---
title: What FDA's FY 2027 Guidance List Means for AI Device Teams
description: FDA FY2027 PCCP and AI lifecycle guide: what a Predetermined Change Control Plan must include and the engineering patterns to operationalize it.
canonical: https://www.baytechconsulting.com/blog/what-fdas-fy-2027-guidance-list-means-for-ai-device-teams
---

The article explains the FDA CDRH FY2027 priorities for AI-enabled devices, what a Predetermined Change Control Plan (PCCP) must contain, and the engineering patterns—versioned data pipelines, isolated test environments (shadow mode) and runtime telemetry with rollback—that operationalize PCCPs so manufacturers can update models after clearance.

A model that retrains on new data can behave differently every quarter. The FDA already lets you plan for much of that change before you ship, and its FY 2027 guidance list shows what it wants to settle next.

[CDRH published the list](https://www.fda.gov/medical-devices/guidance-documents-medical-devices-and-radiation-emitting-products/cdrh-proposed-guidances-fiscal-year-2027-fy-2027) on October 1, 2026, and it covers October 1, 2026 through September 30, 2027. The rules for planned AI updates are not waiting on it. The FDA's guidance on Predetermined Change Control Plans (PCCPs) for AI-enabled devices [has been final since December 2024](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence), and the agency reissued it in August 2025. The list adds a final version of the broader AI lifecycle guidance and a PCCP guidance that extends the idea to all device types.

A PCCP only helps if your systems can keep its promises, which takes versioned data pipelines, isolated test environments and telemetry with rollback.

![Overview of FDA's FY 2027 AI guidance: key figures, what the FDA expects and what teams must build.](https://dzrge5zzbsh6q.cloudfront.net/pccp-guidance-overview.webp)

The guidance asks for a change plan, and teams need pipelines, test environments and telemetry to keep its promises.

## What CDRH Put on the FY 2027 List

The agency sorts its planned documents into an A-list of priority items it intends to publish this fiscal year, a B-list for items it will get to as resources allow, and an "Under Construction" list. According to [Complizen's reading of the October 2026 release](https://www.complizen.ai/post/cdrh-fy2027-guidance-list-what-fda-plans-for-ai-and-510k), the list holds 17 topics in total, and 8 of them sit on the A-list.

Three A-list items matter to anyone shipping AI software:

-   **AI lifecycle guidance.** The final version of "Marketing Submission and Lifecycle Management Recommendations for Artificial Intelligence-Enabled Device Software Functions," which started as a January 2025 draft.
-   **PCCP guidance.** The final version of "Predetermined Change Control Plans for Medical Devices," which started as an August 2024 draft. It extends PCCPs to all device types. AI-enabled devices already have their own final guidance.
-   **Generative AI for mental health.** A new draft on "Generative Artificial Intelligence-Enabled Conversational Devices for Mental Disorders: Evidentiary Recommendations."

The third item shows the agency is already thinking about large language models in patient-facing roles. The first is the one most AI teams should watch, because it covers how you document and manage a model across its life.

The lifecycle guidance and the all-device PCCP guidance are still drafts today, so their final text could change, and a place on the list is no promise of a date. CDRH published 7 of the 11 guidances on its FY 2026 list. The three-part PCCP structure is already settled for AI devices, in the FDA's final guidance and in guidance N90 from the International Medical Device Regulators Forum (IMDRF), so you can build to it now.

Industry has pushed in the same direction. AdvaMed, which represents medtech manufacturers, [commented on October 27, 2026](https://www.advamed.org/industry-updates/news/advamed-encourages-risk-based-fda-approach-to-generative-ai-medical-devices/) on the FDA's discussion paper on generative AI. It encouraged the agency to apply the PCCP framework to generative AI instead of creating separate, more restrictive frameworks, and it said post-market monitoring should stay risk-based.

### The Scale of What Is Being Regulated

As of the September 2026 update, the FDA has authorized 1,614 AI-enabled medical devices. A peer-reviewed cross-sectional study in JAMA Health Forum counted 950 AI/ML devices cleared by August 2024, with 108 clearances in 2023 and 107 in 2024. The pace hit a record 335 authorizations in 2025 alone.

**AI device authorizations per year hit a record in 2025.** AI/ML device clearances and authorizations counted per year. After two flat years, authorizations reached a record in 2025. Source: JAMA Health Forum (2023 and 2024 counts).

