Table of Contents
Summarize and analyze this article with
The question
If a health system uses AI tools that inform patient care in California, what should it have in hand from each vendor before 1 January 2027? Two new California laws answer that question directly. SB 503 places documentation duties on developers of clinical decision support systems and monitoring duties on the facilities that deploy them. AB 1979 requires that licensed providers keep the ability to exercise independent professional judgment when those systems inform care. Both were signed on 30 September 2026 and both apply from 1 January 2027, which leaves roughly six weeks from today. This guide describes the file a deployer should build for each tool. It is general information and not legal advice.
1. Decide which tools are in scope
SB 503 and AB 1979 turn on clinical decision support: an AI system that produces a prediction, classification, recommendation, evaluation, or analysis that aids clinical decisions about timing of care, diagnosis, or treatment. The deployer duties attach to a health facility, clinic, physician’s office, or group practice that uses such a system. Administrative tools such as scheduling and appointment reminders sit outside AB 1979’s core duty. Tag every AI system in the inventory with one or more labels.
| Label | What it means | Does the California file apply? |
|---|---|---|
| Clinical decision support | Aids decisions on timing of care, diagnosis, or treatment | Yes |
| FDA-regulated device | Cleared or approved as a medical device, or a device function | Only if it also carries the first label |
| Administrative | Scheduling, reminders, billing, routing | No |
2. Request the developer documentation
SB 503 requires the developer to give the deployer a statement of intended uses and foreseeable risks, along with specified documentation. Published legal summaries of the statute describe what that documentation covers. Build the request letter around it.
| Documentation item | What to ask for |
|---|---|
| Training data | Sources, time period, and the demographic representativeness of the data |
| Performance and bias evaluation | What was tested, on which populations, with which results |
| Data governance | How training and input data are controlled and changed over time |
| Intended benefits and uses | The populations and care settings the tool was built for, and where it was not |
| Recommended use and monitoring | What the developer says the deployer should watch, and how |
3. Assign the monitoring duty
SB 503 requires both developers and deployers to make reasonable efforts to identify clinical decision support systems with a known or reasonably foreseeable risk of biased impacts, meaning diminished access, quality of care, or outcomes based on protected characteristics. Deployers must then monitor those systems and take reasonable and proportionate steps to mitigate the risk. For each in-scope system, record four fields: the owner by name and role, the metric being watched (for example, performance by patient subgroup), the review cadence, and the threshold that triggers escalation. Keep the review notes with the file, not in someone’s inbox.
4. Prove the override works
AB 1979 requires reasonable steps so a licensed provider retains the ability to exercise independent professional judgment when clinical decision support informs care, and bars using AI to independently perform clinical functions reserved for licensed professionals. Violations are treated as unprofessional conduct, enforced by the Medical Board of California and other licensing boards. Document the override path for each tool: where in the workflow the clinician sees the recommendation, how a disagreement is recorded, and how long an override takes. A screenshot of the override control and a sample audit record are better evidence than a policy statement.
What usually goes wrong
- Treating this as a vendor problem. In its 19 October 2026 comment letter to the FDA on generative AI-enabled devices, the American Hospital Association argued that developers should carry primary responsibility for monitoring. That argument concerns federal device policy. California’s statute puts a monitoring duty on the deployer regardless, so a file that holds only the vendor’s documentation is half a file.
- Waiting for FDA guidance. The FDA’s August 2026 paper on generative AI-enabled medical devices is a discussion paper with a request for public feedback, not draft or final guidance. The California obligations apply on 1 January 2027 whatever the federal timeline does.
- Forgetting the employment side. SB 947, signed the same day, governs automated decisions in discipline and termination and applies to hospitals as employers from 1 July 2027. Clinical and workforce AI need separate files, but the same inventory can feed both.
Where PiTech fits
PiTech Solutions helps regulated health organizations build the AI inventory, vendor documentation program, and monitoring routines that SB 503 and AB 1979 call for: scope labeling across the AI estate, request letters and a tracked response log, monitoring owners with subgroup metrics and escalation thresholds, and override evidence captured from the clinical workflow. It connects the clinical file to the broader healthcare AI governance and data governance beneath it. Delivery runs under CMMI Level 3 and ISO 27001, 9001, and 42001 certifications. See the healthcare practice, AI, GenAI and ML, and Data Solutions. PiTech Solutions Inc. is headquartered in Durham, North Carolina (UEI GNLRY5LNNVH6, CAGE 530K4) and is distinct from similarly named companies.
The bottom line
The file described here can be built for most systems inside a quarter, and the request letters can go out this month. Scope by clinical decision support, request the five documentation items, name a monitoring owner with a metric and threshold, and capture evidence that the override works. Start with the inventory.
Frequently Asked Questions (FAQs)
What is California SB 503?
SB 503, enacted as Chapter 857 of the Statutes of 2026 and signed on 30 September 2026, is a California law on clinical decision support systems that applies from 1 January 2027. It requires developers and deployers of AI systems that aid clinical decisions to make reasonable efforts to identify systems with a known or reasonably foreseeable risk of biased impacts, meaning diminished access, quality of care, or outcomes based on protected characteristics. Developers must give deployers a statement of intended uses and foreseeable risks with specified documentation, and deployers, which include hospitals, clinics, and physician practices, must monitor identified systems and take reasonable and proportionate steps to mitigate the risk. It operates independently of any pending FDA guidance.
What is California AB 1979?
