Superintelligence for Elevator Maintenance: Call Triage and Reports

Superintelligence for Elevator Maintenance: Call Triage and Reports

How elevator maintenance firms can use AI for emergency call triage, maintenance scheduling and technical reports, with limits, logs and human review.

•

Want to apply this in your company?

Book a call with Juan →

Superintelligence, in the sense used by the press and in marketing campaigns, is not a product an elevator company can hire to triage calls or sign technical reports. What works today is an AI agent with clear rules: it classifies calls, suggests maintenance windows, and prepares report drafts from what the technician recorded. It does not decide anything involving passenger safety on its own, and it does not sign technical documents. That responsibility stays with the qualified technical manager.

Key takeaways

  • Calls with a possible trapped person or physical risk cannot depend on automatic classification. They go straight to the on-call technician and to emergency instructions.
  • Maintenance scheduling is a rules problem (frequency, crew, parts, building access), and it works best as a deterministic flow in which the AI summarizes and suggests without promising time slots.
  • The report is a technical document. The AI organizes the checklist and drafts the text, but validation and signature belong to the responsible technician.
  • Every call, suggestion and report change needs a readable record, for internal audit and for the building manager or condominium administrator.
  • A pilot with one call type and a defined group of buildings shows whether the process improves before any expansion.

What "superintelligence" means for an elevator operation

For a maintenance manager, the useful question is not whether the system is superintelligent. The useful questions are what kind of system it is, what data it works with, and what it is allowed to do. Many materials that use the term mix ordinary language models with promises that do not hold up in daily operations. It helps to separate text capabilities (summarizing, classifying, drafting) from real-world actions (opening a work order, booking a technician, changing a contract). The difference between a conversational assistant and an agent with tools is explained in ChatGPT and superintelligence: what it is and what it is not.

Call triage: what the AI classifies and what goes to a person

Triage is the most sensitive part of the process, because a classification error can delay the response to an emergency. The design rule should therefore be conservative: when in doubt, escalate to a person. A simple decision guide could look like this:

  • Signs of a trapped person, cab drop, burning smell, smoke, or water or sparks in the machine room or pit: immediate escalation to the on-call technician, an instruction message to whoever reported it (including contacting emergency services), and no diagnostic attempt by the agent.
  • A door that does not close or open with no one trapped, a stop between floors, or a new or repeated noise: priority call, with a technician on the same shift when the team confirms availability.
  • Routine wear, cab lighting, or a light noise with no change in operation: standard call, placed in the next maintenance window.
  • Questions about billing, contracts, warranty, or liability: routed to administration, with no opinion from the agent.

Keywords help as a first filter, but they are not enough. A resident can describe the same problem in other words, in a voice note, or with a photo. For that reason, the agent should treat any uncertainty as an emergency until a person confirms otherwise. Test scenarios for these rules, including ambiguous messages and attempts to bypass the triage, should be built before the system talks to a real resident, as described in how to test AI agents with mocks and a sandbox.

Maintenance scheduling without promising what the crew does not have

Scheduling usually fails through promises, not calculations. The agent replies quickly but confirms a time slot that no technician has reserved. The flow below prevents that:

  • Register each elevator with model, age, date of the last report, the parts that wear most, and building access rules (front desk, caretaker, hours).
  • Set the visit frequency based on the maintenance contract and on the manufacturer's guidance and the applicable technical standard, as documented by the company.
  • Check the real technician calendar by region, along with parts available in stock.
  • Offer the building manager or caretaker a few possible windows, and confirm none of them until the reservation is saved in the system.
  • Send reminders through the agreed channel, respecting WhatsApp messaging rules such as those covered in the WhatsApp 24-hour customer service window.
  • Record the confirmation, any rescheduling, and the reason for each change, so the team can explain the history later.

Note that the agent does not decide the maintenance frequency or when a unit is fit for service. It carries out the rule the company set and asks for review when something falls outside the standard.

