AI for Elevator Service Companies: Callbacks, Inspections and Permits
How elevator companies can keep callbacks, inspection dates and permit records as structured memory, while licensed staff keep safety decisions.

An AI agent for an elevator service company works best as a structured memory and coordination layer. It can keep callback history, inspection due dates and permit records in one readable place, draft summaries and reminders, and route anything involving a safety judgment to a licensed technician or responsible engineer. It should never decide whether a unit is safe to run, whether a repair meets the applicable code, or whether an inspection has passed.
Key takeaways
- Callback history becomes useful when every return visit is linked to its unit, its controller and the original visit it follows.
- Inspection and permit dates belong in a tracked calendar with a named owner, not in individual inboxes.
- The agent can read, draft, flag and summarize. Licensed staff make safety and compliance determinations.
- A first engagement should cover one role, one set of records and one measurable success criterion.
- Read access, draft creation and write access to status fields should be separated, logged and reviewed on a schedule.
Why callbacks are a memory problem more than a scheduling problem
A callback is a return visit for a fault that was supposed to be resolved. Most elevator service companies already know how to dispatch a technician. What is harder is knowing, at the moment a call arrives, that the same car was visited recently, that the earlier technician suspected a door operator, and that a part is still on order.
That context usually lives in several places at once: the dispatch system, technician notes typed after a visit, photos on a phone, a spreadsheet kept by the office and a manager's memory. When the history is fragmented, callbacks are handled as new problems. The office books a visit, the technician starts from the customer's description, and the pattern is noticed only after a third call.
An agent with structured access to the service record can make those links visible before the dispatcher assigns the visit. It does not need to diagnose anything to be useful. It needs to say which visits are related and what each one recorded.
What belongs in the company record
The agent is only as good as the record it reads. Before any automation, the company should agree on a small set of fields that every visit and every callback uses in the same way. Free-text notes such as a short phrase saying the problem was handled cannot be searched or compared. The useful fields are plain and specific.
- A unit identifier that matches the building's own naming, plus the building and address it belongs to
- Equipment type and controller make and model as recorded on site, with a field that shows uncertainty
- Each visit linked to its unit, with date, technician and the symptom described in the customer's words
- Callback events linked to the original visit they follow, with a reason coded in a short list of categories
- Parts replaced and parts ordered, each tied to the visit that required them
- Inspection due dates by jurisdiction, with the responsible person named
- Permit numbers, the status shown on the authority's documents, and the date that status was last checked
- Summaries of correspondence with the building owner or manager, linked to the source message
This record is also what makes later questions answerable. When a building manager asks what was done to a car last year, the answer should come from linked entries, not from whoever remembers the visit.
A callback record is useful when a stranger to the job could follow the chain from the complaint to the outcome.
Inspection dates and permit records as tracked obligations
Inspection and permit work is a calendar problem with legal weight. A missed inspection window or an expired permit is not a clerical inconvenience, and it often surfaces at the worst moment, usually when a building manager asks for paperwork. An agent can keep these dates visible. It can list what is due in the coming period for each site, flag items with no named owner and prepare a packet of prior records for a person to review before anything is sent.
What the agent should not do is interpret the rules on its own. Requirements differ by jurisdiction and change over time, and the authority having jurisdiction is the source of truth. The safer pattern is to load the requirements the company actually works under from documents it has verified, keep the source and the verification date beside each rule, and have a licensed person confirm the rule set whenever it changes. When the agent states a due date, it should show where that date came from. If it cannot, it should say so and route the question to a person.
Permit status deserves the same discipline. A status recalled from memory is a guess. A status copied from the authority's record, with the time it was checked, is something the team can act on. The same logic applies to quality follow-up. The approach described in AI for Quality Management: Nonconformances and Audit Prep shows how open items stay traceable until someone closes them. Recording which system changed which field, and when, is the operational equivalent of the signal recording described in the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/), and it makes later review far easier.
Where safety-critical decisions must stay with licensed staff
This is the boundary that matters most. An AI agent should never be the party that decides whether an elevator is safe to run, whether a repair meets the applicable code, whether a unit may return to service after a fault, or whether an inspection has passed. Those decisions belong to licensed technicians, qualified engineers and the authorities that issue inspection results. The agent's role is to help those people act faster with better information.
A short decision guide helps the team draw that line in daily work.
