Extension

Every extension is a new pathway map before it is a new purchase.

Extension means one of four things. A new use case. A new care setting. A new operator population. Or a new capability layered onto a use case already running. Each of them sends you back to Phase 2, because each changes the constraints, and constraints are what determine requirements.

The reason extension comes last is arithmetic rather than caution. Extension multiplies whatever the program already is. In a program with functioning archive integration, complete documentation, current privileging, and running quality assurance, it multiplies value. In a program without those things, it multiplies volume that cannot be stored, documented, billed, or reviewed. The failure is faster and larger, and it is usually attributed to the new capability rather than to the sequence.

Where the AI capabilities actually fit

Two distinct capabilities get collapsed under one label, and separating them is what makes the decision clean. Both are real, both are cleared for defined uses, and neither is a program objective.

Guided acquisition answers an operator constraint. It allows clinicians and staff without sonographic training to obtain diagnostic-quality images in defined applications, with reported performance in some applications approaching expert benchmarks and only a small share of studies requiring specialist review. If the pathway map shows that the people available at the point of need cannot reliably acquire the images the decision requires, this is the capability that resolves it. If trained operators are already present where the study needs to happen, it resolves nothing and adds validation and monitoring burden.

Post-processing and automated quantification answer a reader constraint. Automated measurement, preliminary quantification, and standardized presentation reduce interpretation time and reduce variation between readers. If the pathway map shows that interpretation capacity is the bottleneck, that reads happen remotely, or that the decision turns on a threshold where reader variability would change management, this is the capability that resolves it. If the performing clinician interprets at the bedside and acts immediately, the case is weaker and should be made on consistency rather than on throughput.

The two constraints are independent. An organization can have one, both, or neither, and the answer differs by use case within the same organization. A rural clinic doing screening echoes with medical assistants has an acute operator constraint and no reader constraint until volume grows. A cardiology practice with a growing echo backlog has a reader constraint and no operator constraint at all. Buying the same package for both is how organizations pay for capability they do not use and miss capability they need.

There is a third distinction worth holding onto, and it is not about AI at all. Some systems produce a better image or a faster measurement. Some produce a structured, coded inference that can populate a registry, feed a risk model, support a quality measure, or trigger a referral. The first improves a single study. The second determines whether the program's output participates in anything downstream. If any row of your pathway map has a required downstream trigger, that row is specifying structured output, and it will not be satisfied by better pixels.

What extension covers

New use cases. Complete a new pathway map. Sequence by clinical consequence and evidence maturity rather than by clinician enthusiasm.

New care settings. Ambulatory, rural, post-acute, and home. Each brings its own authorization environment, its own connectivity constraints, its own supervision question, and frequently its own payer coverage answer. Extend one setting at a time and re-run the relevant gates.

New operator populations. Subject to the three-gate screen from Phase 4, completed in writing. This is where the market access constraint bites hardest and where the gap between demonstrated clinical capability and payment infrastructure is widest.

New capability on an existing use case. Guided acquisition, quantification, structured output, or longitudinal comparison. Justify each against a constraint on the map for that specific use case, and be willing to conclude that a capability being offered does not address one.

Performance monitoring across your population

An algorithm validated elsewhere has not been validated on your patients. Establish subgroup performance monitoring at deployment rather than after a problem surfaces, covering the demographic and clinical populations your organization actually serves.

The available frameworks are usable. The AHRQ and National Institute on Minority Health and Health Disparities approach divides the algorithm life cycle into problem formulation, data selection, algorithm development, deployment, and ongoing monitoring, with accountability defined at each phase. The AI for IMPACTS framework provides structure for evaluating clinician-facing AI tools across integration, monitoring, performance, acceptability, cost, technological safety, and scalability, which is closer to what a health system actually needs than a purely technical validation.

Monitoring is not a one-time validation exercise. Regulatory expectations increasingly assume total product life cycle oversight, with postmarket surveillance of real-world performance and attention to generalizability across populations. A health system deploying these tools inherits part of that obligation whether or not it plans for it.

Building the evidence you will need

If your organization intends to negotiate rather than absorb the cost of these capabilities, it needs health economics evidence, and in most cases that evidence does not yet exist in the form a payer requires. No cost-effectiveness model or budget impact analysis has been established for AI-guided acquisition by expanded operators in most deployment scenarios. Supplemental payment mechanisms exist for qualifying new technologies, including new technology add-on payment and transitional pass-through pathways, but they are narrow and they are not a substitute for a coding pathway.

Value-based arrangements represent the more realistic near-term route, because they do not require a code that does not exist. A system operating under risk for a population can capture the value of earlier detection and avoided imaging without waiting for fee-for-service infrastructure to catch up. That is a strategic choice available to some organizations and not others, and it should be made deliberately rather than discovered.

What to require from your vendor

Start with the constraint, not the capability. For each capability proposed, which row of our pathway map does it address, and what happens to that row if we decline it.

Cleared intended use stated precisely, including which operator populations are covered. Performance data in the hands of your operator population, in your care setting, on a patient population resembling yours. For quantification specifically, agreement against the reference standard for that measure rather than aggregate accuracy across a bundle of measures. Subgroup performance data, and a direct answer where it does not exist.

Structured, coded output with a published specification rather than a report. Model update and version control policy, including how you will be notified when the algorithm changes and what validation accompanies the change. Postmarket surveillance commitments and what performance data the vendor will share back.

Any health economics evidence supporting the deployment model, and an honest statement where none exists. Companies that treat market access as a commercialization track running in parallel with product development will be able to answer these questions. Companies that treat it as a downstream problem will not, and the difference tells you which relationship you are entering.

Failure mode

Capability is acquired because it was available rather than because a constraint required it. Guided acquisition goes to a department with trained operators. Quantification goes to a use case where the clinician interprets at the bedside anyway. Meanwhile the use case whose entire value depended on triggering a referral is running on narrative reports that trigger nothing. Everyone is using the technology, nobody can explain what changed, and the renewal conversation has no evidence to work with.

The more expensive version of the same failure is extension into an organization that has not cleared Phases 3 through 6. Scan volume increases sharply among operators whose privileging status is undefined, producing studies that cannot be stored reliably, documented completely, billed defensibly, or reviewed at the rate they are generated. The program's exposure grows in proportion to its success.

Gate criteria

Phases 3 through 6 clear before extension begins. A completed pathway map exists for the new use case, setting, operator population, or capability. Every proposed capability traces to a documented constraint on that map. The three-gate screen is documented for every operator population in scope. Subgroup performance monitoring is designed before deployment rather than after. The financial pathway is identified, whether fee-for-service, supplemental payment, or value-based, and stated in the business case.