AI data security for regulated industries, from the answer your AI gives to the records underneath it. Triplets goes deep where a wrong answer is a regulatory event. NoPII and the Vault go wide, because sensitive data in AI traffic and at rest is universal. Find your industry below, at full depth.
Every industry carries some mix of three exposures: the answer its AI gives, the data its AI sees, and the data it stores. Each exposure maps to one product.
Exposure 01
The answer its AI gives.
Triplets certifies AI answers against compiled evidence and refuses past the boundary.
Triplets
Exposure 02
The data its AI sees.
NoPII tokenizes sensitive data in AI traffic before it reaches any model provider.
NoPII
Exposure 03
The data it stores.
The Vault holds cards, fields, files, and customer intake so your systems never do.
The Vault
Each industry section below tells you which exposures dominate, where they actually occur, and what teams in that industry run.
Healthcare AI data privacy
TripletsNoPIIThe Vault
The deepest exposure in the market, on every axis at once. Clinical AI is deployed, not pending; patient data is the most regulated data there is; and a wrong clinical answer has a name attached.
The answerTriplets
A fabricated clinical answer is a safety event.
Clinical questions are multi-document by nature: a formulary, an interaction database, a protocol, a label, and the literature all bear on one answer. Retrieval systems serve whichever fragment ranks highest and let the model fill gaps, producing the failure clinical settings cannot absorb: the answer that is wrong and sounds right. A confident wrong response delivered to a healthcare professional is a reportable event, and governance committees hold approval mandates with no tooling behind them, so launches stall on one question: how do you know it will not fabricate? Teams run Triplets to constrain medication, protocol, and formulary answers to compiled, verified clinical knowledge, answered deterministically where the data allows and refused in writing past the boundary. The 48-hour Trust Baseline gives the governance committee a fingerprinted certificate to gate go-live on, continuous certification keeps it green as labels change, and certificate vocabulary aligns with the GRADE-style evidence hierarchy reviewers already use.
The promptNoPII
The PHI leak nobody mapped.
Trace one ambient documentation session: the clinician greets the patient by name, the patient recites a date of birth, describes a diagnosis, and the transcript, which is now the prompt, lands in a model provider's logs. The same trace runs through chart abstraction, where the record is the input; coding, where the encounter note carries the MRN; and prior authorization, which ingests patient notes and orders by the millions. Enforcement keeps tracing to one finding, a defective risk analysis, and unmapped LLM traffic is precisely that defect. A BAA does not close it; it allocates liability after the data moves. Teams run NoPII in front of ambient documentation, chart abstraction, coding, and prior auth, tokenizing identifiers before prompts reach the model provider, with BAA available and a defensible line for the risk analysis: the provider never held an identifier.
The dataThe Vault
Custody is a statutory duty.
Records, insurance cards, IDs, and consents accumulate across systems never designed to hold them, and every referral inbox becomes an unencrypted PHI archive. Teams run Data Vault for HIPAA-grade field encryption with exact-match lookup on EHR fields, File Vault for records and imaging with expiring, tracked access, Customer Vault to replace fax and email for patient intake, and Card Vault to bill patients without PCI scope in the practice.
Who runs this stack
Health systems, clinics, digital-health vendors, medical information teams, and AI governance committees.
See your own PHI exposure this week. Route one workflow through the NoPII free tier and read the detection log, or bring your governance question to a demo.
The industry where evidence is the product and the patient data is mandated.
The answerTriplets
The evidence base everyone is about to reuse.
A systematic review cycle starts, an HEOR dossier is due, a submission is assembled from synthesis produced months ago, and nobody can say which claims have been superseded or where sources now conflict. Teams run Triplets to compile a therapeutic area end to end, trials, labels, literature, prior reviews, with every claim's state preserved, conflicts reconciled with a verdict on which source governs, and the certificate kept green as evidence lands. One manual systematic review costs roughly $141,000 in labor; a certified corpus is scoped below one review and de-risks every team drawing on the evidence, including CRO and HEOR partners who deliver certification as a service.
The promptNoPII
Mandated workflows on regulated narratives.