Roughly 97 percent of these went through the 510(k) pathway, which rests on substantial equivalence to a predicate, instead of premarket approval (PMA). Radiology accounts for about 76 percent of the authorized devices. That 510(k) share matters later, when we get to the predicate issue.

## Why a Locked Clearance Breaks for Models That Update

Under the old rules, a software change that could significantly affect safety or effectiveness needed a new premarket notification under 21 CFR 807.81(a)(3) for 510(k) devices, or a PMA supplement under 21 CFR 814.39(a) for Class III devices. That works when software changes once a year. It works badly for a model that needs periodic retraining.

Medical AI models in clinical use need that retraining because of sensor hardware changes, firmware updates, shifting clinical definitions and demographic drift, according to a paper in the International Journal of Clinical and Translational Medicine. A model that clears review on last year's population can be sub-optimal on this year's.

Congress addressed this in Section 515C of the FD&C Act, enacted in December 2022 as part of the Food and Drug Omnibus Reform Act (FDORA). It gives the FDA explicit authority to clear or approve a PCCP submitted through the 510(k), De Novo or PMA pathways. The FDA Law Blog described the effect as ending the "lock then submit" bottleneck, because the agency can evaluate in advance the process a manufacturer will use to change its software.

## What a PCCP Contains

A PCCP is documentation filed with a marketing application. It lists the changes you expect to make to the model after clearance, the methods you will use to validate them, and an assessment of how they affect clinical workflows. Once authorized, you can make the described changes without a new 510(k) or PMA supplement, as long as you follow the plan exactly and meet its pre-defined acceptance criteria.

The FDA expects a specific, standalone, versioned section of the submission. ContinuousIntelligence.ai warns that vague claims such as "improves over time" or "adapts to new data" will be rejected. The FDA's final guidance for AI-enabled devices recommends three parts.

### Description of Modifications

This is an itemized, bounded list of what will change in the AI-enabled device software function and how each change should affect performance. For each item you state the rationale and whether the change is applied manually or automatically. You also state whether it goes to every deployed instance or only to certain clinical sites.

Typical items include algorithm parameter adjustments, neural network threshold changes, training data updates on new patient cohorts, and user interface changes that affect AI functionality. Some changes cannot go in a PCCP. A change to intended use, a shift in the indicated patient population, or a higher risk classification needs a new premarket submission.

Change

Fits in a PCCP

Needs a new submission

**Retraining**

Updating weights with new data from the approved intended population

Training on a pediatric cohort when the device is cleared only for adults

**Performance tuning**

Adjusting thresholds to raise specificity above 95% within the cleared indications

Extending the algorithm to detect new diseases or pathologies

**Compatibility**

Support for a new operating system or imaging hardware

Turning a decision support tool into an autonomous diagnostic engine

### Modification Protocol

The Modification Protocol (MP) is the engineering rulebook for every planned change. It describes the verification and validation work that has to finish before a pre-authorized change reaches clinical production. The FDA's guidance breaks it into four parts, and each one maps to something you will have to build.

1.  **Data management.** Collection, validation, bias evaluation, dataset versioning and PHI de-identification. It also covers data sequestration, which keeps training data permanently separate from the independent test data used to evaluate the model.
2.  **Retraining practices.** The ML pipelines, statistical methods and optimization techniques used in an update, such as hyperparameter tuning and weight initialization.
3.  **Performance evaluation.** The exact acceptance criteria a modified model must meet, including non-inferiority margins, confidence intervals, subgroup analysis limits and the reference standard, for example consensus from U.S. board-certified radiologists.
4.  **Update procedures.** Release mechanisms, CI/CD pipelines, rollback triggers for degraded performance, cybersecurity validation, and how labeling is updated so clinicians know which model version is active.

### Impact Assessment

The Impact Assessment is the risk-benefit analysis, tied to ISO 14971 risk management. You compare the baseline device against the version with each modification applied alone, and against the version with all modifications applied together.

It has to cover the harms that updates can bring. Those include unintended algorithmic bias, overfitting to site-specific EHR data, and the effect of false positives and negatives on clinical workflows. The assessment then shows how the controls in the Modification Protocol reduce those risks, so the device keeps a reasonable assurance of safety and effectiveness.

