At 2:14 a.m., a 911 telecommunicator named Priya has four calls holding, a unit requesting backup on a traffic stop, and a new call coming in from a number that has called twice this month. The caller is speaking low and fast, the background noise is loud, and the words that reach Priya clearly are "he's here" and "I can't." In twelve seconds she will have to decide what kind of call this is and what priority to assign it, because the unit she sends, and how fast she sends it, may be the difference between a domestic violence intervention and a homicide.

That decision is one of hundreds a telecommunicator makes per shift. AI is beginning to touch the work. This lesson maps exactly where AI shows up in 911 dispatch and call triage in 2026, what the tools actually do, and where the hard line sits: the line where a human must own the decision because an error on the other side of that line is not a report correction or a redaction complaint. It is a life.

The field terms used in this lesson: BWC is body-worn camera. CAD is computer-aided dispatch, the software platform dispatchers use to receive, log, track, and coordinate emergency calls and unit deployment. A 911 telecommunicator is the trained professional who answers emergency calls, gathers information, and dispatches the appropriate response resource. Triage is the process of classifying incoming calls by urgency and matching them with the correct response type. Call type is the category assigned to an incident in the CAD system, which determines response protocol, priority level, and which units are dispatched.

The Dispatch Console: What the Telecommunicator Is Actually Managing

To understand what AI assistance means in dispatch, it helps to understand what a 911 telecommunicator is managing without it. The workload is not merely high volume. It is multi-channel, multi-task, and time-compressed in a way that is qualitatively different from almost any other professional environment.

A telecommunicator on a busy shift is simultaneously: answering incoming calls, gathering incident information, entering data into the CAD system while the caller is still speaking, monitoring radio traffic from active units, coordinating multiple active incidents at different priority levels, tracking unit availability across geographic zones, relaying updated information to units in the field, and responding to requests from units who need additional information or backup. Each of these streams runs concurrently. The CAD entry for a new call is being made while a unit is asking for a plate check on the radio while a supervisor is updating the status of a priority-one response.

Staffing in public safety answering points (PSAPs, the dispatch centers that answer 911 calls) has been chronically short for years, and call volume has not decreased. The practical result is that telecommunicators are managing more concurrent tasks per shift than human cognitive load research suggests is sustainable without error. The question AI addresses is whether automation can reduce the mechanical load on the operator, freeing attention for the judgment tasks that require a human: assessment, decision, and direction.

Where AI Shows Up in the Dispatch Workflow

AI-assisted dispatch tools are not replacing the telecommunicator. In 2026, they are augmenting specific tasks in the call-handling workflow. The augmentation falls into three categories: transcription assistance, translation support, and classification suggestion.

Transcription Assistance

Real-time transcription of calls is one of the clearest value applications. As a caller speaks, an AI transcription system converts the audio to text in near-real-time, providing the telecommunicator with a running text record of the call alongside the audio. The operational value is in the CAD entry: rather than typing while listening, the telecommunicator can review the transcribed text and confirm or edit it, reducing the dual-task load of simultaneous listening and typing.

The limitations are predictable and should be understood. Transcription accuracy degrades sharply under conditions common in 911 calls: background noise, distressed speech, accented speech, low-volume calls from callers who cannot speak loudly, and calls where the caller is whispering or crying. A call from a domestic violence victim who cannot speak above a whisper is precisely the call where transcription assistance is least reliable and human judgment is most critical. The telecommunicator using transcription assistance must understand this inverse relationship: the calls where the tool works best are the clear, well-articulated calls where the telecommunicator had less difficulty anyway. The calls where the tool is most likely to err are the difficult calls where the stakes are highest.

Translation Support

Real-time language translation is an area where AI has made the access to assistance more immediate. 911 centers have long used human interpreter services for non-English calls, but connecting a three-way call with an interpreter adds time. AI translation tools can provide a faster initial translation, allowing the telecommunicator to gather basic triage information while the interpreter connection is being established, or in some deployments providing a running translation for lower-complexity calls.

The caveat here is significant and deserves to be stated plainly: machine translation errors in a 911 context carry life-safety consequences. A translation that converts "he has a knife" into "he has something" or drops a critical qualifier does not just produce a poor customer experience. It produces an under-resourced response to a dangerous incident. Any agency deploying AI translation in dispatch must have a policy for when human interpreter services are mandatory, and that policy must be protective in its defaults. The convenience of AI translation cannot override the accuracy requirement for life-safety calls.

Classification Suggestion

The third application is call classification, also called call-type suggestion: the AI system analyzes the incoming call text, audio, or metadata and suggests a call type and priority level to the telecommunicator. The suggestion appears as a prompt in the CAD interface, and the telecommunicator confirms, overrides, or modifies it.