Adverse event processing is a regulatory obligation, and each case narrative arrives as prose dense with patient identifiers; AI triage makes that narrative a verbatim prompt. Trial operations leak the same way through EHR pre-screening, monitoring agents reading patient records, and submission drafting. EU trials add the sharper layer: regulators have said a trained model cannot be assumed anonymous. Teams run NoPII between PV and clinical systems and every model call, with deterministic tokens keeping case narratives coherent for causality assessment and a per-transaction audit trail for inspection readiness.
Who runs this stack
Medical affairs, HEOR, regulatory, pharmacovigilance, clinical operations, CROs, and clinical AI vendors embedding certification under their own brand.
Bring your inspection-readiness question to a working session.
The profession where the AI failure modes have already made case law, and where the heaviest AI adoption sits on the worst data.
The answerTriplets
Invented authority fails in front of a judge.
A research tool that fabricates a case or answers from a superseded clause does not fail quietly; it fails in a filing, and the sanctions record shows courts have no patience for it. Firms also sit on genuinely contradictory knowledge, diverging precedents, conflicting clauses, superseded guidance, and a system that silently picks one version is a malpractice generator. Teams run Triplets to ground firm knowledge and research in compiled, verified sources with provenance on every answer, conflicts surfaced with a governing verdict, and refusal in writing past the boundary. Legal AI vendors embed certification so the buyer's governance review becomes the demo.
The promptNoPII
Privilege does not survive a provider log.
Open one mass tort claimant file: name, DOB, SSN, address, complete medical history, employment and financial records, lien data, sometimes a minor's records, and AI-drafted chronologies make that entire file prompt material. Document review sweeps whole custodian mailboxes where the PII is whatever exists; privilege review sends the most protected material a firm holds by definition; deposition summaries carry volunteered details no redaction policy anticipated; intake leaks before an engagement letter exists. With hundreds of thousands of mass tort actions pending at any time, this is the volume center of the profession. Bar ethics opinions now require confidentiality analysis before client data touches a generative tool, and court standing orders govern AI use directly. Teams run NoPII in front of chronology, drafting, review, and summarization tools, tokenizing client names, parties, matter identifiers, and medical details before any prompt leaves the firm, with deterministic tokens keeping thousand-page chronologies coherent.
The dataThe Vault
The client file was never supposed to live in the inbox.
Retainers, IDs, and discovery materials arrive by email and stay there forever, and sharing with co-counsel creates standing access nobody revokes. Teams run Customer Vault for branded no-login intake with e-signatures for retainers, File Vault with ephemeral keys for one-time, self-destructing co-counsel access, Data Vault for encrypted client identifiers that keep conflict checks working, and Card Vault for retainer billing without PCI scope.
Who runs this stack
Mass tort and PI practices, litigation support, immigration practices, firm operations, corporate legal departments, and legal tech vendors.
Give your ethics analysis its missing artifact. The per-request detection log answers the confidentiality question in writing.
The densest regulated data per customer anywhere, plus decisions that must track an external rulebook changing thousands of times a year.
The answerTriplets
Version confusion is an enforcement category now.
Advisory bureaus issue thousands of circulars a year, on the order of twelve thousand discrete changes, and in most states a new loss-cost filing applies automatically to an insurer that takes no action, so missing a circular is an affirmative, unreviewed change to what you have on file. Regulators have cited carriers for calculating on superseded base rates and deviations, version confusion rather than actuarial error; in one case a single rating error ran undetected for three years and ended in more than $18 million of restitution across 80,000 policies with no fine at all, because detection was the failure. Undercharges are forfeited, so a rating error costs twice. Model AI guidance now names reproducibility, traceability, version control, and data lineage as expected properties. Teams run Triplets to make rate filings, bureau adoptions, and coverage logic standing decisions under continuous surveillance, with superseded states tracked, alerts only when a decision's support changes, and point-in-time reconstruction as the exam-readiness artifact.
The promptNoPII
Four regulated categories in one PDF.
An adjuster drops a claims file into the summarization tool, and that single attachment holds medical records, financial details, an identity document, and payment data. Triage and correspondence drafting repeat the transfer across the book of claims; underwriting and fraud carry the same shape. Teams run NoPII as one control point in front of claims, underwriting, and service AI, tokenizing across every category in the same pass, with every detection logged for the market-conduct exam.
The dataThe Vault
Policyholder custody spans decades.
