Before approving an AI finance tool, a medical practice owner needs more than a statement that data is “encrypted.” The useful question is whether that protection covers the financial information the practice will actually share—and what happens when the software, its staff or another service needs to read it.
Bank transactions, provider and administrative payroll, patient balances and entity-level financial reports have different uses and exposure risks. A tool may protect its main database while leaving unanswered questions about uploaded spreadsheets, AI prompts, support attachments or downloaded reports.
Your job is to evaluate the vendor’s representations, not choose its cryptographic implementation. Use the five questions below to request specific answers, identify gaps and decide when an independent technical reviewer is needed. Apply the same review to any prospective provider; this framework does not verify or endorse any vendor’s capabilities.
Start with the financial workflow you intend to approve
Define the proposed use before asking about encryption. “Analyze practice finances” is too broad. “Prepare a location-level expense comparison from general-ledger totals” gives the vendor and your reviewer a clearer scope.
List what will enter the system, where it comes from and what output staff need. Avoid sharing patient-level detail when aggregated financial information can answer the question.
| Proposed finance workflow | Information potentially involved | Scope question before upload |
|---|---|---|
| Cash planning | Payer receipt totals, bank balances and expected payroll | Can monthly or weekly totals support the forecast? |
| Payroll expense review | Provider and administrative compensation | Are employee-level details necessary, or will department totals work? |
| Patient-balance analysis | Balances, aging categories and potentially identifying details | Can the analysis use summarized aging data without patient identifiers? |
| Multi-location reporting | Location revenue, expenses and shared-cost allocations | Which entities and locations should each user be able to access? |
| Bookkeeping support | Statements, invoices and reconciliation documents | Could attachments contain account credentials or unnecessary personal information? |
The HHS minimum necessary guidance explains a standard with exceptions. As a practical review step, minimize patient information in financial analysis and have the appropriate adviser assess applicable obligations. Do not put patient-identifiable information into ordinary Slack messages or a general-purpose AI tool.
sbb-itb-17e8ec9
Encryption terminology: different jobs, not a ranking
A vendor may list several algorithms without explaining their roles. These terms are not interchangeable:
| Technology | What it does | What the owner should understand |
|---|---|---|
| Symmetric encryption | Uses the same secret key to encrypt and decrypt information | Commonly protects stored records and data carried within secure connections. AES-GCM and ChaCha20-Poly1305 are established authenticated-encryption examples. |
| Public-key cryptography | Uses related public and private keys for tasks such as signatures and establishing or protecting shared keys | RSA and elliptic-curve methods serve different roles from bulk-record encryption. Their names alone do not establish safe implementation. |
| Cryptographic hashing | Produces a fixed-length digest from an input; it is not reversible encryption | A hash can help detect changes when checked against a trusted reference. It does not encrypt a payroll file or make patient information anonymous. |
Authenticated encryption combines confidentiality with a way to detect tampering. NIST describes AES in its AES standard and GCM in its authenticated-encryption guidance. These are educational examples, not a recommendation that owners prescribe particular algorithms or key sizes.
Also distinguish encryption at rest, encryption in transit and data in use. A system can protect stored files and network traffic while still making readable information available during AI processing. Specialized techniques such as homomorphic encryption support certain computations on encrypted inputs, but that capability should never be assumed from an “AI-powered” label.
1. What protects every transfer of our financial data?
Ask the vendor to describe the complete path, not just the connection between your browser and its website.
A practice might upload a payroll summary, have it processed by an AI service and receive a variance report. Each transfer needs an explanation: upload, internal processing, third-party processing and delivery of results.
Request answers to these questions:
- Which protocols protect browser, application and service-to-service connections?
- How are the identities of receiving systems verified?
- Does the same protection cover file uploads, scheduled transfers and support attachments?
- Which outside services receive information, and how are those transfers protected?
- Are any transfer paths excluded from the vendor’s encryption statement?
TLS is a protocol used to protect communications; it is not a claim that every subsequent copy remains encrypted. The TLS 1.3 specification describes the protocol, but a vendor still needs to explain where and how it is deployed.
Decision point: If the vendor describes only its public website, request a data-flow diagram covering the proposed finance workflow before approving sensitive uploads.
2. What is encrypted at rest—and what becomes readable during AI processing?
“Encrypted database” may describe only one part of the service. Ask whether the representation also covers uploaded statements, document storage, temporary files, AI prompts, outputs, logs and support systems.
For a practice reviewing payer receipts and patient-balance trends, the distinction matters. A stored source file might be encrypted while a diagnostic log or saved prompt contains readable extracts.
Ask the vendor to identify:
- The stored data categories covered by encryption and any exceptions.
- Where information is decrypted for processing.
- Which employees, administrators and service providers can access readable information.
- Whether inputs or outputs are retained or used for model training or other purposes.
- How those uses are limited by settings and contractual terms.
Encryption at rest does not prevent an authorized application from decrypting records. Nor does it determine whether a staff member should see compensation across every location. Access permissions remain a separate control.
Decision point: Approve a defined use, not unrestricted access to every financial record. If summarized location-level data is sufficient, keep the initial scope there while unresolved questions are reviewed.
3. Who controls the encryption keys, and what happens if that control fails?
Encryption depends on protecting the keys that permit decryption. A familiar algorithm cannot compensate for broadly accessible keys or an untested recovery process.
Owners do not need secret keys or detailed implementation instructions. They need a documented explanation of responsibility and access:
- Does the vendor, a cloud provider or the customer administer the keys?
- How is key access restricted separately from ordinary data access?
- Can support staff or administrators obtain readable practice data, and under what authorization?
- How are key changes, suspected compromise and recovery handled?
