Retention Policies for Clinical Dental AI: How Long an Agent Should Keep Conversation and Inference Logs

You probably think of log retention as a storage-cost question — a knob you turn down when the S3 bill gets loud. However, for a clinical dental AI agent, the retention window is a legal position, and most practices adopt one by accident when they accept a vendor default they never read.

The contract review usually goes well. Someone checks that the BAA is executed, someone else confirms the vendor writes an audit trail, and the deal closes.

What almost never appears in that conversation is a number. How many days does the vendor hold the conversation transcript, the model input, the inference output, and the tool-call arguments that carried a patient's chart data into a prompt?

Why The Retention Gap Exists

Three separate rulebooks govern the data your agent produces, and they were written by people who were not talking to each other. None of them contemplated an LLM transcript.

HIPAA's Security Rule sets a six-year retention requirement — but for documentation of policies, procedures, and required actions, not for PHI itself. There is no federal floor telling you how long to keep the actual clinical record.

That floor comes from state law, and it varies widely. Dental record retention statutes commonly run six to ten years after the last treatment date, with longer tolling for minors — often until the patient reaches the age of majority plus the statute of limitations.

Vendor defaults are the third rulebook, and they are the one actually running in production. A general-purpose LLM platform often defaults to 30 days of abuse-monitoring retention; a SaaS wrapper may default to indefinite retention because deleting data was never a product requirement.

The Four Artifacts Your Agent Actually Produces

Treating "logs" as one bucket is where the analysis breaks down. A production agent emits at least four distinct data classes, and each deserves a different window.

Separating them is the entire exercise. Once they are separated, the retention decision for each one becomes close to obvious.

• Conversation logs. The raw turn-by-turn transcript between a staff member or patient and the agent. This is the highest-PHI-density artifact and almost always the one with the loosest vendor default.
• Inference logs. The model input and output payload — the assembled prompt including retrieved chart context, plus the raw completion. If you run retrieval-augmented generation over a patient chart, the prompt itself contains PHI even when the user's message did not.
• Tool-call traces. The arguments and responses for every function the agent invoked: the patient ID passed to a chart lookup, the CDT code returned by a coding tool, the payload sent to an eligibility check. These sit in your own CloudWatch or OpenTelemetry pipeline as often as in the vendor's.
• Clinical outputs of record. The note, the coded claim line, the treatment-plan draft that a clinician reviewed and accepted. Once a human signs it, it stops being a log and becomes part of the dental record — governed by state statute, not by your retention policy.

That last distinction is the one practices get wrong most often. An AI-drafted clinical note that a dentist accepted is not a log you can age out at 90 days; it is a record entry with a six-to-ten-year statutory life, and it needs to live in the practice management system rather than in the vendor's log store. Work through how that handoff should be structured in clinical note summarization before you write the policy.

What A Defensible Retention Schedule Looks Like

Here is a schedule that holds up in a dental board inquiry and survives a security questionnaire, with the reasoning attached to each line. Treat the windows as a starting position to negotiate from, not as legal advice for your jurisdiction.

Notice what the table does: it pushes the long-retention obligations onto systems you already control and already inventory, and it caps the short-retention obligations inside the vendor. The vendor holds the least data for the shortest time, which is exactly the risk posture you want when the breach notification letter has to name a third party.

The Audit Trail Is Not The Conversation Log

Practices conflate these two constantly, usually because the vendor's dashboard shows them on the same screen. They serve opposite purposes and therefore get opposite windows.

The audit trail answers who touched what PHI, when, and from where. It is a structured event stream — user ID, patient ID, action, timestamp, source IP — and it needs to survive six years in tamper-evident storage because that is what you hand a regulator.

The conversation log answers what was said. It is unstructured, high-volume, and full of incidental PHI that no one catalogued, which makes it a liability to keep and a cheap thing to lose.

Keep the first one long and immutable — S3 Object Lock in compliance mode, or an equivalent write-once store with KMS encryption and CloudTrail coverage on the bucket itself. Keep the second one short and deletable, with a lifecycle policy that runs whether or not anyone remembers it exists.

Why Zero-Retention Is Harder Than It Sounds

Zero-retention inference — the vendor processes the prompt and persists nothing — is the cleanest compliance story available. On Amazon Bedrock, for instance, model inputs and outputs are not retained by the service or shared with model providers by default, which is a large part of why clinical teams land there; the architecture case is covered in more depth in our writeup on running clinical AI on Bedrock.

The complication is that zero-retention removes your ability to debug. When a hygienist reports that the agent produced a wrong tooth number last Thursday, and nothing was logged, your incident review consists of shrugging.

The resolution most operators converge on is a two-tier scheme. Run zero-retention as the default path, and write a de-identified or tokenized copy of the payload into your own store when a specific condition fires — low model confidence, a clinician rejection, a validation-suite failure, or an explicit staff flag.

That gives you an eval corpus without giving you a PHI lake. It also means your regression set is yours: when the vendor pins a new model version, you can replay your own captured cases rather than trusting a release note, which is the core discipline described in building clinical AI evaluation suites.

