by Conor Flynn, Head of Innovation
AI can help regulated organisations draft documents, analyse information and complete routine tasks far faster than before. But faster work is not necessarily better work, and in regulated life sciences that distinction matters.
The question for leaders is not simply where AI can be introduced. It is what the organisation already does well, what is worth improving and what must not be lost in the process. Answering it takes an understanding of how the work actually happens, where professional judgement matters, how results can be verified and which controls are needed.
Our five-part webinar series, AI Without the Chaos, explored these questions with colleagues from Innovation, Quality and Regulatory, Operational Excellence and Project Delivery, and with our guest Steven Byrne, Principal SaMD Expert at Scarlet, a Notified Body. Five lessons emerged that are worth considering before your next AI initiative.
1. Recognise the AI already in your organisation
In most organisations, AI adoption is already happening informally. People use tools to summarise documents, analyse information, prepare reports and support everyday decisions. Some of this is delivering real efficiencies, even though it sits outside any formal programme.
The difficulty is that the organisation rarely has a full picture of which tools are in use, what information is being shared or how outputs are checked. For Quality, Regulatory, Clinical and R&D functions, that creates risks to data security, reliability, compliance and accountability.
Simply restricting access tends to push use further out of sight, which makes transparency harder to build. A better starting point is to find out what is happening: where AI is used, which activities benefit and where guidance or controls are missing.
Approved tools and clear policies are a necessary step, but approval is not the same as fitness for use. An approved tool can still be the wrong tool for a particular task, and a sanctioned use can still produce output that nobody is equipped to check. Guidance and training need to help people judge the task, not only the tool. Without visible support from leadership, use will simply stay hidden.
A question for leaders: Do we know how AI is already being used, and could our people tell an approved use from a fit one?
2. Start with what is worth improving, and what must not be lost
It is tempting to begin by asking where AI could be used. But an opportunity to use AI is not the same as an opportunity to improve the work.
Consider a medical device complaints process with rising cycle times. The obvious assumption is that AI could speed up complaint reviews or automate administration. But where are the delays actually occurring? Are teams working with incomplete information? Are handovers creating waiting time? Are people searching across disconnected systems? Or has the process itself become too complex? Introducing AI before answering those questions risks automating the inefficiency rather than removing it.
Improvement also has a second side. A complaints process can get faster while investigations get shallower, trend signals that matter for patient safety are missed, or reviewers lose their understanding of the product. That is not an improvement, whatever the cycle time shows. Before deciding where AI fits, be clear about what the process must protect: the quality of decisions, the traceability of reasoning and the professional judgement applied to each case.
So map the process with the people who do the work. Identify the steps that create delay, duplication or rework. Agree what better looks like and what must not deteriorate. Only then decide which opportunities need process redesign, which need existing technology and where AI adds value.
This draws on established Lean and Operational Excellence practice, and it matters for adoption too. As Jack Cunningham, Trinzo’s Margin Improvement consultant, puts it: “When you enable people to map their own processes, identify their own pain points and generate their own solutions, then that barrier to change is significantly reduced because, in a sense, the people become the change. They’re the ones driving it.”
A question for leaders: For each process we want to improve, do we know what it must protect as well as what we want to change?
3. Design AI into the work, one task at a time
Once an opportunity is clear, the next step is to decide precisely where AI belongs. A regulated workflow is a sequence of tasks, systems, information sources, checks and decisions, each with its own requirements for accuracy, consistency and oversight.
Generative AI is good at extracting, comparing, retrieving and drafting. Other tasks need deterministic, rules-based checks, verified source data or a decision by a qualified professional. Rather than asking AI to run a whole process, break the workflow into tasks and assess each one. Is it repetitive and time-consuming? Is the information it needs available and reliable? Does it involve interpretation or judgement? What happens if the output is incomplete or wrong?
That last question deserves the most weight, because fluent output can be wrong without any sign that it is. Colm O’Rourke, Trinzo’s Quality and Regulatory expert, put it this way: “There are two critical aspects. One is knowing how to use and interact with AI at a technical level, but the other side is knowing how to validate and review the output in the context of the relevant subject matter … it’s dangerous to have one without the other.”
Three consequences follow:
- Consistency is not accuracy. A tool can give the same answer every time and be wrong every time.
- Omissions are the hardest errors to catch. A summary can be accurate in every sentence and still leave out the adverse event that mattered. The hardest thing to find is the thing that is not there.
- The cost of checking can decide whether a use is worthwhile. A literature summary for a clinical evaluation that takes a specialist longer to verify against the source papers than it would have taken to write is not an improvement.
The same thinking applies to human review. A reviewer is a safeguard only if they can actually check the output: they can see the sources, have the competence to judge the content, and have the time and authority to reject it. A reviewer asked to approve fluent output they cannot reconstruct is providing a signature, not a safeguard.
