters the conversation, and where the conversation usually goes wrong. A general-purpose chatbot pointed at a folder of documents is a liability in clinical research. A plausible wrong answer about a visit window is a deviation with good grammar. If AI is going to participate in trial conduct, the bar is categorically different from consumer AI, and it is worth being explicit about what that bar is.

First, closed-world answering. The system answers only from the approved documents for that specific trial: the protocol, the manuals, the training, the SOPs. If the answer is not there, it says so and routes the question to a human rather than improvising.

Second, citation as a default. Every answer names its source document and section, so the coordinator can verify in seconds and a monitor can trace any piece of guidance to its origin.

Third, scoping. An agent serving a multi-study site must be sealed per trial. The coordinator running three protocols cannot risk an answer from the wrong one, so the agent has to know which trial, which site, and which role it is serving, and answer only within that boundary.

Fourth, auditability. Every interaction logged and reviewable, so that the question “how do you know your sites received and understood this change?” has a better answer than a forwarded email chain. Done properly, the guidance layer becomes inspection evidence rather than inspection risk.

Fifth, and least appreciated: placement. Adoption fails when help requires another login. Guidance has to live inside a system the site staff are already in all day. For most trial work, that means inside the CTMS itself, surfaced when the user logs in, available while they work.

The measurable claim

When information delivery moves from push-and-hope to surfaced-at-the-moment, the effects show up in metrics sites and SMOs already track: deviation rates after amendments, query volume on affected procedures, ramp time for new staff, and the variance between sites running the same protocol. That makes this unusually testable for a new category. Pick one live trial, instrument one cycle, and compare.

The coordinators were rarely the root cause. They have been compensating for a broken delivery chain, brilliantly, for years. It is time the information met them where they work.

Caleb Costa is the CEO of Revinova, which pairs site readiness training with per-trial AI agents embedded in the CTMS, drawing only from the current approved study documents, with the source and version shown.

Placement notes

• For Applied Clinical Trials or DIA submission, drop the bio line’s product description to a neutral one-liner and remove nothing else; the piece intentionally sells the category, not the product, until the final section.
• For LinkedIn, publish as an article and excerpt the “Anatomy of a deviation” section as the teaser post.
• The five requirements in “What just-in-time guidance has to get right” double as your analyst briefing framework and your RFP answer structure. Reuse them everywhere; repetition builds the category.

A contrarian’s case for running directly at the thing everyone else is fleeing

Everyone keeps asking the wrong question.

They gather in conference rooms, post breathless threads, book panels at Davos, and ask: How do we stop AI from taking over? How do we regulate it, slow it, contain it, make sure it stays in its lane? How do we protect jobs, preserve humanity, keep the machines from getting too smart?
Wrong question. Coward’s question.

Here is the right one: What do you plan to do while everyone else is busy being afraid?

I am not here to reassure you that AI won’t disrupt your industry, eliminate your role, or fundamentally alter what it means to be useful in the economy. It will do all of those things. The people telling you otherwise are selling you comfort food. I am not selling comfort food.

What I am telling you is this: the blast is coming, it is already here, and the worst place you can possibly be is fifty yards behind the people running toward it.

The Myth of the Safe Distance

There is a popular coping mechanism among smart, experienced professionals right now. It goes like this: AI is a tool. I will use it selectively. I will stay in control. The fundamentally human things I do (relationships, judgment, creativity) will protect me.

This is a story people tell themselves while the ground shifts under their feet.

The assumption embedded in that logic is that AI will advance far enough to be useful but not so far as to be threatening. That there is some stable middle ground where you get the productivity benefits without the existential pressure. That the blast has a safe radius.

It does not.

The professionals who thrive in the next decade will not be the ones who used AI carefully. They will be the ones who ran straight at it, got burned a little, figured out where the edges were, and came back knowing things the careful people will spend years trying to learn from a distance.

