Test video AI as an evidence organiser: ask it to describe visible events, attach timestamps and flag what the recording cannot establish. A technician should replay those moments, correct the observations and decide what inspection is needed. The useful output is a visual evidence handover that separates what appears in the clip from what a qualified person concludes.
For a maintenance team, the trial should establish whether a reviewer can locate an event and understand the recording’s limits. A plausible explanation of a fault does not demonstrate diagnostic accuracy. Keep the test focused on observations that someone can check against the original footage.
Define the observation task before uploading a clip
Choose a narrow question such as: “What visible steps occur between the operator pressing the control and the equipment stopping?” Asking “What is wrong with this machine?” invites a diagnosis when the recording may show only a symptom.
Google’s Gemini documentation describes video input, questions about video content and references to specific timestamps. These capabilities support a proposed observation workflow; they do not establish maintenance diagnostic accuracy. Source: Gemini video understanding
Write down the distinction your trial must preserve:
- Visible observation: an indicator changes colour after a control is pressed.
- Reported symptom: the person recording says the machine stops during operation.
- Technician conclusion: an inspection establishes the cause and appropriate response.
Ask for the first two separately. Reserve the third for the technician. An unreadable equipment label should remain unreadable, rather than becoming a guessed model number. Likewise, the absence of visible movement does not establish whether power was disconnected or an internal component failed.
Prepare footage that a technician can trace
Select an existing recording that your business is authorised to use. Keep the original and identify the specific copy supplied for analysis. Record the job reference, equipment reference if known, recording date if known, and supplier. Unknown details should remain unknown.
Choose footage showing the relevant sequence without unnecessary surrounding material. Do not ask staff to repeat a hazardous event for the camera. Where another view is needed, let the responsible technician decide whether and how it can be captured safely.
Preserve context when trimming. A close-up may show an indicator but omit the action preceding it. Record the excerpt’s starting point in the original, and state whether returned timestamps refer to the excerpt or original. Otherwise, the reviewer may replay the wrong moment.
The supplied Gemini documentation describes default visual sampling at one frame per second and warns that fast movement or rapid scene changes may be missed. It also documents clipping intervals and custom frame sampling, with an explicit restriction: these customisations are supported only in static processing mode. Assess those settings within that documented scope; they do not prove every event will be captured. Source: Gemini video processing and sampling
Before deployment, check current model access, account and plan eligibility, regional availability and applicable data handling terms. The supplied documentation is living guidance with no stated publication date.
Request descriptions that expose uncertainty
Use an instruction your maintenance lead can review:
Describe the visible sequence in this maintenance clip. Give a timestamp or time range for each observation. Separate visible events from statements made by the person recording. Mark unreadable labels, obscured components and events you cannot establish. Do not infer a fault, recommend a repair or declare the equipment safe. List questions for technician review.
Require descriptive wording. “The housing moves relative to the background” identifies something to inspect. “The mounting is defective” introduces a cause that the clip may not establish.
Hearing someone say “the bearing has gone” should produce a reported statement, not a confirmed diagnosis. If sound matters, the technician should listen directly. Record a conflict between speech and the visible sequence rather than choosing whichever account sounds more convincing.
A consistent format can keep these categories separate. OpenAI’s Structured Outputs documentation describes schema-constrained responses and handling for refusals or incomplete responses. This is a separate implementation option if your workflow uses OpenAI; it does not establish Gemini support or make observations true. A completed field still needs evidence review. Source: OpenAI Structured Outputs
Copy this visual evidence handover
Use this template for each clip. Complete the observation rows first, then let the technician record corrections and decisions. Add a row for each distinct event rather than combining several steps into a single conclusion.