What To Put In The Contract

A BAA establishes that the vendor may handle PHI and must safeguard it. It does not, by itself, tell you what happens to a transcript on day 91.

The retention terms belong in the service agreement or a data-processing addendum, and they should be specific enough that an engineer could implement them without asking a lawyer what was meant. Here is what to require, in order of how often it is missing:

• A named window per artifact class. Not "logs are retained for a commercially reasonable period." A number of days for conversation logs, a number for inference payloads, a number for the audit trail.
• Deletion that is actually deletion. Confirm whether "delete" means purged from primary storage, backups, and any derived analytics store — and get the backup expiry window in writing, because a 35-day backup cycle silently extends a 30-day retention policy to 65 days.
• No training on your data, stated affirmatively. Silence is not a prohibition. The clause should say the vendor will not use customer content to train, fine-tune, or evaluate models, and should bind subprocessors to the same term.
• A subprocessor list with their retention windows. Your vendor's 30-day policy is meaningless if their observability provider holds the same payloads for a year. Ask for the list and the windows.
• Deletion certification and an export path. On termination, you need your audit trail exported in a usable format and written confirmation that everything else was destroyed, with a deadline attached — 30 days is typical.
• Breach notification timing tied to discovery. HIPAA gives covered entities 60 days from discovery to notify; a vendor who takes 45 of those days to tell you has consumed your timeline. Push for notification within 72 hours of discovery.

All of these are negotiable with a vendor that has enterprise customers, and most of them are non-starters with a vendor that does not. The negotiation itself is a useful diagnostic — a vendor who cannot tell you their backup expiry window has not thought about deletion at all.

Where State Dental-Record Law Changes The Answer

The federal picture is stable. The state picture is where a national DSO discovers that one retention schedule does not work across a footprint.

Dental record retention requirements are set by state dental practice acts and health department regulations, and they differ on three axes: the base window measured from last treatment, the tolling rule for minor patients, and whether radiographs and diagnostic images carry a separate, often longer, requirement. A group operating in several states inherits the longest applicable window as its practical floor unless it can enforce per-location policies.

The minor-patient tolling rule is the one that surprises people. In many states the clock does not start until the patient turns 18, which means a chart created for a seven-year-old may need to survive well past a decade — and any AI-generated note accepted into that chart inherits the same obligation.

This is also why the log-versus-record distinction carries real money. If your agent's outputs are treated as records, you have signed up for a decade of retention on a data class you were planning to age out quarterly. Your state board may also have an opinion about AI-generated documentation independent of retention — see how state dental boards are treating AI for where those rules currently stand.

Implementation: Making The Policy Real

A retention policy that lives in a Word document is a retention policy that gets violated within a quarter. The windows have to be enforced by infrastructure that runs without human involvement.

On the AWS side, that means S3 lifecycle rules that transition and expire objects on schedule, CloudWatch log group retention set explicitly rather than left at "Never Expire" — the default that quietly accumulates years of tool-call traces containing patient identifiers — and DynamoDB TTL attributes on any conversation-state table. Each of these is a one-line configuration and each is routinely missed.

Keep in mind that the enforcement point has to sit upstream of your own convenience. Engineers turn off log expiry during an incident and forget to turn it back on, which is why the retention setting belongs in infrastructure-as-code with a policy check in CI rather than in a console someone can click.

Be aware that de-identification is not a checkbox either. Stripping the eighteen HIPAA identifiers from a free-text clinical transcript is genuinely hard, and a naive regex pass will miss the name a patient typed into a chat message about their spouse; the mechanics of handling clinical free text are covered in our piece on dental NLP.

Questions To Ask Before You Sign

The questions below are the ones that produce the most information per minute in a vendor call. They are deliberately specific, because a general question about "your security posture" produces a marketing answer.

• What is your retention window for conversation transcripts, in days? If the answer is a paragraph rather than a number, that is the finding.
• Are model inputs and outputs persisted at all, and where? Ask specifically about abuse-monitoring retention, which is often a separate pipeline with a separate window.
• What is your backup retention, and does deletion propagate to backups? This is the question that most often changes the real answer to question one.
• Which subprocessors see PHI, and what are their windows? Observability, error tracking, and analytics vendors are the usual gap.
• Can you produce a deletion certificate, and on what timeline? If they have never produced one, you will be their first.

Run these alongside the broader evaluation criteria in our guide to selecting dental AI software. Retention is one line in a larger diligence process, but it is the line most likely to be blank.

Frequently Asked Questions

Setting Your Window

The practices that handle this well do not have better lawyers. They separated four artifact classes, assigned a number to each, and enforced those numbers in infrastructure rather than in a policy binder.

If you are scoping a clinical AI deployment and want a second set of eyes on the retention architecture before the contract is signed, the NexV team builds and operates HIPAA-grade clinical agent systems across practice management, imaging, and claims environments every week. Reach out for a working session — we will map your agent's data flows, classify every artifact it emits, and leave you with a retention schedule your compliance officer can defend and your engineers can deploy.