- Can the agent read the full unit history, related callbacks and open permit items? Yes, within the roles it was granted.
- Can the agent draft a callback summary or a handover note for a technician? Yes, and a person reviews it before it is used.
- Can the agent propose a visit window? Yes, but the dispatcher approves the booking.
- Can the agent change a unit's status to in service or out of service? No. A licensed person does this in the system of record, under their own name.
- Can the agent treat a tenant's message as proof that a callback is resolved? No. Customer text describes a problem. It does not confirm a repair.
- Can the agent mark a permit or inspection as passed? No. It records what the authority's document says, with the source attached.
Messages from customers and building managers need particular care. Text from outside the company can contain instructions, deliberately or not, and an agent should treat that text as data to be interpreted, never as a command to act on. The [OWASP Top 10 for LLM Applications](https://genai.owasp.org/llm-top-10/) covers this class of risk, and the permission design should assume it will occur. Write access should be narrow, every write should be logged, and no single message should be able to change a safety-relevant field.
A step-by-step plan for the first engagement
A first engagement should be small enough to finish and specific enough to measure. A reasonable sequence for an elevator service company looks like this.
- Choose one role, such as callback review for one maintenance portfolio, rather than a general office assistant.
- List the records that role needs to read and the ones it should never see, such as payroll or personal contact details beyond what the task requires.
- Agree on the company record fields and correct the most common gaps before connecting any system.
- Write down the approval boundaries: what the agent drafts, what it proposes and what only a named person may do.
- Define one observable success criterion, for example that every open callback has a named owner and a linked original visit.
- Run the role in shadow mode, where the agent produces its summaries and the team compares them with what the dispatcher actually did.
- Review the differences together, adjust the record fields or instructions, and only then widen access.
- Keep a log of every read and write, so that any disputed entry can be traced to its source.
Shadow mode is usually the most valuable step. It shows where the record is thin and where the agent's summary would have misled a dispatcher, which is more informative than any demonstration.
Illustrative scenario: a second callback on the same car
This scenario is illustrative and does not describe a real client. A maintenance company services a mid-rise building with several cars. A callback arrives for one car that stops at the wrong floor. The dispatcher opens the unit history and sees that a technician visited a few weeks earlier for a similar complaint, replaced a sensor and noted a possibly loose connection in the travel cable. The agent places both visits side by side, lists the parts replaced and flags that the earlier note has no follow-up entry.
The agent then drafts a handover summary for the senior technician. The summary states what the records show, what they do not show and the open question about the connection. It does not present a cause as a conclusion, and it does not change the unit's status. The senior technician inspects the car, decides on the repair and records the outcome under their own name. The agent links that outcome to the callback record. The gain is not a faster diagnosis. It is that no one has to rebuild the history from memory before arriving on site.
How Be Human handles this in managed operations
Be Human designs and operates this kind of setup as a managed service. A first engagement defines the single role, the memory and tools it needs, the approval boundaries and the success criterion, all within what the company's licensed staff have agreed to. After launch, managed operations can include monitoring of what the agent reads and writes, recurring evaluations against real callbacks, upkeep of the links to the service system, periodic permission review, incident response and controlled changes when the record or the rules change. Teams in adjacent trades face similar dispatch questions, and Superintelligent Agents for Plumbing Companies: Dispatch and Callbacks walks through one of them.
If you are unsure which task in your operation deserves to be the first one handed to an agent, take the free Be Human AI diagnostic to find it.
Frequently asked questions
Can an AI agent decide whether an elevator is safe to return to service? No. The agent can organize the unit history, flag open items and draft a summary for a licensed technician. The decision, and the record of it, stays with a named licensed person in the company's system of record.
Does the company need to replace its current software to use an agent? Not necessarily. The agent needs read access to the records the team already uses, and it works best once the most important fields are consistent. A first engagement often starts by correcting the few fields that matter most for callbacks and permits.
Which permissions should the agent have at the start? Begin with read access to unit history, callbacks and permit dates, plus the ability to create drafts and flags. Writes to status fields, inspection results or customer-facing messages should require approval from a named person. Permissions should be reviewed on a schedule rather than granted once and forgotten.
How do you know the agent is actually helping? Measure one criterion agreed before launch, such as whether every open callback has a named owner and a linked original visit. Compare the agent's summaries with what dispatchers actually did during a shadow period, and keep a log of every change so that errors can be traced and corrected.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