## From Document to Runtime

The three-part document is the easy half. The hard half is that every sentence in the Modification Protocol becomes a promise your systems must keep on every release. The FDA is moving compliance from a one-time filing to something that runs in production.

[Legacy monoliths](https://www.baytechconsulting.com/blog/a-5-step-legacy-code-audit-that-keeps-the-business-running) and manual QA struggle here. The patterns below match the MP's pillars one to one. Data management maps to versioned pipelines. Performance evaluation maps to isolated test environments. Update procedures map to telemetry, drift detection and rollback. All three rest on the same [DevOps efficiency](https://www.baytechconsulting.com/services/devops-efficiency) habits as any well-run release process.

Area

Traditional software

PCCP-ready AI

**Data**

Ad-hoc queries and manual cleaning

Versioned pipelines, immutable artifacts, lineage tracking

**Release testing**

Staging server with static test data

Shadow-mode runs on live clinical data

**Monitoring**

Help desk complaint log

Runtime telemetry, drift detection, automated rollback

## Pattern One: Versioned Data Pipelines With Lineage

ContinuousIntelligence.ai says the FDA treats data provenance and governance as primary regulatory requirements. TMA Solutions adds that generic ML prototypes often fail the move into clinical production because they were built without versioned data pipelines and handle data wrangling ad hoc.

The fix is to treat every dataset and every feature as a versioned, testable, immutable artifact. In practice the pipeline reads from hospital systems over HL7 and FHIR, runs deterministic PHI de-identification, and lands the data in a compliant cloud environment.

### What the Lineage Record Must Answer

Add a model registry on top, with lineage tracking. If an auditor or a clinician questions a specific AI-influenced decision, you need to show which version of the raw data trained the model version that produced that prediction. Without that chain, every retraining cycle becomes manual and error-prone, and it cannot satisfy the data management expectations in the Modification Protocol.

A few design choices follow from this.

-   **Link weights to data.** Tie the production model weights to the exact training datasets and validation results with cryptographic links.
-   **Separate the test set at the storage layer.** The MP calls for data sequestration, so the independent test data should be unreachable from the training job. Permission boundaries are better than team discipline here.
-   **Record de-identification as a pipeline step.** Because the MP covers de-identification of PHI, the step should be versioned code with logged output, not a one-off script.
-   **Keep bias evaluation in the same pipeline.** The MP includes bias evaluation as part of data management, so it should run, and be recorded, with each dataset version.

## Pattern Two: Isolated Test Environments and Shadow Mode

A modified model has to prove itself under real conditions before clinicians see it. Healthcare teams commonly do this with shadow mode. The candidate model runs in parallel with production, takes in live EHR data and computes predictions, but its outputs go to a separate database and stay hidden from the clinical interface.

![Four-step flow showing a candidate model running in shadow mode on live data without touching patient records.](https://dzrge5zzbsh6q.cloudfront.net/shadow-mode-steps.webp)

The candidate is checked against the protocol on real hospital data while clinicians never see its outputs.

This lets you compare the candidate against the baseline model and against the statistical thresholds in the PCCP's performance evaluation protocol, using real hospital data instead of clean pre-market datasets.

### How Strict the Isolation Has to Be

Hadrian, a security vendor, gives a warning from outside healthcare. [It reports that in July 2026](https://hadrian.io/blog/what-ai-penetration-testing-needs-to-run-safely-at-speed) OpenAI disclosed that two of its models escaped an isolated test environment and breached Hugging Face's production systems while working on a benchmark task. Anthropic then reviewed more than 140,000 evaluation runs and found cases where its models reached the open internet and compromised real organizations. Hadrian's conclusion is that limits must be enforced outside the model, because prompts alone cannot guarantee safety.

For medical AI, that points to hard network guardrails, pre-provisioned authentication and strict rate limits, so a model under test cannot touch live hospital networks or exfiltrate PHI. Containers such as Docker and API sandbox environments give you lightweight, platform-agnostic isolation.

A workable shadow setup looks like this:

1.  Mirror live inputs to the candidate model through a one-way path.
2.  Write candidate outputs to a separate store that the clinical interface cannot query.
3.  Block the candidate from writing to patient records at the network and permission level, not just in application code.
4.  Run the candidate against the acceptance criteria in the MP and keep the results with the release record.

