Only 12% of agentic AI initiatives reached production over the past year. The challenge isn’t the technology, but ensuring AI can operate within highly regulated environments. Natosha McNeal, of agentic platform Camunda, investigates.
Life sciences is not short of ambition when it comes to AI. The budgets are in place, and pilots are running widely across research and manufacturing.
However, AI that has reached live operations remains rare. For example, recent data from the healthcare sector highlights that only 12% of agentic AI use cases reached production over the past year.
The reason it persists is that life sciences carry a responsibility most sectors do not have.
Mistakes have significant consequences, from regulatory exposure to worst-case scenarios impacting patient safety. The question for life sciences leaders is not whether to adopt AI, but whether they can trust it inside clinical and operational processes that allow no room for error.
Trust is the real barrier
The AI models are becoming increasingly capable, and potential use cases span the entire sector, from faster drug discovery to more efficient supply chains.
But trust remains a barrier, set against the regulatory environment the life sciences sector operates within.
As operations run under the Good Practice quality rules (GxP), an audit can come at any time, and a compliance failure carries financial and legal penalties. Any computer system used in a regulated process must also be formally validated before it goes live.
There is also no tolerance for error, because consequences reach patients. In a field such as cell therapy, a single misstep is a significant safety event, so caution about handing decisions to software is entirely rational. AI must earn its place once the foundation can be trusted.
Agents can ’t be left to run alone
This is where agentic AI in particular runs into difficulty. An agent reasons and then acts, and the same instruction can take different paths depending on context and the tools available.
That adaptability can be a benefit, but it also makes an agent hard to defend in a regulated setting. Left on its own, an agent can take an action nobody signed off on and leave no record. For an organisation that must account for every decision to an auditor, that is not a sustainable position.
The instinct in some industries is to push toward greater autonomy, letting agents handle more with less intervention.
But in life sciences, that same instinct to chase fully autonomous agents is actually what holds the sector back. An agent that cannot be supervised or tied to a defined process will never be cleared to operate where the stakes are real.
Orchestration as the enabler
For most life sciences organisations, the missing piece is architectural. An individual agent, whether built in-house or bought from a vendor, cannot provide this control alone, so what is needed is an orchestration layer that sits above the agents and the systems they draw on.
This is the role of agentic orchestration. Rather than letting an agent act freely, an orchestrated process places enforceable steps between an AI ’s decision and the action.
This provides a crucial step for important operations such as batch releases, where a human will likely be required to approve before the process is complete.
The agent still does the work, but inside the guardrails the business sets, and on top of the validated systems a company already runs, rather than forcing a costly rip-out and revalidation.
The same logic applies wherever the stakes are high. In a quality-control laboratory, this approach can safely automate manual processes to connect systems such as the laboratory information management software and electronic lab notebooks with instrument data, so the right information is available at each step which can lead to reduced turnaround time.
In a clinical trial, where data is scattered across many systems and moves between teams without losing its chain of evidence, the same orchestration coordinates those handoffs, so information reaches clinicians and reviewers at the point they need it, while every action stays recorded for audit and submission.
Because the controls are built into the process rather than the agent, autonomy can rise gradually as each part of the workflow earns trust, which is what makes it safe to give AI real responsibility where it matters most.
Separating success from stalled pilots
Agentic AI capabilities will keep advancing, and the pressure to use it will only grow. The life sciences organisations that move ahead will not be the ones running the most pilots.
They will be the ones that re-engineer the underlying process first, so that when AI is ready to take on more, it steps into something that can already be trusted and audited.
That foundation is what turns AI from a promising experiment into a dependable part of life sciences operations, and it separates the companies that scale from the ones that stall.