Distance is not safety. Distance is just a slower kind of obsolescence.

What “Destroying Humans” Actually Looks Like

Let’s be honest about what AI is actually destroying, because it is not humans. It is a particular version of humans that most people were not that excited about to begin with.

It is destroying the human who spends four hours reformatting a report no one reads. The human who sits in a meeting to explain a slide deck that should have been an email. The human whose entire professional identity is built on knowing something that can now be retrieved in eleven seconds. The human who charges for effort rather than outcomes. The human who confuses being busy with being valuable.

That version of you? Gone. And honestly, good riddance.

What AI cannot destroy is the human who builds things other humans need. Who understands a customer’s problem better than the customer understands it. Who makes the call when the data is ambiguous. Who has enough scar tissue to know which risks are worth taking and which numbers are lying. Who can sit across from someone in genuine distress and make them feel like they are not alone.

The destruction being promised is real. The victim they have cast for the role is not.

The Blast Has a Logic

Here is what people miss when they treat AI as pure threat: it is not random. It follows a pattern. It eats tasks before it eats roles. It eats roles before it eats functions. It eats the bottom of every skill distribution before it touches the top.

Which means the blast has a logic, and if you are willing to run toward it, you can read that logic faster than everyone retreating from it.

Right now, in real time, AI is showing you exactly what the market values. Watch what it automates first. That is the market telling you, in the clearest possible language, what it was only tolerating about you, what it was never truly paying for.

Watch what it cannot automate. That is your blueprint.

The people running away from AI are, functionally, running away from the most accurate market signal of their careers. The people running toward it are getting a graduate-level education in what actually matters, paid for by the discomfort of everyone who stayed behind.

Running Toward the Blast is Not Recklessness

Let me be precise about what I mean, because there is a version of this that is just hubris with better vocabulary.

Running toward the blast does not mean abandoning judgment. It does not mean deploying AI carelessly, trusting models blindly, or confusing novelty with value. The person who runs toward the blast and gets killed by it is not brave. They are just dead.

Running toward the blast means being the person who finds out what it actually does before they have an opinion about it. Who runs the pilots while everyone else is writing white papers about whether to run pilots. Who builds fluency in something threatening because fluency converts threat into leverage. Who accumulates real experience in the space where everyone else is still debating whether to show up.

It means accepting that you will be wrong about some of it. That you will bet on tools that fizzle, strategies that miss, approaches that do not work. And it means knowing that those losses are worth more than the clean record of someone who never risked being wrong because they never risked anything at all.

The People Who Win the Next Decade

They are not AI researchers. They are not necessarily technical.

They are the people who understand a domain deeply enough to know what a good AI output looks like and what a dangerous one looks like. Who have spent enough time with real customers, real problems, and real consequences to know where the model is full of it. Who can translate between the language of business and the language of systems without losing the meaning in either direction.

They are the people who got uncomfortable early, on purpose, before they had to.

They are the people who, right now, while this essay is being written and debated and shared and dismissed, are already three iterations past the version of AI everyone else is still arguing about.

They ran toward the blast.

They came back knowing something.

One More Thing

The people most afraid of AI are, with some exceptions, the people who built their careers on doing things that should have been automated years ago. The fear is not really about AI. It is about what AI makes visible: that a lot of what passed for expertise was really just friction dressed up as value.

That is not a comfortable thing to say. It is also not a comfortable thing to be.

If you read that last paragraph and it does not apply to you, then you have nothing to fear. Run toward the blast. It will not destroy you. It will clarify you.

If you read it and felt something uncomfortable move in your chest, then you have a choice. You can keep building your case for why AI is dangerous and unfair and moving too fast and needs to be stopped.

Or you can figure out what you actually do that matters, and go do more of it, faster, with better tools, before someone else does.

The blast does not care which one you choose.

Neither does the market.

The author runs an AI agent platform company. He is either right about this or extremely motivated to believe he is.