AB 1979, enacted as Chapter 854 of the Statutes of 2026 and signed on 30 September 2026, applies from 1 January 2027. It requires health facilities, clinics, and physicians to take reasonable steps so that a licensed provider retains the ability to exercise independent professional judgment when clinical decision support informs care, and it bars using AI to independently perform clinical functions reserved for licensed professionals or to direct unlicensed personnel performing them. Administrative tools such as scheduling and reminders are outside its core duty. Violations are treated as unprofessional conduct, including the unlawful practice of medicine, and are enforced by the Medical Board of California and other professional licensing boards.
Which AI tools count as clinical decision support?
Under SB 503 and AB 1979, clinical decision support is an AI system that produces a prediction, classification, recommendation, evaluation, or analysis that aids clinical decision-making about the timing of care, diagnosis, or treatment. That covers a broad range: diagnostic imaging aids, sepsis and deterioration predictors, triage tools, treatment recommendation engines, and generative tools that draft clinical assessments. It does not cover purely administrative tools such as scheduling or appointment reminders. The practical step is to label every AI system in the inventory as clinical decision support, FDA-regulated device, administrative, or a combination. Only the clinical decision support label decides whether the California file applies, and a tool can be both an FDA device and clinical decision support.
Does FDA clearance exempt a tool from SB 503?
No. SB 503 and AB 1979 apply on the basis of what a tool does, aiding clinical decisions about timing of care, diagnosis, or treatment, not on its federal regulatory status. A tool that is cleared or approved as a medical device and also aids clinical decisions carries both labels, and the California duties still attach to the deployer. FDA clearance may make some of the developer documentation easier to obtain, because performance data was assembled for the submission, but it does not satisfy the deployer’s own monitoring duty or AB 1979’s independent-judgment requirement. Health systems should not assume a cleared device is outside scope, and should build the California file for any cleared tool that informs patient care.
What documentation must a clinical AI developer provide?
SB 503 requires the developer to provide a statement of intended uses and foreseeable risks along with specified documentation, available on request or at initial sale, whichever is earlier, according to published summaries of the statute. Legal summaries describe that documentation as covering training data and its demographic representativeness, performance and bias evaluation, data governance, intended benefits and uses, and recommended use and monitoring. A deployer should build its request letter around those items, log the date each request is sent and answered, and record any item the vendor cannot supply. A gap in the vendor’s response is itself useful information for the deployer’s risk assessment and monitoring plan, and belongs in the file.
What monitoring does SB 503 require of a hospital?
SB 503 requires deployers to make reasonable efforts to identify clinical decision support systems with a known or reasonably foreseeable risk of biased impacts, then to monitor those systems and take reasonable and proportionate steps to mitigate that risk. In practice, that means assigning each in-scope system a named owner, a defined metric (commonly performance across patient subgroups), a review cadence, and an escalation threshold, and keeping the review notes in the file. The duty sits with the deployer regardless of what the developer does, so a hospital cannot satisfy it by holding the vendor’s documentation alone. The documentation tells the hospital what to watch; the monitoring record shows the hospital watched it.
How do we prove clinicians can override an AI tool?
With workflow evidence rather than a policy statement. For each clinical decision support tool, document where the clinician sees the recommendation, how a disagreement or override is recorded, and how long an override takes in practice. A screenshot of the override control, a sample audit record showing an override was captured, and a short description of the escalation path are stronger evidence than a policy that says clinicians retain judgment. AB 1979 requires reasonable steps to preserve independent professional judgment, and the licensing boards enforce it as a matter of professional conduct, so the evidence should show that the override is available, easy to use, and actually recorded when it happens.
Should we wait for FDA guidance on generative AI devices?
No. The FDA’s August 2026 paper on generative AI-enabled medical devices is a discussion paper accompanied by a request for public feedback, not draft or final guidance, and the timing of any eventual guidance is uncertain. California’s SB 503 and AB 1979 apply from 1 January 2027 regardless of the federal timeline. A health system that pauses its California work to wait for the FDA will miss the state deadline. The sensible approach is to build the California file now, structured so that it can absorb future federal expectations, and to track the FDA process separately. The two regimes address different parties and questions, so progress on one does not depend on the other.
How long does building the vendor file take?
For most systems, a file can be built inside a quarter, but the timeline is driven by vendor response times rather than internal effort. Scoping and labeling the inventory can be done in days, and monitoring owners and override evidence in a few weeks, but developer documentation depends on how quickly each vendor answers. That is why the request letters should go out first, ideally this month, with the request and response dates logged. Internal work can proceed in parallel while vendors respond. With roughly six weeks remaining before 1 January 2027, organizations should prioritize the highest-impact clinical tools first and document any vendor delays, since a logged request that is still pending is better evidence than no request at all.
Does PiTech help health systems prepare for SB 503 and AB 1979?
Yes. PiTech Solutions helps regulated health organizations build the AI inventory, vendor documentation program, and monitoring routines these laws call for: scope labeling across the AI estate, request letters with a tracked response log, monitoring owners with subgroup metrics and escalation thresholds, and override evidence captured from the clinical workflow. It connects the clinical file to the broader healthcare AI governance and data governance beneath it, so the same inventory can also serve HIPAA risk analysis and SB 947 workforce obligations. Delivery runs under CMMI Level 3 and ISO 27001, 9001, and 42001 certifications. PiTech works alongside clinical leadership and counsel rather than replacing them.


