A hospital team trials a tool that suggests treatments and asks if the FDA will care. In the future ai in healthcare, the answer turns on function: tools that analyze images or signals face device oversight, while some advisory software does not.
Readers who want the wider context can start with our guide on how is biotechnology used in medicine, then return here for the AI-specific outlook and a readiness checklist.
Common misunderstandings about AI oversight
Most confusion about where clinical AI is heading comes from treating all software as one category. Regulators separate tools by what they do with data, and that split shapes which products reach routine use first. Hospitals that plan around the function of a tool, rather than its marketing label, avoid the most expensive surprises.
| Misunderstanding | What the evidence supports | Planning use |
|---|---|---|
| Decision support software is exempt from FDA review | A clinical decision support tool avoids device regulation only if it meets criteria, including that it does not process or analyze medical images or signals. | Check each tool against the criteria before purchase, not after deployment. |
| Pathology AI will sit outside device rules | Most predictive AI applications in pathology are expected to remain under FDA oversight because of how medical images are classified. | Treat computational pathology as a regulated purchase with vendor evidence requirements. |
| Generative AI is unregulated | Generative AI that does not analyze images but gives single recommendations might see reduced enforcement. That is a possibility, not an exemption. | Keep governance in place even where enforcement may be lighter. |
| The 21st Century Cures Act removed software from regulation | It excluded certain decision support software from the device definition; existing digital health policies still apply to software functions that are devices. | Classify every software function individually. |
Figures and conditions come from the FDA clinical decision support software guidance page and the College of American Pathologists summary of the updated guidance.
Regulatory timeline for clinical decision support
The regulatory path is a sequence of legal changes and guidance updates, and the dates below are the ones that anchor current planning.
- December 13, 2016: the 21st Century Cures Act was enacted and amended section 520 of the Federal Food, Drug, and Cosmetic Act to exclude certain decision support software from the device definition, according to the FDA.
- Criteria in section 520(o)(1)(E): the FDA guidance applies these criteria to decide which clinical decision support is non-device. Software outside them stays within device oversight.
- 01/13/2026: the College of American Pathologists reported that the FDA had released updated clinical decision support software guidance covering oversight of software for health care professionals.
The practical lesson is that the 2016 exclusion was narrow and the later guidance refines it. A health system should expect classification questions to be decided tool by tool. It should not assume that a favorable reading of one product carries over to another.
Where AI becomes routine first
Uses are most likely to become mainstream where the task is well defined, the output is reviewed by a clinician and the regulatory category is clear. The evidence points to imaging, decision support and supervised automation, in roughly that order of maturity.
Imaging and computational pathology
Imaging is the most established area. One industry review notes that AI has made a substantial impact in medical imaging, particularly in X-rays, MRI and CT scans. Computational pathology follows the same logic of pattern recognition on image data, which is also why it stays regulated. Image analysis cannot qualify for the non-device decision support route, so pathology teams should expect validation evidence and device-style oversight to remain part of adoption.
This is a point in favor of pathology AI for cautious buyers: regulated status means a vendor must meet an external bar. It is a cost for teams hoping for fast, lightly reviewed rollouts.
Clinical decision support
Decision support that consolidates information for a clinician is the category with the most regulatory room. Where a tool meets the statutory criteria, it can sit outside device rules. Where it analyzes images or signals, it cannot. Clinicians remain the decision makers in both cases, so the sensible internal rule is to document how the tool presents its reasoning and how a clinician can override it.
Language models and agents
Large language models raise a different question, because they often produce a single recommendation without analyzing an image. The CAP summary indicates such non-image generative applications might see reduced enforcement. That is a reason to pilot, not a reason to skip governance. For concrete applications, see Generative AI Use Cases In Healthcare: 7 High-Impact Applications For 2026 – Biolifetechnologies.
Agentic AI adds autonomy. One automation provider’s outlook on AI in healthcare advises always including human-in-the-loop verification for AI agent actions. That advice suits any workflow where an agent acts on records, orders or messages, and it is a sound default until a unit has local evidence that a task is safe to leave unchecked.
Readiness checklist for health systems
Preparation is mostly governance and measurement, not technology purchasing. The same automation provider offers practical direction that fits clinical and operational settings. Use the list below as a working sequence.
- Classify each tool by function. Record whether it analyzes images or signals, gives advice to a clinician, or acts autonomously. This decides the regulatory route and the review depth.
- Build governance in from day one. The provider recommends designing clinical and operational AI applications with governance built in from the start, rather than adding it after a pilot succeeds.
- Keep a human check on agent actions. Assign a named reviewer for each automated action, and write down when that reviewer can step back.
- Set KPIs before go-live. The same guidance says to set key performance indicators and measure results as automations run. Choose measures that reflect patient and staff impact, not only throughput.
- Revisit classification when guidance changes. Updated FDA guidance on decision support means earlier readings of a tool’s status may need review.
The order matters. Classification comes first because it determines how much evidence the vendor must supply. Governance and human review follow, and KPIs come last because they need a defined workflow to measure.
What to expect as personalized care expands
AI-supported decisions will increasingly sit alongside advanced therapies and individualized treatment planning. Readers following that thread can continue with Cell and Gene Therapy Explained: Breakthrough Treatments, Risks, and The Road Ahead (2026) – Biolifetechnologies. The shared lesson is that capability and oversight advance together. Tools that fit a recognized regulatory category and a supervised workflow will reach routine use sooner than tools that need new rules or unsupervised autonomy.
Frequently asked questions
| Question | Answer |
|---|---|
| Which statute section sets the non-device decision support criteria? | Section 520(o)(1)(E) of the FD&C Act sets the criteria that decision support software must meet to fall outside the device definition. |
| Does FDA policy cover software used by patients or caregivers? | Yes, where the software function meets the definition of a device. The FDA states that existing digital health policies continue to apply, including for patient or caregiver use. |
| What is an AI agent? | An AI agent can act autonomously and make decisions without human intervention, which is why human-in-the-loop verification is recommended for its actions. |
| Where can I find the FDA decision support guidance record? | It is filed under docket number FDA-2017-D-6569 on the FDA guidance page. |
