Skip to main content

Call Us: Oregon and SW Washington: (503) 353-0100: Contact Us | Business Hours: 7:00 AM to 5:00 PM Mon-Fri

Technology & tools / MAR-HY ARTICLES

Keep Unknown Cases Visible in HVAC Maintenance AI

Originally published February 25, 2025
Updated November 5, 2026 5 min read

A maintenance tool should have a supported way to say that it cannot assess a case. Missing data, unfamiliar equipment, or an input outside the evaluated application should not silently become “normal.” Define how unknown cases are displayed, routed, and reviewed before relying on a summary accuracy figure.

Distinguish a result from no supported result

A system may produce a classification, a risk estimate, an alert, or a statement that required information is unavailable. Those outputs need different meanings in the maintenance process. An empty field or quiet dashboard should not be assumed to mean equipment is healthy.

Ask the provider which situations cause the tool to withhold a result and which situations it may not recognize as outside its scope. A product’s ability to output a score does not prove that the score is supported for every input.

NIST’s AI trustworthiness discussion connects validity and reliability to intended use and expected conditions, and addresses human intervention and resilience. Apply that context to the actual HVAC use rather than treat a general model label as evidence of readiness.

Read accuracy together with coverage

Consider an invented evaluation of one hundred cases with independently established outcomes. The tool provides an answer for sixty cases and withholds an answer for forty. Among the sixty answered cases, fifty-four are correct and six are incorrect.

Correctness among answered cases is 54 divided by 60, or 90%. Coverage is 60 divided by 100, or 60%. The evaluation establishes fifty-four correct answers across the original hundred cases; it does not establish ninety correct answers across all hundred.

Hypothetical answered and unanswered cases
CategoryCountInterpretation
Answered correctly54Confirmed correct within the example’s evaluation method.
Answered incorrectly6Requires examination of the error and its consequences.
No answer40Remains outside the answered-case score and needs a defined process.

Do not automatically count the forty unanswered cases as either correct or wrong. Preserve them as a separate outcome and explain the denominator. These figures are teaching arithmetic, not vendor performance claims.

Understand the tradeoff

A tool that declines difficult cases may improve its answered-case score while leaving more work for people. That can be a useful design choice if the limits are visible and the service team can handle the remaining cases.

A tool that answers nearly everything may provide higher coverage while making more consequential mistakes. The right evaluation considers both the errors and the unresolved workload, not a single percentage.

Use representative cases and an appropriate independent evaluation method. If the unanswered cases concentrate in particular equipment or conditions, the headline average can conceal a gap in the exact work the contractor expects to support.

Give the unknown queue an owner

Define who reviews unavailable or unsupported results, how urgency is assessed, and what maintenance process applies meanwhile. Do not remove required inspections or other work merely because an AI service was expected to help prioritize it.

Keep equipment safety and technical decisions with qualified people using the applicable instructions and procedures. An unknown model result is not a reason for an unqualified customer or employee to perform extra testing.

Make the customer-facing explanation accurate. “The service could not assess this case” is different from “no issue was found.” Avoid allowing the interface to turn uncertainty into reassurance.

Use a configuration-change example

In an invented installation, the equipment or control configuration changes after the monitoring tool was evaluated. The data continues arriving, but the earlier validation may no longer describe the new arrangement.

The responsible team identifies the change, checks the supported application, and uses its approved maintenance process while the provider and qualified reviewers resolve the tool’s suitability. It does not assume that continued connectivity proves continued validity.

Retain which model or service version produced earlier results. A later update should not silently rewrite the meaning of an old output or conceal a period when coverage was uncertain.

Test loss and recovery

Prioritize unresolved cases through the appropriate maintenance and safety process, not merely in the order an AI screen lists them. An unanswered case involving a consequential function may deserve different attention from a low-consequence information gap. Qualified people need to establish that distinction for the actual equipment.

Keep the reason for abstention visible where available. Missing inputs, an unsupported model, and an unavailable service call for different fixes. A generic “AI error” label can route all three to the wrong person.

Do not hide the queue by relaxing the tool’s thresholds until it answers everything. Any change to model settings or acceptance criteria needs the appropriate evaluation and authorization. More outputs are not automatically more useful outputs.

Compare coverage before and after a change using a suitable common basis. If the equipment population or evaluated cases also changed, the difference may not describe an improvement in the model itself.

Ask how the tool shows missing inputs, unavailable service, an unsupported equipment identity, and recovery after a gap. Use approved tests and protect customer information.

Verify that the relevant people receive a usable exception rather than only a technical error message. If the connection resumes, determine what happened to cases that arrived or changed during the interruption.

Review the workload created by abstentions and failures alongside useful results. A service that leaves a large unowned queue is incomplete even if its answered-case score is high.

Bring Mar-Hy exact equipment and documentation questions where product information is needed. Keep the AI provider’s claims, the contractor’s maintenance responsibilities, and qualified technical judgment distinct.

PROSTOCK AT MAR-HY

PROSTOCK PROFESSIONAL QUALITY HVAC PARTS AND SUPPLIES

Mar-Hy Prostock HVAC parts and supplies store in Milwaukie

Support for the next service call.

Mar-Hy Distributors adopted the Prostock concept to enhance our support of local HVAC contractors. Since opening our Prostock store, we have received positive feedback from contractors on how they have improved response time to their customers. Our HVAC contractors have been able to stay ahead of the competition with access to our extensive inventory of Prostock Parts.

Quality parts. Contractor support.

Our Prostock replacement parts department is committed to improving the business of our HVAC contractors by providing quality services, competitive pricing, and products and supplies that earn customer trust and build lifetime customer satisfaction. We value our contractors and will continue to provide top of the line service for all of your HVAC needs.