### The ARPA-H ADVOCATE Example

The ARPA-H ADVOCATE program is a leading worked example. It is building autonomous clinical AI agents for heart failure management. Kaiser Permanente and Johns Hopkins Applied Physics Laboratory teams are embedding agents into Epic workflows entirely in shadow mode. The program produces real-world evidence and deployment blueprints the FDA can use before agents get permission to change care plans.

The stakes are concrete. A single wrong medication adjustment could cause kidney injury or arrhythmia. Stanford is designing a disease-agnostic supervisory agent to watch the patient-facing AI after deployment. It uses a three-stage pipeline with escalating compute (outlier filtering, rule-based screening and deep-research auditing) to catch out-of-distribution behavior and write inspectable rationales for each action. The [FDA is involved in the program](https://www.cosmhq.com/resources-posts/arpa-h-advocate-fda-authorized-autonomous-clinical-ai), and the shadow-mode deployments span Kaiser Permanente's 21 medical centers. They are meant to show how PCCPs and post-market surveillance have to evolve for autonomous clinical AI.

## Pattern Three: Runtime Telemetry, Drift Detection and Rollback

Post-market performance monitoring is described as an explicit expectation in the PCCP guidance, according to Quality Smart Solutions and IntuitionLabs. Greenlight Guru is blunt about what does not count. A standard hospital IT complaint log does not meet the FDA's expectation for performance monitoring, which is active, real-time drift detection.

For the update-procedures pillar, [a complaint log will not pass](https://www.greenlight.guru/blog/pccp-fda-requirements-samd-ai) as a monitoring system, so telemetry has to be part of the product. In practice that means sidecar logging that captures model latency, processing throughput, shifts in input data distributions and output confidence scores. Those streams feed observability dashboards that track precision, recall and calibration across patient demographic slices.

### Safety Invariants and Automatic Rollback

Telemetry only helps if something acts on it, so check it in real time against fixed safety limits. For example, if an updated algorithm's specificity falls by more than a set threshold, such as 5 percent, the system should revert automatically to the last validated version, or escalate to a human clinician.

![Checklist of three rollback requirements with a 5 percent specificity drop as the example trigger.](https://dzrge5zzbsh6q.cloudfront.net/rollback-safety-invariants.webp)

The threshold is an illustration; your own number belongs in your Modification Protocol.

Your rollback trigger has to match the number in your Modification Protocol. Three parts make this work in a real system.

-   **One source of truth for thresholds.** The same acceptance criteria that gate a release in the pipeline should drive alerts in production. Two sets of numbers will drift apart.
-   **A rollback you have tested.** The previous validated version must stay deployable, with its data lineage intact, so reverting is a release of a known artifact.
-   **Labeling tied to the version.** The MP requires labeling updates so clinicians know which model version is active, so the active version should be exposed to the interface from the registry.

Telemetry also feeds your paperwork. The periodic performance reports for regulators draw on this data, and it shows the model is operating inside its authorized parameters.

## What a Cleared PCCP Looks Like in Practice

### a2z Radiology AI: A Locked PCCP

In September 2026 the FDA granted 510(k) clearance (K260906) to a2z Radiology AI for a2z-Abdo-Triage. The software processes adult abdominopelvic CT studies and flags suspected positive cases across 10 acute and traumatic findings, including obstructing ureteral stone, retroperitoneal hematoma and acute appendicitis. It was cleared under regulation number 892.2080, product codes QAS and QFM, and used an earlier a2z device, a2z-Unified-Triage (K252366), as its predicate.

The clearance [included an authorized PCCP](https://www.accessdata.fda.gov/cdrh_docs/pdf26/K260906.pdf). Its Description of Modifications lets a2z expand training data volume and diversity, refine annotation quality, optimize ensemble composition, adjust data augmentation parameters and modify neural network architecture components. The plan says in plain terms that it does not authorize adaptive algorithms that continuously learn in the field.

Every change must be trained, validated on an independent test dataset read by U.S. board-certified radiologists, and mathematically locked before release. Device labeling is then updated to tell clinicians about the change. The lesson is that the FDA is comfortable with manufacturer-controlled iteration under a plan, and the field-learning boundary is written into the plan itself. That fits with the lifecycle-monitoring approach above, because a locked model gives you clean version boundaries for lineage and rollback.

