Don’t Veer for Deer: Why Drug Development Needs Steady Steering in the Age of AI and Real-Time Data.

Authored by

Michigan summer comes with its own version of ICH guidance. 

Watch for orange barrels. Respect sudden storms. Leave room for boats, bikes, construction trucks, and the occasional driver on I-94 convinced they have been granted priority review. And, of course, remember the classic Michigan road advice: 

Don’t veer for deer. 

At first, it sounds wrong. If a deer jumps into the road, every instinct tells you to swerve. To move fast, protect yourself and avoid the object in front of you.  

But the driving advice exists for a reason. Swerving can turn a manageable hazard into a far bigger accident. You can lose control, cross into another lane, hit a tree, or cause a chain reaction. The safer answer is usually the harder one: stay steady, keep both hands on the steering wheel, brake firmly, and avoid overcorrecting. 

That advice feels very relatable to drug development right now. 

The “deer” are everywhere. 

FDA is exploring real-time clinical trials. AI-enabled technologies are moving quickly into clinical development. Real-world data are being used more visibly across regulatory contexts. Review models are becoming more dynamic. Sponsors are being asked to move faster, explain more, document better, and make decisions in an environment where the road is changing while they are driving. 

This can all be categorically called progress. But progress can still become a hazard if the response is panic steering. 

The question is not whether every sponsor should chase every new FDA curve ball, AI-enabled tool, real-time data model, or modernization pilot. Many will not apply. Some will be premature. Others may create more complexity than value for a specific program. The real question is whether sponsors can determine what matters, integrate it deliberately, and keep control of the program. 

The first curve ball: Real-time data may change the tempo 

FDA recently announced steps toward real-time clinical trials, including proof-of-concept clinical trials that report endpoints and data signals to the agency in real time. FDA also released a Request for Information on a proposed real-time clinical trial pilot program. The agency has described the RFI as focused on how AI-enabled technologies could improve efficiency, speed, and quality of decision-making in early-phase trials. 

These are not trivial initiatives for sponsors.  

Traditional drug development has long depended on staged review gates. The data are collected, cleaned, analyzed, interpreted, and then packaged for regulatory discussion. Even when timelines are compressed, the model still assumes a certain sequence. 

Real-time clinical trial data challenges the old paradigm. FDA’s proof-of-concept work with AstraZeneca and Amgen, including Amgen’s STREAM-SCLC study in limited-stage small cell lung carcinoma, suggests a future where regulators may see defined safety signals and endpoints much earlier in the development process. That can accelerate decision-making, but it also changes the discipline required. 

Real-time data are inherently messy. They are generated while the trial is still ongoing, before every query is closed, reconciliation is complete, and downstream interpretation has been resolved. That does not make the data less valuable. It means the evidentiary framework must be stronger from the start. 

The recent Nature Medicine retraction of the Chinese LungTIME-C01 NSCLC study is a useful cautionary counterpoint. The issue was not simply whether the result was surprising. The concerns involved study registration, protocol history, endpoint changes, eligibility criteria, sample size, reported design, source data, and discrepancies between protocol versions. That is exactly where confidence is either built or lost. 

Real-time does not mean less rigor. It means more discipline around what is considered a signal, what remains provisional, what has been validated, and what should or should not drive regulatory decision-making. 

If the data are going to move faster, the protocol, SAP, source documentation, safety review process, data flow, audit trail, and regulatory interpretation need to be stronger earlier. A dashboard cannot create evidentiary discipline. AI cannot compensate for ambiguity in the data story. And no pilot program, however innovative, replaces the need for disciplined regulatory strategy and experienced judgment. 

Faster data do not reduce the need for judgment. They increase it. 

The AI curveball: Not one deer, but a herd 

AI is one of the biggest issues in drug development right now, and perhaps the better analogy is not one deer in the road, but a whole herd. 

It is not entering the industry as one neat, contained tool. It is showing up everywhere at once: in drafting platforms, comment-generation tools, literature reviews, data review, coding, safety signal detection, quality checks, dashboards, and operational workflows. Each tool may be helpful in isolation. The problem begins when all these outputs start moving through a program without a clear connection to the clinical strategy, statistical rationale, safety interpretation, regulatory history, and submission plan. 

That is where we are starting to see real complexity. 