Premium payments mean stored cards, stored cards mean PCI scope, and claim evidence still arrives as email attachments. Teams run Card Vault for premium billing through the existing processor via our proxy, Data Vault for policyholder records searchable across lines of business, File Vault for claims evidence with expiring access, and Customer Vault for agency intake of applications, disclosures, and claim documents without an engineering project.
Who runs this stack
Carriers, TPAs, MGAs, insurtech vendors, actuarial and filing teams, and the consultancies that manage filings on carriers' behalf.
See what one claims file actually contains. Run a redacted sample through the live playground.
Regulated data under PCI DSS, GLBA, and SOX, with AI features the market already expects.
The answerTriplets
A misquoted rate is a compliance exposure the moment a customer acts on it.
Rates, fees, eligibility, and disclosures are exact values living in documents that change on their own schedules, and a retrieval system that surfaces the outdated page quotes a rate that no longer exists, in writing, to a customer who will act on it. Teams run Triplets to answer product terms deterministically from verified values with superseded versions tracked as states, refusal past the boundary, and, for compliance teams, standing obligations monitored against changing regulation with point-in-time reconstruction. Wealth management points the same machinery at plan documents and the rules they track.
The promptNoPII
That paste is the leak.
A support agent pastes a case into the copilot: name, full account number, transactions, complaint, hundreds of times a day per team. Fraud narratives leak by design, since the narrative is the transaction history with a name attached; compliance review leaks in bulk; and engineers paste connection strings and API keys into internal copilots daily. Teams run NoPII to tokenize financial identifiers in transit with deterministic tokens preserving multi-step reasoning, and secret detection across fourteen credential pattern families in the same pass, on by default.
The dataThe Vault
The card table was never supposed to exist.
Recurring billing means stored cards, and the processor's vault solves storage by locking your data inside one vendor. Teams run Card Vault to tokenize at capture and charge through any gateway via our proxy, with optional Luhn-passable tokens checkout code accepts unchanged, so PCI scope collapses and processor choice stays; Data Vault for searchable encrypted account fields; File Vault for KYC documents. For platforms and vertical SaaS embedding payments, this is the standard path: offer payments to merchants while the vault carries custody, then extend the same integration to personal data.
Who runs this stack
Banks, fintechs, payment platforms, marketplaces, vertical SaaS, wealth managers, and fraud and support operations.
Run one week of copilot traffic through the free tier. The detection dashboard tells you what has been leaving the building.
Custody as a statutory duty, and AI arriving in casework.
Agencies hold tax records, justice data, license information, and citizen files across departments with different sensitivity rules, and AI-assisted casework and correspondence now move citizen data into prompts. Teams run Data Vault to encrypt records at the field level with per-department scopes and distributed tracing, File Vault for documents with expiring access, and NoPII in front of casework AI so citizen identifiers never reach a commercial model endpoint. It is worth noting the reference architecture argument: the strictest reviewers in government run their own internal LLMs inside hardened, non-training enclaves rather than sending raw data to commercial endpoints; tokenization delivers the same property without building the enclave.
Who runs this stack
Tax authorities, courts and justice agencies, licensing bodies, and the vendors serving them.
Education
NoPIIThe Vault
Student data at enrollment scale, under FERPA and state law.
Universities and school systems hold student records, financial aid data, transcripts, and staff HR files, with AI tools now reading student communications and records for advising, admissions triage, and support. Teams run Data Vault to batch-encrypt records through enrollment surges at up to 5,000 secrets per request, File Vault for transcripts and IDs shared across departments through ephemeral keys instead of email, and NoPII in front of advising and support AI so student identifiers stay out of provider logs and erasure requests resolve at the vault.
Who runs this stack
Registrars, financial aid offices, IT, and edtech vendors.
Customer support and SaaS
NoPIIThe Vault
Volume, streaming, and erasure rights that must resolve somewhere.
Support AI runs at conversation speed: the customer types name, order number, and address into chat, ticket history carries every prior interaction, and summarization, drafting, and routing each re-send the thread to the model in real time. Once customer data is distributed across provider logs, GDPR and CCPA erasure requests become promises you cannot keep. Teams run NoPII in the streaming path with SSE support so the chat experience is unchanged, deterministic tokens keeping multi-turn sessions coherent, and one-click token purge resolving right-to-erasure at the vault. SaaS platforms holding customer records back this with Data Vault, and those billing subscriptions with Card Vault.