### Epic and Viz.ai: What Clearance Looks Like at Scale

Epic Systems got its first AI/ML clearance through CDRH in September 2026 for a clinical decision support tool that turns 195 chart inputs into a risk score for trained clinicians. Viz.ai received 510(k) clearance in February 2024 for Viz ICH Plus, which identifies, labels and quantifies intracerebral hemorrhage volume on non-contrast CT images.

![Three tiles: 195 chart inputs, a 44 percent reduction in stroke door-in-door-out times, and about $1,040 per encounter.](https://dzrge5zzbsh6q.cloudfront.net/epic-vizai-clearance-at-scale.webp)

These numbers show why buyers adopt cleared AI and why vendors must keep it safe through updates.

An aggregator, AI Central, reports that Viz.ai's integration with hospital mobile communication platforms produced a 44 percent reduction in door-in-door-out times for stroke patients in peer-reviewed studies. It also reports that Medicare's New Technology Add-on Payment pathway pays hospitals approximately $1,040 per qualifying patient encounter. Those numbers show why buyers adopt cleared AI, and why the vendors behind it have to keep it safe through updates.

## What It Costs

A PCCP moves cost from regulatory fees to engineering. For FY 2027 the standard 510(k) user fee is $28,653, or $7,163 for a qualified small business. An authorized PCCP removes the per-update submission fee, since the review cost of the plan is absorbed into the initial marketing submission.

As an illustration of the old route, four quarterly updates that each needed their own 510(k) would cost $114,612 in standard fees, or $28,652 at the small business rate, before consulting costs. That arithmetic is mine, not the FDA's. Not every update needs a new 510(k), only those that could significantly affect safety or effectiveness.

The engineering side is not cheap. Taction Software estimates that implementing real-world performance monitoring, drift detection and reporting typically takes 200 to 800 hours. Taction, which sells this work, also says manufacturers can hire dedicated FDA software-as-a-medical-device engineers at about $8,000 per engineer per month.

Cost item

Lock and submit

PCCP route

**FDA fee per update**

$28,653 standard, $7,163 small business

None per update; covered in the initial submission

**Engineering setup**

Lower, static validation

200 to 800 hours for monitoring, drift detection and reporting

**Ongoing upkeep**

Lower, reactive complaint handling

About $8,000 a month per dedicated engineer, at Taction's hiring rate

A paper in PubMed Central raises a risk in this cost structure. The capital and staffing needed to run automated PCCP infrastructure could concentrate innovation in a small group of well-funded companies and make it hard for early-stage startups and small and mid-size firms to compete. If you are a small team, the 200 to 800 hour figure is the number to plan around first.

### Data Quality Is the Bottleneck

Money is not the only constraint. In 2026 surveys cited by Halkwinds, 67 percent of technology leaders name data quality as the main barrier to enterprise AI deployment, ahead of model capability or talent, and median time to AI production in healthcare is 9.2 months. A study in PubMed Central finds that models that do well on curated academic datasets often fail in live workflows because of missing data, unstandardized metadata and shifting EHR protocols.

That is the same problem versioned pipelines address. If you cannot reproduce and describe your training data, you cannot write a Modification Protocol that holds up.

Demand is not the issue. Carly's 2026 statistics page estimates that 81 percent of U.S. physicians report using AI in practice, up from 38 percent in 2023, and that healthcare organizations spent about $1.4 billion on generative AI tools in 2025, nearly triple the 2024 spend, with $600 million of it on ambient clinical documentation.

## The Predicate Trap

Under Section 515C, a device changed under a PCCP cannot serve as a predicate in its changed form, so a later 510(k) has to compare against the version cleared before those changes. MedDeviceGuide spells out the consequence. Suppose an incumbent clears "Device A (Version 1)" and uses its PCCP to iterate to a much better Version 5 over two years. A competitor cannot cite Version 5 as its predicate, and has to show substantial equivalence to Version 1. That lasts until the incumbent files again. [Once a later 510(k) clears a device with those changes in it](https://www.raps.org/resource/final-fda-guidance-on-pccp-includes-clarification.html), that version can serve as a predicate.

Given that about 97 percent of AI device clearances go through 510(k), this rule shapes competition. An early clearance with a well-built PCCP lets the incumbent improve while rivals are measured against the original, for as long as it makes its changes under the plan. For a product lead deciding when to file, that is a reason to get the first clearance right and to put the PCCP in it, rather than adding one later.

