The cleaning inspection has existed in K–12 facilities management for decades. An inspector walks a building. They score each area on a one-to-five scale. The aggregate score tells the district whether the custodial operation is performing at standard.
APPA Level 3 is the common threshold — ordinary tidiness, the floor is clean, the surfaces have been wiped, the fixtures are functional. A Level 4 or 5 score in a space means something went wrong. A building scoring consistently below Level 3 means the program has a problem.
The inspection is not glamorous work. Neither is the data it generates. But it is the most structured operational data a facilities team produces, and most districts are doing almost nothing with it.
The scores go onto a clipboard, into a spreadsheet, or into a PDF that gets emailed to a supervisor. Nobody reads the trend. Nobody knows whether the cafeteria has been below standard for three consecutive months. Nobody knows whether the same custodian's spaces consistently score lower on Fridays. The data exists. The pattern is invisible.
What we're building
This is a tool Mesa AI is building for K–12 facilities operations. APPA inspections are already standard in many districts — supervisors walk buildings, score each space on a rotating schedule, file a form. The problem is everything after the score: the inspection ends when the form is submitted, the data sits in a folder, and the operational response is based on whoever happened to know that a building had a problem — not on a pattern anyone had actually reviewed.
The goal is a mobile-first tool that lets the inspector score each area on their phone, generates a structured summary in real time, flags spaces below standard, and routes the report to the right supervisor with an action recommendation — before the inspector leaves the building.
This is not a complicated use case. It is a well-defined one. The scoring standard already exists. The spaces are already defined. The threshold is already agreed. The only thing missing is a system that captures the score where the inspector is standing and does something useful with it before the data goes cold.
The inspection form is the data contract
Before anything is connected to any AI layer, we treat the inspection form the same way we treat any operational data input: define exactly what is being captured, by whom, in what format, and against what standard.
The APPA space classifications become the data schema. Each space type — classroom, restroom, corridor, common area, exterior — has its own scoring criteria. The five-point scale is written out in plain language inside the mobile form so the inspector is not scoring from memory. The form requires a photo for any space scored at Level 4 or 5. An unscored space cannot be submitted.
This sounds like form design. It is actually data governance. The model that reads these scores cannot do anything useful with them if the input is inconsistent. If a Level 3 in one inspector's judgment is a Level 2 in another's, the trend data is noise. The form enforces the standard — not because a manager is watching, but because the form will not let an ambiguous entry through.
The result is a consistent inspection record for every building, every visit, with a photo attached to every below-threshold score. That had not existed before. Supervisors had been making operational decisions based on institutional knowledge and direct observation. Now there is a data layer that is current, structured, and queryable.
What the model does
The AI role in this tool is narrow. It reads the submitted inspection scores. It generates a structured summary: aggregate score for the building, spaces below standard listed by area and score, comparison to the previous inspection if one exists, and a draft recommendation — which spaces need immediate attention, which are trending toward a problem, and which are stable.
The recommendation is a draft. It routes to the facilities supervisor, not to the custodial staff. The supervisor reads the inspection summary, reviews the photos attached to the flagged spaces, and decides what action to take. The model does not assign a crew member. It does not schedule a follow-up. It does not post to the work order system. It drafts and routes. The supervisor decides and acts.
This is the same gate that runs on every other system this practice has connected AI to. The model does not post to the system of record. The human approves before anything changes in the operation.
The reason this matters in a school district setting is the same reason it matters in any field service operation: the AI has seen the scores, but it has not seen the building. The supervisor who reviews the summary knows that the cafeteria always looks worse on the first day back from a holiday weekend. They know that the inspector who scored that building yesterday tends to score corridor floors low when they are damp. They know that the space flagged as a Level 4 is being renovated and the score is expected. The model cannot know any of this. The gate is not a formality — it is where the operational judgment lives.
What it changes
With the tool active, supervisors review inspection data they have not been able to see before: trend comparisons by building, by space type, and by inspection date. Buildings that have been below standard for multiple consecutive visits become visible as a pattern, not as individual incidents.
The custodial coaching conversation changes. Instead of a supervisor walking a building from memory and giving general feedback, the conversation becomes: here are the three spaces that have scored below standard on the last four inspections, here are the photos from each visit, what is happening in these areas. That is a different conversation. It is specific, it is documented, and it is hard to dismiss.
A district also gets, for the first time, a building-level quality baseline it can compare across sites. Which buildings are consistently above standard. Which are consistently below. Which supervisors' buildings are trending up after a coaching intervention and which are not. This is the kind of operational data that changes how a facilities program is managed — not because it replaces the judgment of the people running it, but because it makes the patterns visible that were always there and always invisible.
The right scope for a first field AI tool
A facilities inspection is not a complicated AI use case. The standard exists. The scoring criteria exist. The threshold is agreed. What has not existed, in most operations, is a system that captures the inspection in a structured format, generates a summary in real time, and routes it to the right person before it goes into a folder.
The model does not replace the inspector. It cannot evaluate whether a floor is clean or whether a restroom is stocked. Those are judgment calls made by a person standing in the space. What the model can do is read the score that person entered, compare it to the last score for that space, flag it if it is below standard, and draft a summary the supervisor can read before the inspection crew has left the building.
That is the right scope for a first field AI tool. Not dispatch. Not scheduling. Not crew management. Start where the data is structured, the standard is agreed, and the failure mode is a supervisor who doesn't know what they don't know.
The inspection is the first dataset a facilities operation has that is clean enough to run an AI on. Use it.