For example, many teams are now using AI to generate comments or suggested edits. On the surface, that sounds efficient. The comments may be well written, fast, and even technically reasonable. But drug development is rarely about one isolated point. A comment on a protocol, SAP, clinical summary, safety section, or briefing document may sound correct and still be wrong for that program. 

It may not reflect what was already discussed with the agency. It may miss why a prior decision was made. It may not understand the tradeoff between statistical purity and operational feasibility or it may fail to recognize that changing one sentence in one document can create a conflict with another document, a prior response, a planned endpoint strategy, or the final submission narrative. 

That is the real issue. AI can produce language. It can summarize and detect patterns. It can help teams move faster. But it does not automatically understand the history of a program, the nuance of an agency interaction, the reason a difficult decision was made, or the downstream consequences of a suggestion that appears harmless in isolation. 

FDA’s draft guidance on the use of AI to support regulatory decision-making for drugs and biological products makes this point in a different way. The focus is not simply on whether AI can be used. The focus is on the credibility of AI-generated information or data when it may support decisions about safety, effectiveness, or quality. FDA and EMA have also collaborated on guiding principles for good AI practice in drug development, which reinforces that AI use must be appropriate, risk-based, documented, and explainable. 

This is key because AI is not just a technology question. In drug development, it becomes a regulatory credibility question. 

If AI is used to generate, process, classify, summarize, detect, or interpret information that may influence a regulatory decision, a sponsor needs to be able to explain what happened. What did the model do? What data did it use? What was reviewed by humans? What was validated? What were the limitations? Was the model fit for the intended purpose? Was the output reproducible? Was there a risk of bias or error? And most importantly, did qualified people make the final judgment? 

These questions are beyond an academic exercise. They are reflected inFDA meetings, information requests, advisory committee preparation, inspections, and submissions. 

This is why I worry less about AI itself and more about disconnected AI use. A sponsor can have multiple tools generating comments, summaries, analyses, and dashboards, but if those outputs are not being governed by experienced people who understand the program, the result is not intelligence. It is noise with better formatting. 

I do believe AI can be extremely useful inside a disciplined framework where it is connected to the evidence strategy and reviewed by people who understand the science, the data, the regulatory history, and the end game. Otherwise, the program can spend more time reconciling AI-generated activity than advancing the actual development path. 

That is the herd-of-deer problem. 

The issue is not that AI is crossing the road. It is that several AI tools may be crossing at once, each from a different direction, each creating a different reaction, and none of them carrying the full context of the program. 

In that environment, the answer is not to panic, and it is not to pretend the road is empty. The answer is to keep both hands on the wheel: use the tools that are fit for purpose, govern them carefully, bring experienced humans into the review, and keep the regulatory strategy steady. 

The next deer in the road: Real-world data still have to earn their place 

Real-world evidence is another area where the road is changing. 

FDA’s CDER page now reports submissions containing RWE that meet defined reporting criteria, and FDA has continued to describe the role of real-world data and real-world evidence across the medical product lifecycle. 

This clearly shows that they have embraced RWE as a part of the regulatory landscape. 

But here is the danger: Sponsors often confuse data volume with evidentiary strength. 

From a regulatory standpoint, data volume is not the same as evidentiary strength. A registry, EHR extract, claims analysis, or dashboard may be valuable, but only if it can support questions about provenance, completeness, bias, endpoint validity, missingness, and fitness for purpose.  

The FDA curve ball is that RWE is becoming more available and more visible, but the evidentiary standard has not disappeared. 

More data sounds promising but poorly contextualized data can hurt. 

The final hurdle: Speed makes weak integration visible sooner 

Faster review models are another deer in the road. 

They sound attractive because everyone wants time back. Sponsors want faster decisions. Patients need faster access. Investors want faster clarity. Teams want to know whether the program is moving forward or needs to change direction. 

FDA’s modernization efforts reflect that pressure. The agency has highlighted actions to accelerate and modernize clinical development from IND-stage work through late-stage pivotal trials. FDA also has programs such as Split Real Time Application Review, or STAR, which was created under PDUFA VII and is intended to shorten the time from complete submission to action date for qualifying applications. 

The industry often hears “faster review” and focuses on the word “faster.” 

But the more important word may be “ready.” 

Speed does not solve weak integration. It reveals it sooner. If the clinical story is moving in one direction, the statistical rationale in another, the safety interpretation is still unsettled, and the submission strategy is being built separately, a faster review model will not make the program stronger. It will simply bring the disconnects to the surface earlier. 