- How does the vendor preserve access to legitimate historical records when keys change?
A claim such as “customer-managed keys” needs context. It may describe administrative control without establishing that the application cannot decrypt data. Ask your technical reviewer to explain the actual boundary.
Recovery also has a finance consequence. Losing access to historical payroll support or reconciliation records can disrupt the practice even if no information was disclosed.
Decision point: Require clear responsibility for key administration and recovery. Escalate unclear answers to an independent technical reviewer rather than selecting the implementation yourself.
4. Does protection extend to backups, exports and departure from the service?
Follow a report beyond the vendor’s main application. An exported payroll spreadsheet, emailed attachment or locally saved patient-balance report may no longer have the protections described for hosted storage.
| Copy or destination | Question to ask | Practice-side action |
|---|---|---|
| Backups and recovery copies | Are these encrypted, who can restore them, and how is recovery tested? | Confirm how records can be recovered when needed. |
| Spreadsheet or PDF exports | Does protection persist after download, or depend on the destination? | Use approved storage and limit recipients. |
| Email or support attachments | What transfer and storage protections apply? | Avoid attaching unnecessary sensitive details. |
| AI processing copies | Are prompts, outputs and temporary copies retained by other providers? | Confirm the scope before sharing data. |
| Account closure | What can be exported, what is deleted, and how are backup copies handled? | Obtain usable records before ending access. |
The IRS business-recordkeeping guidance supports keeping complete business records and reconciling bank records. Vendor deletion settings should therefore be coordinated with your CPA and applicable recordkeeping requirements—not treated as a universal destruction schedule.
Decision point: Do not approve a service until the practice understands how it will retrieve usable financial records and protect exported copies.
5. What independent evidence supports the encryption claims?
A marketing page, badge or cloud-provider logo is not enough to establish how the specific service operates.
Ask for evidence relevant to the product and workflow under review, such as a current independent controls report, technical assessment, penetration-test summary or cryptographic validation where claimed. These materials answer different questions; none should be treated as a blanket guarantee.
Have a qualified reviewer examine:
- Scope: Does the evidence cover this product, its AI processing path and relevant service providers?
- Timing: What period or system version was assessed, and what has changed since?
- Findings: Were relevant weaknesses identified, and what supports remediation?
- Limitations: Which controls, systems or customer responsibilities were excluded?
Use vendor-authorized testing arrangements. Practice staff should not probe a provider’s systems without permission.
Decision point: If evidence is unavailable or excludes an important processing step, document the gap. Choose whether to defer approval, narrow the data scope or obtain further review rather than silently accepting the uncertainty.
Copyable vendor-questionnaire template
Send this with a short description of the intended workflow. Ask for written responses and supporting documents, not patient or employee records.
| Review area | Vendor response requested | Evidence or clarification requested |
|---|---|---|
| Proposed scope | Identify data accepted, outputs produced and all processing providers. | Data-flow diagram and service-provider list. |
| 1. Transfers | Explain protection for uploads, internal connections, external processing and report delivery. | Protocol documentation and exceptions. |
| 2. Storage and processing | Identify encrypted stores, plaintext processing locations, retained prompts and outputs, and permitted secondary uses. | Architecture summary and applicable terms. |
| 3. Keys | Identify key administrators, authorized access, key-change procedures and recovery responsibilities. | Key-management control summary. |
| 4. Copies and exit | Explain backup protection, export handling, recovery, deletion and account closure. | Recovery evidence and exit procedure. |
| 5. Verification | Identify independent assessments and their scope, date and limitations. | Relevant reports or review summaries, subject to confidentiality terms. |
Record the reviewer, decision date, approved data scope, unresolved gaps and who can authorize any expansion. This creates a useful purchasing record without requiring the owner to become a cryptographer.
Turn a vague claim into actionable follow-up questions
Illustrative vendor claim: “Our AI platform uses bank-grade AES-256 encryption to keep all your data secure.”
Do not interpret that sentence as proof of end-to-end protection. Ask:
- Does this describe storage, transfers or both?
- Which files, databases, logs, backups and AI processing providers are covered?
- Where can information become readable, and who can access it?
- Who controls the keys?
- What protection remains after a report is downloaded?
- What independent evidence supports this statement for the service we would purchase?
A useful response identifies boundaries and exceptions. Repeating the algorithm name does not resolve them.
Make a scoped approval decision
Conclude the review with one of three outcomes: approve the defined workflow, approve only a narrower scope, or defer pending evidence. Assign someone to revisit the decision when processing providers, data uses or export arrangements change.
Encryption is an important control, but it does not establish HIPAA compliance. HHS cloud guidance explains that a cloud provider maintaining ePHI can be a business associate even without the decryption key. Appropriate agreements and applicable safeguards still matter; HHS does not certify cloud products.
The practical goal is defensible approval of a finance workflow: useful analysis, appropriately limited information and evidence that matches the vendor’s claims.
FAQs
Should a medical practice choose the vendor with the largest encryption key size?
No. Key sizes are not directly comparable across different cryptographic methods. Evaluate coverage, implementation, key management and supporting evidence with a qualified reviewer rather than ranking vendors by a number in a brochure.
Does hashing make patient or payroll information safe to upload?
Not by itself. Hashing is not encryption, and a hashed identifier is not automatically anonymous. Predictable inputs can sometimes be guessed and compared. Avoid assuming that replacing names with hashes makes a dataset suitable for unrestricted sharing.
Can we start evaluating a vendor before sensitive-data approval is complete?
Yes. Use synthetic demonstration data or carefully reviewed, non-identifying financial summaries. Do not connect live accounts or upload patient-identifiable information merely to test a feature. Expand the scope only after the relevant technical, contractual and privacy review is complete.