Visual evidence handover
- Job reference: [reference]
- Equipment reference: [verified reference / unknown]
- Original recording: [file identifier and accessible location]
- Recording date and supplier: [details / unknown]
- Analysed copy or excerpt: [identifier; original start offset if trimmed]
- Timestamp basis: [original recording / excerpt]
- Reported symptom: [supplier’s statement; not a confirmed cause]
- Trial settings: [model/version, processing mode, applicable sampling settings and instruction version]
| Time or range | Evidence type | Observation or reported statement | Limit or uncertainty | Technician review |
|---|---|---|---|---|
| [MM:SS–MM:SS] | [visible / audible / reported] | [literal description] | [obscured, unreadable, ambiguous or none identified] | [confirmed / corrected / unsupported; explanation] |
- Missing evidence: [what the recording does not show]
- Duplicate check: [existing job or clip reference / no match found / unresolved]
- Questions for technician: [specific questions tied to observations]
- Technician decision: [further information / inspection required / other reviewed outcome]
- Approved next action and owner: [action, responsible person and agreed timing]
- Reviewer and review date: [details]
- Fault conclusion: [not established from this clip / technician conclusion with separate supporting evidence]
Handover rule: No repair instruction, safety clearance or fault claim leaves this record without the responsible technician’s review.
Work through ordinary, unclear and duplicate cases
All identifiers, timings and events below are hypothetical examples, not test results.
Ordinary case: A clip labelled MC-A shows an operator pressing a control at 00:08. An indicator illuminates at 00:10, and visible movement stops at 00:24. The AI records these events without naming a cause. The technician replays each range, confirms the sequence and requests an inspection using the team’s established procedure. The handover says “movement stops”; the fault field remains “not established from this clip”.
Missing evidence: Another clip begins after the equipment has stopped. The AI cannot establish the starting sequence. Expected handling is to record that gap and ask the technician whether an operator account, existing operating record or safely obtained additional footage would help. The system should not reconstruct an unseen start-up. An absent label also leaves the equipment reference unresolved until someone checks it.
Ambiguous evidence: Glare makes an indicator appear red in one moment and orange in another. The output should mark the colour as unclear. The technician checks the original footage and, if necessary, the equipment directly. Until resolved, the handover must not describe a specific alarm state. A precise timestamp can help locate the glare without resolving its meaning.
Duplicate submission: A second upload has a different filename but appears to contain the same event as MC-A. A coordinator compares source files and job references before linking it to the existing record. If identity remains uncertain, retain it as a possible duplicate for review. Repeated uploads must not become evidence of repeated failures; a confirmed duplicate belongs with the existing event.
Score the trial against technician-reviewed evidence
Have a technician prepare a reference record before comparing it with the AI output. Identify supported events, relevant limits and statements that cannot be verified. Comparing two AI summaries would not provide the same independent check.
For each trial clip, record:
- Supported observations captured and relevant events missed.
- Unsupported causes, invented labels or incorrect sequence claims.
- Whether timestamps lead to the stated evidence.
- Whether missing information and uncertainty were preserved.
- Whether duplicate handling kept the job record accurate.
Agree acceptance rules before reviewing results. A proposed rule is to return any handover containing an unsupported fault claim for correction. Keep technician review mandatory during the trial. Record review effort if efficiency matters, including time spent correcting summaries; do not assume generation saves time.
Start with handovers only. If later integration creates work orders or sends messages, review those actions separately. n8n documents tool-level approval: approval allows execution and denial cancels it. That gate requires workflow configuration; video analysis does not automatically provide it. Source: n8n human review for AI tool calls
FAQ: testing maintenance video AI
Can we say the AI found the fault if the technician agrees?
State what the technician established and reference the supporting inspection evidence. Agreement with an AI suggestion alone does not establish causation. Keep the original observation separate from the later diagnosis so the record shows how the conclusion was reached.
What should happen when a precise timestamp points to an unclear event?
Replay the surrounding interval. Precise formatting does not make the description reliable. Correct the row, widen the range if appropriate, or mark the observation unsupported. Preserve the uncertainty if the original recording cannot resolve it.
Do we need a custom agent for the first maintenance clip trial?
A manual trial can establish whether the handover is useful. Use the AI agents versus automation comparison to assess the workflow, and the custom AI agent definition to clarify terminology before commissioning an integration.
If your business needs help making a reviewed handover repeatable, explore custom AI agents, the custom agent workflow guide and AI automation services, then get in touch with an approved sample clip and your technician’s review criteria.