That is why the end game cannot be treated as document assembly. By the time a program is moving toward a major FDA meeting, an accelerated review pathway, an information request, an advisory committee, or an NDA or BLA, the evidence story has to be coherent. Clinical, statistics, safety, data management, medical writing, regulatory strategy, and submission operations all have to be working from the same map. 

In Michigan driving terms, speed is only helpful if the car is aligned. If the wheels are pulling in different directions, pressing the accelerator is not progress. It is risk. 

The same is true in drug development. Faster pathways can be valuable, but only when the program is ready to move at that speed. 

Near the end, there is no room for disconnected excellence. The end game requires integration. 

Why this is important for sponsors 

In a changing regulatory environment, the natural instinct is to overcorrect. 

A new FDA pilot appears, and suddenly everyone wants to rethink their operating model. A new AI tool emerges, and the first question becomes, “Where can we use it?” A new data source becomes available, and the temptation is to add it before asking whether it is fit for purpose. A faster review pathway creates excitement, but the evidence package may not yet be ready to move at that speed. 

That is the drug development version of veering for deer. 

The answer is not to ignore what is happening on the road. That would be reckless. AI Real-time data, Real-world evidence and FDA modernization are all here to stay and important. Sponsors should take them seriously. 

But taking them seriously does not mean swerving every time something new appears. 

It means seeing the change early, understanding whether it truly applies to the program, and responding with control. Sometimes that means adjusting the strategy. It may mean slowing down long enough to assess the implications and deciding that a new tool, dataset, or pathway is interesting but not relevant to the specific question in front of the team. 

That is where regulatory strategy must be practical, not academic or impulsive. 

Regulatory strategy is not just reading guidance and quoting it back. It is understanding how today’s decision will look later, in a Type B meeting, an FDA information request, an integrated summary, a safety update, an advisory committee briefing book, a labeling negotiation, or an NDA or BLA review. 

This is also why end-game experience is so important. 

The end of development is not a formatting exercise. It is the point where every prior decision is tested. The endpoint choice, the SAP language, the safety interpretation, the missing data strategy, the RWE rationale, the AI governance approach, the submission architecture, and the regulatory history all have to come together into one coherent evidence story. 

By then, it is too late to pretend the road was straight all along. 

Good teams know how to keep moving without panic steering. They know when to brake, when to adjust, and when to hold the wheel steady. They do not ignore the deer, but they also do not let every deer determine the direction of the car. 

That is what sponsors need now: not less innovation, but steadier judgment around when and how to use it. 

Helping sponsors stay steady 

At MMS, we spend much of our time helping sponsors manage exactly this kind of complexity. 

Our work across the development lifecycle gives us a practical view of how early decisions show up later in the submission package. 

We are often involved when the road is narrow and the stakes are high: briefing books, regulatory responses, integrated summaries, NDA and BLA readiness, advisory committee preparation, safety narratives, transparency deliverables, and submission execution. 

This experience means that when a sponsor is considering AI, we are not only asking whether the tool is useful. We are asking whether the use is governed, documented, validated, and explainable. 

When a sponsor is considering real-world data, we are asking whether it is fit for purpose, traceable, interpretable, and aligned to the regulatory question. 

When a sponsor is moving toward real-time or accelerated review concepts, we examine whether the evidence story is ready to move at that speed. 

When a sponsor is preparing for the end game, we don’t start on document drafts from templates. Instead, we begin with the label and more, whether the science, data, safety, regulatory rationale, and submission architecture can withstand review. 

Therein lies the difference between activity and strategy. 

The road ahead 

AI will keep advancing. FDA will keep modernizing. Real-world evidence will continue to grow in relevance. Real-time data may become a more serious part of clinical development. Review models will continue to evolve. 

The challenge is not innovation itself but rather how innovation is integrated within the regulatory pathway. 

The companies that succeed will not be the ones that chase every flashing object. They will be the ones that understand the relevance, and stay disciplined enough to keep the program moving. 

The deer may be AI or real-time data or a new pilot, a new guidance, a new technology, or a new regulatory expectation. 

The answer is not to ignore it. We must see it early, brake with purpose, keep the wheel steady, and stay on the path to approval. 

In drug development, just like on a Michigan road, the danger is not always the deer. Sometimes it is the overcorrection.