AI Inspection Reports
Control how AI drafts and refines inspection report prose with template briefs, refinement, and guardrails.
What the AI Generates
Control how AI drafts inspection report prose by setting a template brief, refining the report with free-text instructions, reviewing clauses, and managing open questions.
For the full report lifecycle, attestation, and emission workflow, see Inspection Reports.
AI inspection reports use a hybrid pipeline. The system assembles an evidence bundle from checklist answers, inspection findings, photos, video keyframes, transcripts, and template context. A deterministic process creates the report structure, an AI prose pass smooths the narrative, and a separate AI pass generates open questions for evidence gaps.
Report blocks per finding
Each finding can contain the following blocks:
| Block | What it contains |
|---|---|
| Observation | The observed condition or state. |
| Deviation | The difference between the observed condition and the expected standard. |
| Consequence | The potential result or impact of the deviation. |
| Recommended action | The action recommended to address the finding. |
| Evidence references | References to the evidence supporting the finding. |
| Media references | References to photos or video evidence associated with the finding. |
Open Questions
The AI generates open questions when the available evidence does not support a confident conclusion. Questions surface missing or ambiguous information instead of asking the AI to guess.
The question types are:
- Placement: Evidence could not be linked to a specific finding.
- Media-failure: A media item was excluded, failed, or timed out.
- Content: A finding has a grade, but the available evidence does not fully support it, such as when a required photo is missing.
Image budget
The AI does not receive every image in an inspection. Each report draft sends at most 40 images. Every photo counts as one image, and every video counts as one image, its first keyframe. The limit is fixed and does not grow with the number of findings. When an inspection has 40 or fewer of these, all of them are sent, except media that has not finished processing. When it has more, the system takes one image per finding in turn, so a finding with many images cannot crowd out the others. It starts with the highest-priority findings, such as graded findings, and prefers photos over video keyframes. The image selection is recorded with the AI run so reviewers can understand which visual evidence informed the draft.
What You Control
Template AI brief
The template AI brief provides tone, emphasis, and narrative guidance for the report. Use it to indicate how the report should present finding content.
The brief is treated as untrusted input. The AI can use it to shape prose, but it cannot override regulated facts or the report's factual structure.
Refinement instruction
Use a free-text refinement instruction to adjust report prose, such as changing emphasis, shortening a section, or expanding an explanation. Each refinement creates a new report version, so earlier versions remain available.
The system re-runs factual verification after every refinement. It fails closed if regulated facts change, and it does not create a duplicate version when the result is identical to the current version. Refinement instructions can contain up to 2,000 characters.
Clause-level review
Review generated clauses individually or as a group:
| Action | What it does |
|---|---|
| Accept | Marks a clause as reviewed without changing its wording. |
| Correct | Replaces or deletes inaccurate text and creates a new version. |
| Accept all | Marks all remaining inferred clauses as reviewed at once. |
| Restore | Returns to an earlier version and creates a new latest version from that content. |
Open Questions
Answer an open question when you have the missing context. Answering starts an AI refinement using your response as an instruction.
You can also dismiss a question when no further action is needed. For media-failure questions, retry the affected media instead of answering the question directly.
Report language
Report language is derived from the inspection, report, and template context. The AI writes generated questions and report prose in the resolved language.
Guardrails
The AI cannot change the report's regulated or factual structure. The deterministic process owns the structure, while the AI prose pass only smooths the narrative.
The AI cannot change:
- Grades
- Measurements
- Standard references
- Numbers
- Evidence references
- Clause count
- Finding structure
AI-generated report content cannot be sent as a customer document until the required review and attestation state is complete.
Plan Limits
Inspection report drafting, refining, placement, and related AI runs consume the inspection token pool. This pool is separate from the meeting summary pool, and the two pools do not share quota.
| Entitlement | Inspection Token Pool |
|---|---|
| No entitlement | 0 |
| Trial | 250,000 |
| Inspection add-on | 1,250,000 |
Inspection entitlement can come from a paid Inspection add-on, an active trial, chain or enterprise membership, or a super-admin override. Without an entitled Inspection capability, the pool is zero.
Inspection AI requires the Inspection module. Reading reports requires report read access, and editing or refining reports requires the appropriate report management permission.
When the Inspection pool is used up, the report page, the report refine chat, and the offer proposal panel show a notice titled "Inspection AI limit reached for this cycle". The notice says that new generation is paused, gives the date the allowance resets, and confirms that everything already generated stays usable. The guidance depends on who is looking:
- Users who can manage billing see a Chat with support button.
- Other users are told to contact their admin or owner to review the limit.
- Users in a chain company see a notice titled "AI limit reached for this cycle" instead. It says whether the shared chain pool or the company's own chain limit was reached, and that only Vadio can lift it.
The AI summary allowance is not affected, so meeting summaries keep working.