The aim is to put the right capability on the right task. Sometimes that is a simple assistant supporting an existing process. Sometimes it is integration with established systems, structured checks and designed review points. Either way, changes can be introduced and tested one task at a time, building confidence before anything expands.
A question for leaders: For each AI-supported task, who checks the output, and could they detect a material error?
4. Match the route and the controls to the consequence
Responsible adoption in regulated life sciences is risk-based. That does not mean every AI use needs the same controls, or a controlled workflow at all.
Most AI-assisted work should remain ordinary professional work: a competent person uses a tool, checks the result and answers for it, to the same standard as any other work. Some uses do need a designed workflow with defined checks, records and monitoring. Some work should stay unassisted, because the judgement or the learning is the point. And some uses are not worth pursuing, because the output cannot be verified at a reasonable cost or an error could not be contained.
The choice should not follow the sophistication of the technology. A capable AI tool used by a professional who checks the output is still professional work. What should change the route is a material shift in who exercises the judgement, who has authority to act on the output, how dependent the process becomes on the tool, or whether an error would be caught before it caused harm.
Steven Byrne drew the same line: “If AI is being used by a competent person and the oversight is inherently baked into that process, there’s lower risk of you missing something. Whereas if you’re using an agentic workflow, an automated system that doesn’t have a human in the loop by default, you then have to define: how do I put a human in the loop or on the loop to review that output?”
He also warned that the ground moves: “The problem with the AI models is they’re updated underneath your feet.” A use that performs acceptably today may need reassessment as the model, the inputs or the intended use change.
Where a use does need formal controls, five questions matter most:
- Intended use: what task is AI supporting, and how will the output influence later decisions?
- Verification: can correctness be established, by whom, and at what cost?
- Traceability: are the inputs reliable, and can the sources behind an output be identified and checked?
- Oversight: who reviews the output, what competence do they need, and can they reject it?
- Monitoring and governance: how will changes in performance be detected, and are approved tools, responsibilities and supplier considerations documented?
One theme of our discussions with Steven was proportion: risk-informed controls rather than burdensome ones that discourage organisations from exploring worthwhile uses. The objective is not governance for its own sake. It is responsible progress with justified confidence.
A question for leaders: Could we explain why each AI use sits where it does, and what would make us move it?
5. Build the capability that makes it last
A successful pilot is a first step, not an outcome. A solution that works technically does not automatically become a valuable, sustainable part of how the organisation operates.
People need to understand how it supports their work, have confidence in its outputs, know their responsibilities and be able to recognise when something is wrong. The underlying process has to be robust enough to carry it. Leaders need to look past whether it works to whether it delivers the intended improvement: less time on administration, more consistent workflows, less rework and more capacity for higher-value work.
Working well with AI is also a skill. A useful output often involves providing context, refining instructions, challenging assumptions, comparing results and applying experience. The value may lie less in speed than in better work: a well-founded decision reached sooner, or an issue caught that would otherwise have been missed.
Kathryn Cullen, Trinzo’s Head of Client Delivery & Operations, has seen this in applying AI to project delivery: “We need to enable experienced professionals to use their experience and knowledge on the projects rather than getting bogged down in the administration and the governance that is required. That way, they’re getting more value to the project quicker.”
There is a further question alongside this. Some of the work AI takes on is the work through which junior professionals build their judgement: reading complaint files, drafting a first assessment, being corrected by someone more experienced. Removing it saves time now and can thin the pipeline of expertise later. That is not an argument for keeping unnecessary administration. It is an argument for deciding deliberately how expertise will keep developing when the first draft no longer comes from a person.
Start with uses that address real needs and can show value. Involve the people who will use them. Learn from what works and what does not, then scale what has earned it.
A question for leaders: Beyond the number of pilots underway, what evidence do we have that AI is making the work better?
From experimentation to work you can stand behind
Making AI work in regulated life sciences is as much an organisational challenge as a technological one. It brings together AI capability, Quality and Regulatory expertise, operational improvement and the judgement of the people closest to the work.
There is no single right starting point, but there is a common principle: begin with the work that matters, be clear about what must be protected, make sure the output can be checked, and build on demonstrated value. Whatever route a use takes, a professional still owns the result. The machine can answer, but it cannot be answerable. You can.
Our webinar series AI Without the Chaos explores each of these themes through practical demonstrations, expert perspectives and real-world considerations for Quality, Regulatory, Clinical and R&D teams.
As AI takes on more evidence-related work, a further question comes into focus: how do we know an AI-supported conclusion is reliable, and can we show the evidence and reasoning behind it? That is the focus of Trinzo’s next webinar series, Method Over Magic: Building AI You Can Defend starting on 10 November. Register to attend via the link below:
