The FDA's AI Religion: Elsa Dons A HALO
By Louis Garguilo, Chief Editor, Outsourced Pharma
The FDA calls on your CDMO; inspectors are fortified by data in AI models, and appear practically clairvoyant.
“It won’t be like the ‘old days,’” says David Grote of GrayMatters Partners, an organization that advises on quality systems and inspection readiness.
“Until today, when the FDA inspected a facility, you pretty much knew what they were going to look at. They’d review 15 or 20 SOPs and certain batch records, your messy deviations. You and your CDMO would have your playbook ready. Your tour route was planned – where to stand, who would talk in each area. In certain regards, it was a surface-level examination of systems and facility.
Now the FDA has Elsa, an AI-enable data architecture allowing the agency to access all your electronic records – maybe hundreds of thousands of pages. The FDA will pull out from those documents where things don't match.
This batch record doesn't sync with that SOP. An operator's training record says she wasn’t trained on certain procedure. These CAPAs are not properly closed; there’s a string of them that appear related to one another. AI makes those connections, prioritizes, and informs the inspectors of areas of concern.
It is a different world we're in, and now it’s an arms race.
When regulatory agencies adopt AI, your CDMOs have no choice. They must adopt AI themselves to gain the same superior powers to evaluate their own systems.”
And now, along with Elsa, the FDA has a HALO.
Elsa With A HALO
It is described as an agency-wide generative-AI system to assist FDA employees, such as scientific reviewers and investigators. Elsa resides in a secure government-cloud environment.
FDA assures us any information submitted by drug makers and partners is not used to train the models. At launch, FDA mentioned Elsa would, among other activities, identify high-priority inspection targets.
A few months later in December, FDA announced deployment of agentic AI capabilities for use by all employees. Agentic AI at the FDA is designed to handle complicated, multistep workflows, including those related to “inspections and compliance,” as well as premarket reviews, post-market surveillance, and all review validations.
Then in May of this year, the agency announced Elsa 4.0, and put a HALO on it – a data platform officially deemed Harmonized AI & Lifecycle Operations for Data.
This consolidates more than 40 previously disparate application, submission, data and portal systems across its centers.
Elsa integrates with HALO facilitating FDA personnel queries across all that data, and builds workflows without having to manually feed documents into the AI system. (That would have been a real bummer.)
(To be clear, my more spiritual halo analogy breaks down a bit here – technically speaking, Elsa the AI sits on top of HALO the data-collection system. Oh well.)
FDA Is Explicit: AI Into Inspections
When FDA first launched Elsa in 2025, it had already been employing it to “identify high-priority inspection targets.” We are told the upgrade to Elsa 4.0 includes a deployment of agentic AI capabilities to all employees.
So as I write this in October 2026, how’s it all working?
It’s a bit unfair to ask that specifically of the HALO addition, just a few months since its existence was announced. Stories from manufacturing floors will assuredly begin to arise soon.
Regarding Elsa, so far there’s no definitive identification of a facility or inspection that was enhanced by AI. (If someone knows differently, let us know.)
Here’s a suggestion:
FDA should start to talk about uses and results of AI-enhanced inspections as soon as feasible. How about in (near) future some comparisons of pre-Elsa and post-Elsa inspections?
Here’s Grote’s advice:
Drug sponsors should be undertaking serious discussions about how external partners are ready, willing, and able to respond in AI-kind to Elsa + HALO wielding FDA inspectors.
Early Issues Now Overcome?
There’s another rationale the FDA to provide us as much detail as possible on HALO and future upgrades or AI adjustments.
The agency must continue to assure the industry of the accuracy and fairness, and efficacy of the systems, and the security of all that facility data now in the hands of AI tools.
Is the FDA’s AI tools improving quality – and inspection/audit productivity – at the facility-by-facility level?
Not long after the initial Elsa launch, FDA found it was capable of producing “false citations and hallucinated information,” certainly not unusual as initial rollouts go, but still raising early questions about human verification.
The FDA (quickly, I’ll add) then moved through Elsa 2.0, 3.0 and now on to Elsa 4.0, at each step emphasizing human oversight.
By the way, where does the technology originate?
It appears the FDA created Elsa internally as a fit-for-purpose system; the agency says the AI was developed in a secure, government-cloud environment. Because Elsa is an LLM-powered application, the FDA is building the secure application, interfaces, data environment and agency-specific capabilities around underlying AI technology.
As for HALO – and again, HALO is a data architecture, perhaps not best described as “AI” itself – it was also internally developed for the most part, but based on FDA budget items, heavily supported by outside contractors.
Will Elisa + HALO change the scale and sophistication of inspections and audits?
The FDA is simultaneously experimenting with such tactics as one-day inspectional assessments, something piloted in April (2026).
The agency is moving forward with all the tools of the times. You and your CDMOs need to stay apace. As Grote says, we are in an arms race. But this one should be a positive build out for everyone involved, including patients.