Dispatch Agents for Locksmiths: Lockout Calls, Key Records and Limits
How a locksmith company can build a dispatch agent with key-code and job-record memory, approval limits, and no unsupervised access to vehicles or codes.

A dispatch agent for a locksmith company takes incoming lockout and rekey requests, checks them against job history and key-code references in its own memory, proposes a technician and time window, and sends anything touching codes, vehicle access or pricing to a person. Build it around one narrow role, a memory of job records and key-code references held under staff control, and tool permissions that stop short of unsupervised access to vehicles or codes. The agent can get better as underlying models improve, but each gain should change how it reasons, never what it is allowed to do.
Key takeaways
- Start with one dispatch task, such as triaging lockout calls in your service area, rather than a general front-desk agent.
- Keep key codes and cut references in a controlled store that the agent reads by reference, never as free text it can repeat.
- Give the agent read access to job history and draft-only actions. A technician confirms anything that touches a vehicle or a code.
- Define a success criterion you can observe, such as calls routed correctly without a manual correction during a weekly review.
- Treat a model upgrade as a change to test against past calls, not as a reason to grant new permissions.
Why dispatch is the right first role for a locksmith company
Lockout calls are urgent, varied and easy to get wrong in small ways. A caller may describe a vehicle by its color rather than its model, a building entrance may need a specific access window, and the same address may already appear in last season's records under a different name. A dispatcher who handles these calls all day is the bottleneck, and the mistakes that cost the most are usually routing mistakes: the wrong technician, the wrong window, or a job booked into a day with no spare capacity.
That makes dispatch a good first role. The inputs are mostly text and structured fields, the outputs are proposals a person can check, and the success criterion is easy to observe. A general assistant that answers every question on the phone line has far more surface area and far more ways to make an unsafe commitment. Narrow scope is what keeps the first version reviewable.
What memory the agent needs, and what it must never hold
The agent needs memory that makes its proposals consistent with your operation. That usually means job records with the address reference, service type, technician assigned, completion notes, and any prior visit to the same property or vehicle class. It also needs the vehicle classifications you already use, such as lock family or year range, stored as categories rather than as a long free-text history.
Key codes and cut references need separate handling. The agent should see an identifier that points to a record in a secured vault, and it should know that a record exists and who is authorized to retrieve it. It should never hold the code itself in a memory document, a prompt, or a conversation log. If a caller asks for a code, the correct answer is that a verified technician will handle the request in person or through your approved procedure.
Customer personal data should be limited to what dispatch needs. Apply the privacy rules that cover your customers, such as the LGPD in Brazil or the GDPR in Europe, and set a retention period for call transcripts that matches your own policy.
The agent should know that a key record exists, never what the key is.
Illustrative scenario: A caller says they are locked out of a sedan. The job history shows the same vehicle was serviced last spring by a different technician, and a key-code reference is held in the vault. The agent surfaces the prior visit, notes that the reference exists, offers two time windows, and flags the call for a dispatcher. It does not reveal, request or summarize the code. The dispatcher confirms the proposal and the verified technician handles the rest.
Approval boundaries: what the agent does alone and what needs a person
The most useful document you will write before launch is a short decision guide that says which actions the agent may take on its own, which it may only propose, and which it must never attempt. Keep it specific enough that a new dispatcher could apply it on the first day.
- Reading the intake, checking job history and confirming service area: the agent may do this alone.
- Proposing a technician and a time window: the agent drafts, and a dispatcher confirms before anything is sent to the customer.
- Any request involving a key code, a cut reference or unlocking a vehicle: a verified technician handles identity and authority outside the agent.
- Pricing, refunds, fees or exceptions: a person approves each case.
- Changing vault contents, tool permissions or the decision guide itself: the agent never does this, and a named staff member makes the change.
These limits belong in the tools, not only in the instructions. OWASP's authorization guidance describes the principle plainly: access decisions should be enforced where the action happens, not assumed from what a prompt says. If the dispatch tool cannot retrieve a code, no wording from a caller can make it do so.
