Submit when the Science is Done
Learn how to make agentic AI trustworthy in clinical trials, with your experts in control, always.
In AI and the Economics of Time: Part 2 we discussed how the future winners in AI may not be the organisations that automate the most work, but those that unlock the most opportunity by compressing time and releasing capacity.
If acceleration creates value, then surely faster must be better? This is not assured!
One of AI’s greatest strengths is its ability to generate answers at speed, but it also introduces a challenge that is rarely discussed outside technical circles.
When information is incomplete, AI does not always stop and ask for more evidence. Instead, it attempts to produce the most likely answer from the information available. Sometimes that answer is correct. Sometimes it is not!
AI may not always distinguish clearly between:
….. what is known
….. what has been inferred
….. and what is simply missing
The result is not necessarily an obvious error. It may be a well-written, plausible, and convincing answer that is partially supported by evidence and partially constructed through inference.
This phenomenon is commonly referred to as hallucination.
In everyday applications it is an inconvenience. In regulated industries, and certainly in clinical development and regulatory submissions, it is a significant source of risk!
Let’s consider a hypothetical integrated clinical summary prepared using AI:
95% of the required information is available and correctly connected. The remaining 5% consists of disconnected datasets, inconsistent terminology, missing metadata, or incomplete links between supporting documents.
Rather than highlighting every missing connection, AI generates the most coherent narrative it can from the available evidence.
The resulting document reads well, the conclusions appear logical, nothing immediately suggests that information is missing. However, several weeks later, during regulatory review:
……clinical conclusions are challenged
……the supporting evidence cannot be located
……key analyses need to be repeated
……and the submission package must be revised
So, while the initial AI output was produced quickly, the delay came afterwards when the organisation had to establish whether it could be trusted. Investigation, validation, and rework can quickly outweigh the time saved through automation.
Every inferred relationship creates uncertainty. Every unsupported conclusion creates additional review. Every missing connection creates another validation step.
Imagine an AI-enabled submission process that reduces effort by 30%. This is an impressive achievement, but then uncertainty introduces:
….. three additional review cycles
….. two weeks of medical verification
….. further statistical validation
….. and additional regulatory quality checks
The labour saving remains, but much of the value created through acceleration disappears. The bottleneck has simply moved from generating information, to validating information.
As we established in Article 1, the greatest economic value comes not from saving hours but from compressing timelines. Anything that recreates delay destroys part of that value.
The instinctive response to this challenge may be to limit AI usage. However, the real challenge is not the model, but the environment in which the model operates.
AI performs best when the underlying information is complete and connected. When we can establish:
….. the original source of the data
…..what has happened to it since it was created
…..which transformations, calculations or analyses have been applied
…..ultimately, why the resulting answer can be trusted
The fewer the gaps in the information landscape, the fewer assumptions AI must make. The fewer assumptions it makes, the less uncertainty in the output. Less uncertainty means less validation and rework, and this means less time given back.
The answer is therefore not simply to deploy AI, but to deploy AI within an environment where the source and history of the underlying information are fully traceable.
Provenance matters because it allows acceleration to be achieved without sacrificing confidence.
The economic risk of AI is therefore not simply that it may produce an incorrect answer, it is that uncertainty can recreate the very delays AI was intended to remove.
An output generated in seconds creates little economic advantage if it then requires weeks of investigation, reconciliation, and validation.
AI creates value through acceleration. Uncertainty gives some of that value back.
In clinical trials, AI is only as good as the evidence it can see. The organisations that will realise the greatest return from AI will need to do more than generate answers quickly.
The winners will run AI on a fully deterministic, reproducible and traceable system, where every output is grounded in evidence and overseen by experts. This is what protects the time gained through acceleration.
This leads to one final question to complete our vision of AI and the Economics of Time.
“What happens when an organisation can not only generate an answer faster, but can trust it sooner?”
We will address this question in our final article.
At Verisian, our focus is on creating an infrastructure that maintains provenance from source data to output, reducing uncertainty and enabling AI to operate within a trusted information environment.
If you are interested in AI and the Economics of Time, I would love to chat.
Carol Collins has over 30 years leadership experience in the global Life Sciences and Healthcare industry, with 20 years in clinical development. Throughout her career she has had a key focus on innovation and transformation to accelerate outcomes, create value and bring life-saving medicines to patients. She is a strategic advisor to Verisian.
See Verisian in action. Share a few details and we’ll be in touch to schedule your demo.
Your request has been received. A member of the Verisian team will contact you shortly to arrange a demo.