Technical reports: the draft is not the document

The report is where AI is most tempted to be useful and where it most needs to be contained. The recommended process separates three roles. The technician records measurements, photos, and checklist items in the field. The agent organizes that data and produces a draft text. The responsible technician reviews, corrects, and approves, and only then does the document go to the archive and to the client.

Design rules that apply to this flow:

  • If a required field is empty, the draft marks the item as pending. The agent does not fill in measurements that do not exist.
  • No item becomes compliant without a matching record from the technician.
  • An approved report cannot be changed by the agent. Any correction creates a new version, with author, date, and reason.
  • The signature follows the process the company already uses for technical documents, never an automated shortcut.
  • Every interaction with the draft is logged: what the agent received, what it generated, and what a person changed.

This design follows risk management principles found in the NIST AI Risk Management Framework, which recommends identifying, measuring, and controlling risks across the whole life cycle of a system, not only at launch.

Illustrative scenario: a passenger elevator call in a commercial building

Illustrative scenario: the front desk of a commercial building receives a WhatsApp message saying a person is trapped in the passenger elevator and the cab stopped between two floors. The agent recognizes the emergency signal and does not try to discuss time slots or suggest a restart. It sends an alert to the on-call technician through the defined channel, replies to the front desk with emergency instructions, and logs the time of the message, the time of the alert, and the person who took over. The following day, the technician visits the site and fills in the checklist on a phone. The agent generates the report draft with pending items marked. The responsible technician corrects one point, approves it, and the document is archived as a version.

The value of this process lies less in response speed than in clarity: every step has an owner, every decision has a record, and no step was left with the agent just because that was convenient.

Data, LGPD, and access control

An elevator operation handles personal data of residents, building managers, caretakers, and visitors, and cab images can show people. Brazil's data protection law (LGPD) requires the company to collect only what is needed for a stated purpose, to inform people about the use, and to protect the data. In practice, the agent should see the technician calendar without seeing the building's financial history, and it should have read-only access to past reports without the ability to alter them. Guidance on what a system may do with data and on enforcement comes from the Brazilian data protection authority (ANPD). Transferring data to services hosted outside the country needs its own review, covered in international data transfer under LGPD when using AI providers.

There is also a specific security risk. Messages from residents can contain instructions designed to push the agent into ignoring its rules. The OWASP Top 10 for LLM applications describes this type of attack, and the main defense is the same as in the design above: the agent does not have permission to do what the rules forbid, whatever the message asks.

How to apply the Juan Carlo method to this case

The method has four stages, and each has a clear role in this operation. First, map the current process: who receives the call, through which channel, how an emergency is identified today, and how long it takes for a technician to respond. Second, limit data and actions: the agent reads what it needs, opens calls, and suggests windows, but it does not change contracts, close work orders, or approve reports. Third, pilot a real flow with logs and a metric agreed in advance, for example the share of calls classified correctly in a monthly review by the team. Fourth, hand over the operation with documentation, and decide whether to expand, adjust, or stop. When the operation needs ongoing support, it can be delivered as a managed service through Be Human.

Frequently asked questions

Can an AI decide on its own whether a call is an emergency? We do not recommend it. The AI can suggest a classification based on keywords and the text of the report, but any sign of risk should be treated as an emergency and escalated to a person on call, who confirms and leads the response. When in doubt, the system should escalate.

Can the agent sign the maintenance report in place of the technician? No. The agent can organize data, draft the text, and flag pending items, but validation and signature belong to the qualified responsible technician, through the process the company already uses. Any change after approval creates a registered new version.

How long does it take to put an agent like this into operation? It depends on the size of the portfolio, the quality of equipment records, and the technician calendar, and there is no standard timeline that fits every company. The safest path is to start with one call type and a group of buildings, with logs and human review, and expand only when the agreed metric shows consistent results.

What data should an elevator maintenance agent not access? It does not need detailed financial data, residents' personal documents, or the full contract history. Access should follow the minimum needed for each task, with read or write permissions defined by role, which is what the LGPD requires when it asks that data collection and use have a clear purpose.

Sources

elevator maintenanceAI agentsfield servicetechnical reportsoperations