When Carlos Miranda’s team at Evergen began testing an AI agent to support donor tissue eligibility screening, the objective was clear – validate the technology before moving toward production. What emerged instead was a more uncomfortable finding – the testing process exposed weaknesses in the manual workflows the agent was meant to support.
Miranda explains:
We were seeing discrepancies between what the agent was identifying and what the human identified. When we went back and reviewed those, most of them – I would say about 90 percent or more – were either discrepancies because we didn’t have enough information in our documents to guide the agent to make the determination, or the human made an error. It was eye-opening for us to look back and say, this agent is already pointing out issues in our process, and this is only in test mode.
For Evergen, a regenerative medicine contract development and manufacturing organization (CDMO) working with human and animal donor tissue, this is not a theoretical insight. Eligibility determinations sit at the intersection of regulatory compliance, operational reliability, and patient outcomes. Errors at this stage carry material consequences.
Transformation under constraint
Evergen, formerly RTI, is in the midst of a wider business transformation following a series of acquisitions that expand its portfolio. The organization now operates as a CDMO, providing end-to-end services spanning manufacturing, quality assurance, and design for medical implants and devices.
Miranda leads Artificial Intelligence (AI), automation, and analytics at Evergen, having joined two years ago to build the capability from scratch within the Business Transformation Office. His background includes more than 15 years delivering AI and automation programs in large enterprises, but the mandate here is shaped less by scale ambitions and more by constraint.
The organization wants to grow volume and revenue without expanding headcount at the same rate. Workforce reduction is not the goal. Instead, Evergen aims to maintain what Miranda describes as an “agile footprint”. He notes:
Automation and AI has been a way of helping us maintain that agility without adding bloat.
Some attrition, Miranda acknowledges, is inevitable as disruptive technologies are introduced. He describes this largely as cultural rather than structural. For backend roles, Evergen now asks candidates about experience with generative AI tools – not as a requirement, but as an indicator of adaptability.
The 400-criteria challenge
The use case that earned Evergen an honorable mention at last year’s UiPath AI25 Awards focuses on donor tissue eligibility screening. The manual process requires analysts to assess more than 400 criteria across multiple documents to determine suitability. Despite years of training, the work remains highly susceptible to error.
Miranda is explicit that the organization is not attempting broad, indiscriminate automation:
We’re not really a hyperscaler. We’re not trying to automate everything that comes in our path. We want to pick and choose specifically use cases that we know are going to have a big impact.
That selectivity also shapes how the automation is designed. Rather than targeting individual tasks, Evergen focuses on end-to-end process transformation – including validation, auditability, and governance – rather than point efficiency.
Designing for compliance first
In tissue processing and medical device manufacturing, regulatory compliance is not a downstream concern. It shapes system design from the outset. UiPath was selected in part for its ability to log every action taken by the automation. Miranda explains:
The fact that it can log every single click that the automation is doing is a big factor for us when it comes to quality assurance, when it comes to compliance.
The resulting solution, named NEST, embeds human-in-the-loop validation at critical stages. Agents generate recommendations, but final determinations for GXP-controlled processes remain with human reviewers. Quality assurance teams are involved early, educated on the technology, and included in solution design.
Using UiPath’s Action Center, tissue processors validate both data extraction accuracy and eligibility determinations through an Azure portal interface. The automation runs in the backend, but all activity is logged and traceable. As testing progresses, attention shifts from the technology to the documentation itself. Miranda observes:
If you uncover the layer of the technology and really think about the process itself, it’s also questioning the process that we put together from a manual perspective and the documents that we’re feeding the agent. Are these documents easy enough for another person to read? And if they’re not, then how can we assume an agent is going to be able to also make sense of it?
Rather than feeding raw material into the system, Evergen invests time upfront in restructuring guidelines so they are optimized for AI consumption. Context, structure, and clarity become explicit design considerations. Miranda elaborates:
You need to guide it in a way that it will understand context behind it.
From internal capability to external interest
The initial deployment delivers a 2.5 percent operational expenditure reduction – enough to self-fund the initiative. That metric matters more than headline speed gains. Miranda emphasizes:
The ultimate metric was, can we self-fund this and can we also potentially have this as a service?
Before NEST reached full production, Evergen began receiving inquiries from partners interested in using the screening capability for their own donor workflows. The interest is earlier than anticipated and suggests broader applicability beyond internal use.
Next on the roadmap is expanding NEST from pre-screening into full medical director review. While pre-screening may involve around 100 pages per donor, full screening can involve thousands, including medical charts, lab results, and autopsy reports. That process currently takes around 90 days – a timeline Miranda believes can be reduced without removing clinical judgment from the loop.
My take
Evergen’s experience offers a useful counter to the dominant agentic AI narrative, which often frames progress in terms of autonomy, speed, and scale. In regulated environments, those measures are secondary. Documentation quality, auditability, and governance determine whether systems can be deployed responsibly. This is shown by how the organization responds when an AI agent surfaces errors. Discrepancies are not treated as a model tuning problem, but as evidence that existing processes deserve scrutiny. That response requires a higher level of organizational maturity than simply improving accuracy metrics.
The emphasis on documentation is especially important. In many organizations, procedural documents function less as complete instructions and more as prompts for experienced staff who know how to interpret them (worse still, as an afterthought). That tacit knowledge never appears in the workflow, but it keeps the system running. Agents cannot rely on that context. When AI fails in those conditions, it exposes the gap between what the process claims to be and how it actually operates – revealing where process discipline has been assumed rather than designed.
Compliance-first design choices reinforce discipline. Human validation is positioned as a structural feature rather than a transitional safeguard. That reflects a realistic understanding of how trust is earned in regulated systems – prioritizing visibility, traceability, and accountability over speed.
The business case is one to learn from, built to hold up under scrutiny. A 2.5 percent operational expenditure reduction is modest, but also sustainable and self-funding. It provides options, allowing the capability to evolve into a service rather than remaining a one-off efficiency project. AI is unforgiving of weak process discipline. Where documentation, oversight, and executive ownership are strong, agents can extend human capability. Where they are not, AI will surface uncomfortable truths long before it delivers efficiency.