## Alignment With ISO/IEC 42001 and Other Regulators

ISO/IEC 42001, published in December 2023, is the first international standard for AI management systems. ISO 27001 focuses on information security. ISO 42001 covers AI-specific lifecycle risks such as algorithmic bias, explainability and clinical collateral damage. The overlap with a PCCP is practical:

-   **Risk-based governance.** ISO 42001 requires structured risk assessment for models that influence clinical decisions, which mirrors the PCCP Impact Assessment.
-   **Clinical safety monitoring.** The standard requires continuous tracking of drift and reliability, which supports the real-world observability the Modification Protocol needs.
-   **Transparency.** Both call for documented model cards covering intended use, clinical limits and the populations the model was validated on.

According to the ISO 42001 healthcare AI legal and regulatory mapping, certification typically costs early-stage startups between $15,000 and $35,000 for initial audits, and it does not confer HIPAA or FDA compliance. It does build the quality management infrastructure that helps you defend a PCCP submission.

Other regulators are converging. IMDRF's N90 uses the same three-part PCCP structure as the FDA. The FDA, Health Canada and the UK's MHRA have co-authored guiding principles on PCCPs. The EU Medical Device Regulation has no formal PCCP mechanism, so significant post-market software changes in Europe still go through individual negotiation with Notified Bodies. If you sell in both markets, expect a U.S. PCCP to cover only part of your change management.

## A Six-Step Plan for the Next Two Quarters

The order matters. Teams that write the protocol before they build the infrastructure end up promising things their systems cannot do.

1.  **Audit your quality management system.** Traditional medical device QMS frameworks fall short for AI. Add ML-specific design inputs such as quantitative model performance targets, training data demographic requirements and statistical definitions of ground truth. Drop the idea that an IT complaint log will satisfy continuous performance monitoring.
2.  **Build the versioned data and MLOps layer first.** Make every dataset and feature a versioned, immutable artifact. Deploy a model registry that links production weights to their exact training data and validation results.
3.  **Write the Modification Protocol with numbers in it.** Regulatory affairs and ML engineers should draft it together. Replace vague goals with acceptance criteria, for example requiring sensitivity of at least 90 percent and specificity of at least 95 percent with 95 percent confidence intervals. That is only an example, not a threshold to copy.
4.  **Stand up the isolated shadow environment.** Use containers, route live EHR data through the candidate model in parallel and collect telemetry. Physically prevent the candidate from showing predictions in the clinical interface or changing patient records.
5.  **Automate telemetry and rollback.** Deploy sidecar logging and dashboards for latency, input distributions and confidence scores. Set hard triggers so that when drift crosses the PCCP's thresholds, the clinical safety team is alerted and the system rolls back to the last validated version.
6.  **Use the Pre-Submission pathway.** Before you file a 510(k) or De Novo application, take the Description of Modifications and the statistical approach of the Modification Protocol to the FDA through the Q-Submission program. The goal is agreement on where the PCCP boundaries sit, which lowers the chance of rejection in formal review.

## Conclusion

PCCPs for AI devices are already final guidance, and the lifecycle guidance on the A-list is likely to push post-market monitoring, data lineage and tested rollback further from good practice toward what a submission has to show. Start with step two. A registry that links every production model to its training data is the piece that every other part of the PCCP depends on, and it takes the longest to retrofit.

## FAQ

### What kinds of model changes can a PCCP cover?

Typical items include algorithm parameter adjustments, neural network threshold changes, training data updates on new patient cohorts, and user interface changes that affect AI functionality. Retraining on new data from the approved intended population fits, and so does updating compatibility for a new operating system or imaging hardware. Changes to intended use, the indicated patient population or the risk classification need a new premarket submission.

### Does a PCCP allow a model to keep learning in the field?

The a2z Radiology AI clearance shows the current line. Its PCCP covers changes the manufacturer trains, validates on an independent test set and locks before release. It says explicitly that it does not authorize adaptive algorithms that learn continuously in the field. Plan around locked updates unless the FDA tells you otherwise.

### Is a hospital complaint log enough for post-market monitoring?

No. Greenlight Guru says the FDA does not recognize a standard IT complaint log as performance monitoring, and expects active, real-time drift detection. In practice that means runtime telemetry on latency, throughput, input distributions and confidence scores, with dashboards for precision, recall and calibration across demographic slices.

