Every large regulated enterprise wants in on AI right now, and most of them are still stuck at the pilot stage. The reason usually is not the model itself – the model tends to work fine in a demo. What breaks the pilot is what happens after the model produces an output, when someone at the SEC or in a state insurance commissioner’s office asks a very concrete question and the answer had better not be “we trust the system.”
Regulated industries operate under scrutiny most software categories never have to think about:
- HIPAA in healthcare. FinCEN and NYDFS in banking
- The NAIC Model Bulletin, now adopted across 25+ US states in insurance
- The EU AI Act’s new obligations for high-risk systems
All of these push in roughly the same direction: if an AI-influenced decision ends up in front of a court, an examiner, or a journalist, the burden falls on the enterprise to reconstruct exactly what happened and why. That reconstruction is the actual product requirement. The model itself, however capable, is only a means to it.
The problem with black-box outputs
The current generation of generative models was built to be excellent at giving answers and, for the most part, not so good at showing its work. That is a fair trade-off for a consumer chatbot, but it is a much worse trade-off for a document processing pipeline where a decision might route a loan, deny a claim, or move a patient’s record into the wrong care pathway. When someone eventually asks “how did we get here,” pointing at the model is not going to help you.
With Hyperscience, our Explainable AI extends beyond model transparency to make the entire AI pipeline auditable end to end. Understanding how a model arrives at an output is important, but in regulated environments, organizations also need to demonstrate how that output was generated, validated, and acted upon throughout the workflow. The source document is preserved. The specific location of the extracted value on the page is recorded. Confidence is logged at the field level, not just as a document-wide average. Human interventions are timestamped, logged, and attributed. When an examiner arrives, you’re producing evidence rather than improvising explanations.
There are several big industries where we see this play out regularly.
Banking: KYC that holds up under review
KYC and AML programs generate an enormous volume of documents. Passports, utility bills, bank statements, paystubs, and the increasingly complicated beneficial ownership structures that come with corporate customers. FinCEN’s Customer Due Diligence Rule and, in New York, NYDFS Part 504 both require banks to prove they verified who they onboarded, not merely to assert it.
Modern IDP handles the extraction part well enough. The harder problem tends to show up two or three years later, when someone flags an account and the compliance team has to explain exactly how a particular officer’s name got captured from a particular filing.
Hyperscience can show the bounding box on the page, the confidence score at that specific field, and the reviewer who signed off, so everything is compliant in front of a regulator. Lacking the ability to produce this evidence chain creates a gap that becomes very expensive during an audit.
Insurance: Keeping automation defensible
Claims processing is one of the messier corners of insurance. A single adjudication can pull from a fifty-page policy, adjuster notes that are sometimes handwritten, medical bills and photographs. In that landscape insurers are under real pressure to close claims faster without opening themselves up to bad-faith litigation or regulatory action.
Field-level confidence scoring is what makes automation defensible in that environment. High-confidence extractions on routine fields can flow through automatically, while low-confidence or ambiguous ones (a smudged signature, a mismatch between two documents in the same case) get routed to a human reviewer along with the specific reason for the routing. This is exactly how the Hypercell is built.
This matters even more today, because the NAIC Model Bulletin on AI – adopted in 25 states as of mid-2026 – now explicitly requires insurers to document how AI-influenced decisions were made and to maintain human oversight on adverse determinations. “The model was 87% confident” is no longer a defense on its own. What works is being able to show which field, from which document, was reviewed by which human reviewer, at what time.
Healthcare: PHI (Protected Health Information) extraction, patient safety, and the paper trail
Healthcare is where the stakes tend to be the highest and the margin for error the smallest. Extracting Protected Health Information from unstructured records like referrals, prior authorization forms, and discharge summaries has to be right, and when it is not right, it has to be traceable. A transposed digit in a dosage is not just a compliance risk in the abstract; it is a patient safety event.
HIPAA has always required audit trails. What has shifted more recently is ONC’s HTI-1 rule, which came into effect over the past year and added specific transparency requirements around predictive decision support, including source attribution.
Systems that were previously getting away with vague explanations now have a much harder bar to clear. With Hyperscience, an explainable pipeline clears it by design, because every PHI element carries its own explainability: which document, which field on that document, which model version processed it, and which human, if any, confirmed the extraction.
The new bar
AI transparency was, until recently, treated as a differentiator vendors talked up rather than a capability buyers depended on. That is no longer the case, and the shift is happening quickly. Between the EU AI Act, the state insurance bulletins, HTI-1, and the general direction regulators are moving in, an AI system without explainability cannot make it from pilot to production in any regulated environment worth being in.
The upside is that the two goals of modernizing and staying compliant are no longer in conflict the way they once seemed to be. When every data point is traceable back to its source, every confidence score is visible, and every automated decision is auditable, you get the operational leverage you were promised, and you get the paper trail your regulator was going to ask for anyway. That is the version of AI we are building at Hyperscience, and the version we think will actually reach production.