Who runs this stack
Support operations, SaaS platforms, BPOs, and CX tooling vendors.
Protect your first million support tokens free. Two lines of code, live before the next shift starts.
Recruiting assistants read resumes with names, addresses, and visa status; feedback summarization reads performance reviews; people analytics reads compensation by employee, and it is exactly the data internal AI pilots reach for first, usually before legal has seen the tool. Teams run NoPII to tokenize employee identifiers, salaries, and personal details before people workflows touch a model, with the audit trail giving legal the control layer to approve the program, and Data Vault to hold HR records encrypted and searchable underneath.
Who runs this stack
People operations, HR tech vendors, and payroll platforms.
Get the people-AI pilot approved. Bring legal to a demo; the audit trail does the convincing.
Cards at checkout, customer data in personalization, and catalogs that must tell the truth.
Commerce runs on stored cards and customer profiles, with AI now writing product content and answering shoppers. Teams run Card Vault with hosted and iframe forms in 13 languages that embed in any checkout, including Magento and WooCommerce storefronts, and gateway proxy to any processor, so checkout stays and PCI scope goes; Data Vault for customer profiles; NoPII in front of personalization and service AI carrying order and profile data; and Triplets for catalog accuracy at the scale where manual review stopped working, certifying product claims against source data.
Who runs this stack
Retailers, marketplaces, e-commerce enablement platforms, and catalog teams.
Travel, hospitality, and property management
The Vault
Cards, IDs, and documents collected across locations and channels, controlled by no single system.
Hotels and travel agencies hold passports, itineraries, and traveler cards; property managers process applications with IDs, signed leases, and deposit payments; restaurants and event businesses keep corporate cards on file for recurring orders. Teams run Customer Vault for branded, no-login intake across locations with a shared staff inbox and expiring document sharing, and Card Vault for cards on file charged through the existing processor, with Twilio Pay capturing phone bookings straight from the IVR.
Who runs this stack
Hotel groups, travel agencies, property managers, and hospitality operators.
Aviation, cruise, clubs, and card-on-file operations
The Vault
The operating model is cards on file, collected over years, across channels.
Private aviation, cruise lines, member clubs, dental groups, and home services businesses run on stored cards living in spreadsheets, legacy systems, and terminals nobody fully trusts. One representative deployment in private aviation: existing cards imported into the vault in bulk, a branded card-management experience for account holders, staff access behind enterprise single sign-on, every charge routed through the operator's existing processor. The cards moved into certified custody; the operation did not change. This is the pattern Agent Vault productizes as a fixed-scope, eight-week engagement including migration.
Who runs this stack
Operators with card-on-file at the center of the business, and the platforms that serve them.
Ask about Agent Vault. Eight weeks from scattered custody to vaulted custody.
The most sensitive financial documents a person owns, collected mostly by email.
W-2s, bank statements, and financial records arrive as attachments every spring, and completed returns go back out the same way, with AI now assisting review and drafting in between. Teams run Customer Vault to collect client documents through encrypted, branded intake and deliver returns through protected expiring links, File Vault underneath for storage with a full access trail, and NoPII in front of document-review AI so client financial identifiers never reach provider logs.
Who runs this stack
Accounting and tax firms, advisory practices, and the platforms serving them.
Manufacturing, quality, and compliance-driven operations
Triplets
Specifications that supersede each other, and answers that must come from the current revision.
Manufacturers run on specs, revisions, and quality documentation that contradict each other across versions, and an AI answering from the superseded revision is an operational failure with a quality record. Compliance and risk functions in any industry face the mirror problem: standing obligations tracking regulation that changes without notice. Teams run Triplets as the reconciliation ledger: specifications and obligations compiled with states, conflicts surfaced with a governing verdict, monitored continuously with alerts only on support changes, and point-in-time reconstruction for the audit.
Who runs this stack
Quality and regulatory teams in manufacturing, and compliance and risk functions across industries.
EVERY OTHER INDUSTRY
Do not see your industry?
If you hold regulated data, send it to AI, or answer questions where being wrong is expensive, the trust layer fits. The sections above are where we go deep, not the limit of where we work. Tell us your industry and we will map the three exposures to your stack in one conversation.