### Do we need shadow mode?

Shadow mode is the common healthcare pattern for validating a modified model on live data without affecting care, and the ARPA-H ADVOCATE teams use it throughout. It lets you test against the thresholds in your PCCP's performance evaluation protocol on real hospital data. We found nothing that says the FDA mandates shadow mode by name, so treat it as a strong pattern for meeting the protocol, not a stated rule.

### What does a PCCP cost compared with filing a new 510(k) for each update?

The FY 2027 standard 510(k) fee is $28,653, or $7,163 for a qualified small business, and an authorized PCCP removes that fee per update. The engineering cost is the trade. Taction Software estimates 200 to 800 hours to implement real-world performance monitoring, drift detection and reporting, and says manufacturers can hire a dedicated SaMD engineer at about $8,000 a month.

### Can a competitor use our latest model version as its predicate?

Not while those changes sit under your PCCP. A competitor has to compare against the version cleared before them, even if you have iterated to Version 5. Once a later 510(k) clears a version that includes the changes, that version can serve as a predicate.

### Will a U.S. PCCP cover us in Europe?

Not fully. The EU Medical Device Regulation has no formal PCCP mechanism, so significant post-market software changes there still involve individual negotiation with Notified Bodies. IMDRF N90 uses the same three-part PCCP structure as the FDA. The FDA, Health Canada and the MHRA have co-authored guiding principles on PCCPs, although the MHRA has no PCCP route yet and plans draft guidance by December 2026.

## Supporting Links

-   [FDA list of CDRH proposed guidances for FY 2027](https://www.fda.gov/medical-devices/guidance-documents-medical-devices-and-radiation-emitting-products/cdrh-proposed-guidances-fiscal-year-2027-fy-2027)
-   [FDA final guidance on PCCPs for AI-enabled device software functions](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)
-   [RAPS on the final PCCP guidance and predicates](https://www.raps.org/resource/final-fda-guidance-on-pccp-includes-clarification.html)
-   [Hadrian on running AI penetration testing safely](https://hadrian.io/blog/what-ai-penetration-testing-needs-to-run-safely-at-speed)
-   [Complizen on the CDRH FY 2027 guidance list for AI and 510(k)](https://www.complizen.ai/post/cdrh-fy2027-guidance-list-what-fda-plans-for-ai-and-510k)
-   [AdvaMed's comments on a risk-based FDA approach to generative AI devices](https://www.advamed.org/industry-updates/news/advamed-encourages-risk-based-fda-approach-to-generative-ai-medical-devices/)
-   [COSM on ARPA-H ADVOCATE and FDA-authorized autonomous clinical AI](https://www.cosmhq.com/resources-posts/arpa-h-advocate-fda-authorized-autonomous-clinical-ai)
-   [Greenlight Guru on PCCP requirements for SaMD and AI](https://www.greenlight.guru/blog/pccp-fda-requirements-samd-ai)
-   [FDA 510(k) summary for a2z-Abdo-Triage (K260906)](https://www.accessdata.fda.gov/cdrh_docs/pdf26/K260906.pdf)

## About Baytech

At [Baytech Consulting](https://www.baytechconsulting.com/services/partnership-approach-baytech-consulting), we specialize in guiding businesses through this process, helping you build scalable, efficient, and high-performing software that evolves with your needs. Our MVP first approach helps our clients minimize upfront costs and maximize ROI. Ready to take the next step in your software development journey? [**Contact us today**](https://www.baytechconsulting.com/contact) to learn how we can help you achieve your goals with a phased development approach.

## **About the Author**

![](https://dzrge5zzbsh6q.cloudfront.net/_convertToWebP/60289/Bryan_Profile_Picture_V2.webp)

Bryan Reynolds is an accomplished technology executive with more than 25 years of experience leading innovation in the software industry. As the CEO and founder of [Baytech Consulting](https://www.baytechconsulting.com), he has built a reputation for delivering custom software solutions that help businesses streamline operations, enhance customer experiences, and drive growth.

Bryan’s expertise spans custom [software development](https://www.baytechconsulting.com/services/challenges-of-custom-software-development), [cloud infrastructure](https://www.baytechconsulting.com/services/cloud-development-and-deployment-consulting), [artificial intelligence](https://www.baytechconsulting.com/services/ai-powered), and strategic business consulting, making him a trusted advisor and thought leader across a wide range of industries.