Your AI Read the Label. Was It the Right Unit?

November 9, 2026

By Anna with Oppy
Cover: AI-generated editorial illustration, not an equipment identification record.

A homeowner sends a crisp photo of an equipment label. Your AI extracts the model and serial number. The office adds them to the job.

One small complication: there are two condensers beside the house. The photo shows the other one.

This hypothetical failure is not an OCR problem. Every character could be correct. The association could still be wrong.

Before extracted text becomes an asset fact, someone must establish which physical unit it describes. A sharper photo cannot answer that question by itself.

The new signal: inferred is not approved

On November 3, Snowflake described Business Ontology, marked public preview soon, as a layer of business definitions with owners and approval status. Its announcement says approved definitions take precedence over conflicting inferred context.

That is a product announcement, not evidence that equipment identification has been solved. But the distinction is useful: what an agent can infer and what a business accepts as official are different things.

BCG's October 21 insurance architecture article makes a related recommendation: define what must happen before a process advances, including required outputs and stage gates. We can borrow that discipline without borrowing an insurance system.

For a residential service office, the gate is simple: do not promote label text into a confirmed equipment identity until a qualified person accepts the association.

Give one Oppy the comparison job

Oppy's Functions guide documents Analyze Image and Extract Text (OCR). Its agent setup walkthrough documents supplied knowledge and role-specific agents.

Those capabilities support a proposed internal evidence-review assistant. They do not establish a native asset registry, authenticated photo capture, parts database or equipment-matching guarantee.

There is a reason to keep the packet small. Inman's October 19 survey report says 29% of agent respondents withheld client information from AI because of privacy concerns. Its 460 responses span several industry roles; the agent subgroup size is not disclosed. This is self-reported behavior, not a population estimate or equipment study. Supply only appropriately authorized images and the job details needed for comparison.

Start manually. Staff supplies one approved job reference, the existing asset record and the photos. Use a supported, approved integration only if your organization has verified its access and behavior. Otherwise, a human moves the review into the work order.

The agent's job is to preserve observations, compare supplied evidence and expose discrepancies. Not diagnose the equipment. Not select a part. Not order anything.

Supply a pair, not just a crop

Give the agent:

A context image: the unit and its surroundings, with the intended unit identified by the person collecting the evidence.
A detail image: the label on that same unit, when safely accessible.
A capture note: who supplied the images, when they were received, the reported capture time if known, and the reported location.
The approved record: asset ID, recorded location, model and serial if available, plus the record's source and last-confirmed date.

Do not ask anyone to open equipment, cross a barrier or approach a hazard to improve the packet. Missing evidence is an acceptable result.

A staff member's statement that the images belong together is evidence to preserve, not proof of authenticity. Neither a timestamp nor a familiar-looking patio authenticates an asset.

Keep observation separate from acceptance

The draft review should contain four short sections:

Observed: visible label text, preserving uncertain characters and pointing to the source image.
Reported: the supplied location and claimed relationship between the images.
Compared: agreements, discrepancies and missing fields against the approved record.
Unresolved: the specific question a qualified technician must answer before the association is accepted.

Even a fully consistent packet remains provisional. Silence from the reviewer is not confirmation.

If the record says one serial and the image shows another, preserve both. Do not helpfully repair the record. That is how a tidy database acquires an untidy relationship with reality.

Copy this draft-only instruction

Role: Internal equipment-evidence reviewer.

Inputs: One approved job reference, one supplied asset record,
context image, label image, and capture notes. Identify missing inputs.

Preserve three separate categories:
OBSERVED: what is visible in each image.
REPORTED: what the submitter says about location, time, and identity.
RECORDED: what the supplied asset record states, with its source/date.

Transcribe visible label text. Mark uncertain characters explicitly.
Do not complete a partial serial number from the existing record.
Do not infer that two images show the same unit merely because
model, appearance, or address seems consistent.

Compare the supplied evidence. Return a draft with:
- image references and observed text
- agreements and discrepancies
- missing evidence
- one focused question for the qualified human reviewer

Always label the asset association PROVISIONAL.
Do not diagnose, select parts, certify compatibility, purchase,
change inventory, update asset identity, dispatch, or contact customers.
A qualified human must accept or reject the association separately.

A prompt is not an access control. Review actual enabled tools and permissions before testing. Oppy's documented Public/Admin gating restricts who can invoke functions; it does not make a write function read-only. Do not connect this pilot to operational write actions. If that boundary cannot be enforced in your setup, perform the comparison manually.

Test the readable mistakes

Start with synthetic or appropriately authorized examples, not a live customer decision. Include:

A perfectly readable label from the neighboring unit.
A detail image with no evidence connecting it to the context image.
A replacement unit whose label conflicts with an older record.
A missing or unreadable label that should remain unresolved.

Have a qualified technician establish the reference judgment independently of the AI draft.

For the pilot period, report the missed wrong-unit rate: technician-confirmed wrong-unit cases the agent failed to flag, divided by all technician-confirmed wrong-unit cases in the reviewed sample. Show both counts. If there are no confirmed wrong-unit cases, the rate is undefined, not zero.

Also report unresolved packets divided by all reviewed packets, and record the human review time. An agent that doubts everything may avoid false certainty while creating an impressive new clerical department.

These are proposed tests, not measured Oppy results.

Keep the useful boundary

The goal is not to make an AI employee a technician. It is to give the technician a cleaner comparison without quietly converting an image into authority.

If you are building with Oppy, try one internal, draft-only equipment review first. Confirm the unit before anyone acts on its identity.

Reading the label is useful. Knowing what it belongs to is the job.