There is a particular type of engineer who moves from hyperscale cloud to regulated healthcare, and there are not very many of them. The skills overlap less than people assume. Cloud teaches you how to handle scale; healthcare teaches you that scale is the easy part when every action has to be defensible. Manu Agrawal made that move, and the playbook she is building from it is starting to attract attention.
Agrawal currently leads agentic AI work at Oracle Health. She got there by way of Amazon Web Services, where she spent her senior engineering years on systems most of the cloud industry uses every day. She was a founding member of an AWS service, a critical contributor on CloudFront, and worked on AWS Rekognition and AWS Textract. Her most consequential AWS project was the one that brought all of that together.
The Bedrock chapter
Amazon Bedrock Data Automation was Amazon’s first multimodal unstructured-to-structured data offering. The pitch was that customers could push documents, images, audio, and workflow data into the system and get structured output back, with multiple AWS AI services orchestrated under a single hood. Agrawal owned that project end-to-end.
“I did a zero-to-one, meaning I wrote the product strategy, as well as pricing, as well as full architecture of the system called Bedrock Data Automation, which was Amazon’s first multi-model, unstructured-to-structured data offering, which combined all the AWS AI services under a single hood,” she said.
That is the type of role where you have to think about what the market wants, what the architecture can support, and what enterprise customers will actually pay for, all at once. The customers running on those platforms include Intuit and Genesys, among other large organizations with workloads that cannot afford downtime.
The Oracle Health pivot
Agrawal moved to Oracle Health to apply the operational discipline she developed at AWS to a domain with different rules. Healthcare AI cannot fail gracefully. The systems she is building support care gap closure, clinical workflow automation, cohort discovery for medical research, and AI-assisted trial intelligence — work that runs inside healthcare providers and public-sector healthcare organizations operating across multiple countries.
The technical foundations transfer. Distributed systems thinking, observability discipline, the operational reflexes you develop running infrastructure at AWS scale: all of it applies. The novel pieces are clinical interoperability, governance constraints that did not exist in her previous work, and a much higher bar for human oversight.
What carries over is the instinct that AI capability is not the hard problem. The hard problem is everything around it: how the system is orchestrated, how decisions are logged, where the data flows, and how the AI hands work back to a clinician when it should.
The playbook
Agrawal’s playbook for reliable autonomous systems is starting to look like a coherent set of principles. Build for production from day one, not for the demo. Treat governance as architecture, not as compliance. Design human oversight into the system structurally. Assume the AI will be wrong sometimes, and engineer for what happens when it is.
That last point is what separates her approach from the bulk of enterprise AI work in 2026. Most teams design for the happy path and patch the failure cases later. Agrawal’s view is that you should design for failure cases first, because they determine whether the system can be deployed at all.
It is an unfashionable position in some quarters of the AI industry, where speed and capability still get more attention than reliability. It is also the position that tends to win in any domain where the cost of a mistake is high.
Healthcare is one of those domains. Financial services, public-sector services, and any environment dealing with regulated data are others. The market for engineering leaders who can build for that constraint is growing faster than the supply. Agrawal’s work reflects the growing demand for engineering leaders capable of operationalizing trustworthy AI systems in regulated and mission-critical environments.