Callers, text messages and voicemail transcripts can contain instructions aimed at the agent, such as a request to ignore its rules. Treat all incoming text as data to be read, not as commands. The OWASP cheat sheet on prompt injection prevention covers this class of risk in detail, and it is worth giving that section to whoever reviews your configuration.
A step-by-step plan for the first dispatch agent
A first version should be small enough to evaluate completely. The following sequence keeps each step observable.
- Choose one call type, for example residential lockouts within your core service area.
- Export a set of past calls with the outcome your dispatcher actually chose, including the technician, the window and any correction.
- Write the decision guide: what the agent may read, what it may propose, and what it must never attempt.
- Build the agent with read-only job records and key-code references, and with no tool that retrieves or displays a code.
- Run it in shadow mode, where it drafts proposals that dispatchers compare against their own decisions without sending anything to customers.
- Review every disagreement in a weekly session, and fix the rules, the records or the routing logic before touching the prompt.
- Move to live proposals only after the agent matches your dispatchers on the cases you care most about.
Notice that the first fixes are usually in the records and rules rather than in wording. A dispatcher who keeps correcting the same routing error often reveals a missing field in the job history, such as access windows for a building type, that the agent cannot infer.
How the agent improves as models get stronger, without getting more access
A more capable model can read messy intake text more accurately, handle an unusual request with fewer errors and explain its reasoning more clearly. Those gains matter for dispatch quality. They do not change what the agent is authorized to do, and they should not be used to justify a wider tool list.
The mechanism that makes improvement safe is a fixed evaluation set. Anthropic's guidance on evals for agents describes a practice of defining tasks with known correct outcomes, running them on every change, and reading the failures rather than only the average score. Apply that to your calls: keep the past calls you exported, add every correction a dispatcher makes, and rerun the whole set before you change a model, a prompt or a tool description. A model swap that improves three cases and breaks two important ones should not ship.
Also keep permissions in a separate review. Each quarter, or whenever the decision guide changes, confirm that the agent's tool list matches the guide line by line. Permission drift is easier to miss than a bad answer, because nothing visibly fails.
Observability and incident response for field calls
Every proposal should leave a trace that a person can read later. For each call, record the intake, the job records and vault references the agent consulted, the proposal it made, and the dispatcher decision that followed. The OpenTelemetry primer describes the general idea of tracing a request through its steps, and the same logic applies to an agent's chain of reads and proposals.
Write an incident procedure before you need it. If the agent proposes a visit to the wrong address, or if a caller's message appears to have changed its behavior, the response should be to pause the affected tool, preserve the trace, review what the agent read, and correct the records or rules before turning the tool back on. Log access to the vault as well, and review those logs on a schedule.
How Be Human's managed approach handles this
Be Human designs these pieces as one system. A first engagement defines the dispatch role, the job and key-record memory it needs, the tools it can use, the approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations against your own call history, integration maintenance, permission review, incident response and controlled changes when a model or workflow is improved. The agent keeps its identity, memory and operating context under your control, so the decision to expand its role stays with your team.
If you want to find the first task in your dispatch or field operation that is worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI dispatch agent handle lockout calls without a dispatcher? Not safely as an unsupervised system. It can take intake, check job history, propose a technician and a window, and draft the reply, but a dispatcher should confirm proposals until the agent has matched your decisions on the cases that matter most. Any step involving a vehicle, a code or a pricing exception stays with a person.
Should key codes be stored where the agent can read them? No. The agent should see a reference that points to a record in a secured vault, and it should know whether that record exists and who may retrieve it. Keeping the code itself out of memory, prompts and logs means a caller cannot pull it out through conversation, and it limits the damage from any configuration mistake.
Does a stronger model mean the agent needs more access? No. A more capable model may read messy requests more accurately and explain its proposals more clearly, but it does not need new tools to do that. Expanding permissions should be a separate decision, made by a named person after reviewing the decision guide and the evaluation results.
What should we measure before trusting the agent with more call types? Measure how often its proposals match the dispatcher's decision on a fixed set of past calls, and review every disagreement to find whether the cause is the rules, the records or the wording. Add new call types only after the current ones stay correct across model changes and permission reviews.
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.