This is where the governance line in dispatch is most clearly drawn. Classification suggestion assists; it does not decide. The telecommunicator owns the priority. The reason this line is drawn here, and not at transcription or translation, is the asymmetry of consequence. A transcription error degrades record quality. A priority misclassification can mean the wrong unit type arrives, arrives too slowly, or arrives without adequate preparation for the threat environment. In the scenario at the top of this lesson, a system that classifies "he's here, I can't" as a welfare check rather than a domestic violence emergency has assigned the wrong resource to a potentially lethal situation.

In dispatch, human ownership of the priority decision is not a preference. It is the design requirement, because the cost of an automated misclassification is measured in lives.

The Classification Problem: Why AI Gets Call Type Wrong

To use AI call classification assistance responsibly, telecommunicators need to understand how the classification models work and where they fail. The failure modes are not random; they are predictable, and knowing the patterns makes the human override reflex more reliable.

AI classification models are trained on historical dispatch data: past calls, their recorded types, their outcomes. The model learns to match incoming call patterns to call types based on that history. The failure modes cluster in three areas.

First, underrepresented incident types. If the training data contains few examples of a particular call pattern, the model will classify calls matching that pattern less accurately. This is an issue for emerging or unusual incident types, for calls that do not fit standard patterns, and for calls from demographic groups or geographic areas that were underrepresented in the training data. A model trained primarily on urban dispatch data may perform less accurately for rural call patterns. A model trained on calls from a specific region may perform less accurately when the agency it is deployed in has a different demographic mix of callers. These are not theoretical concerns; they are the documented behavior of classification models in production.

Second, ambiguous or partial information. Real 911 calls frequently begin with incomplete information. The caller is distressed, does not know their location, is using a cell phone with poor signal, or is reporting a situation they themselves do not fully understand. AI classification models work best with complete, clear input and perform less reliably when the information is fragmentary. A model suggesting a call type at second 15 of a call that will be better understood at second 45 may be wrong in a way that has already influenced the telecommunicator's response preparation. Early suggestion can anchor the telecommunicator to a wrong classification before adequate information has arrived.

Third, the caller's word choice versus the underlying event. Call classification models are partially text-based, and caller word choice does not always match incident category. A caller reporting a "family argument" may be reporting a domestic violence situation with assault. A caller reporting a "suspicious person" may be reporting an active medical emergency. A caller who says "I think someone is trying to break in" may be reporting an active burglary in progress. The distance between the caller's framing and the actual incident type is where AI classification is most likely to err in a consequential direction. The model classifies on words; the telecommunicator must assess intent, tone, background noise, and context simultaneously.

The Telecommunicator's Override Skill

Given the failure modes above, the most practically important skill a telecommunicator develops with AI call classification assistance is not learning to accept the suggestion. It is learning when and how to override it confidently. This is a skill that needs to be taught explicitly, practiced in training, and reinforced in QA review, because the default human tendency with an authoritative-looking automated suggestion is to accept it rather than contradict it.

Researchers who study human-automation interaction have documented a failure mode called automation bias: when a human and an automated system disagree about a classification, the human tends to defer to the automated system even when the human's judgment is correct. In low-stakes settings this is mildly inconvenient. In a 911 dispatch center, automation bias on a misclassified call can mean a welfare check unit responds to an active domestic violence situation without the appropriate resources.

The professional antidote to automation bias in dispatch is a standing override protocol: every AI classification suggestion is a starting point, not a conclusion. The telecommunicator confirms it against their own assessment of the call before it is entered into the CAD. This confirmation is not skepticism for its own sake; it is the professional standard for a role where the human judgment is both legally required and operationally irreplaceable. The CAD entry is the record that determines the response. The person making that entry owns the entry.

What should trigger an override? A telecommunicator should treat the AI suggestion as questionable and apply active reassessment when: the caller's tone or word choice does not match the suggested category; background sounds are inconsistent with the call type (sounds of a struggle on a reported welfare check, sounds of traffic on a reported home invasion); the caller's level of distress is disproportionate to the suggested priority; the caller is known to the center from previous calls with a different pattern; or the call is in the first thirty seconds and information is still incomplete. None of these triggers require certainty about the correct classification. They require slowing down the acceptance of the suggested classification until the information supports it.

The CAD Entry as Evidence

Everything entered into the CAD system is a record. It is timestamped, logged, retained, and subject to subpoena. In a use-of-force investigation, a wrongful death lawsuit, or a post-incident review, the CAD record is one of the first documents examined. The call type, the priority level, the time of dispatch, the unit assigned, and the transcript of the call are all in that record. The CAD entry made by the telecommunicator is evidence in the same way a police report is evidence.

This means that the governance standard for AI assistance in CAD entry is analogous to the governance standard for AI assistance in report writing: the AI suggests, the human confirms, and the human is accountable for what enters the record. A CAD entry that reflects an AI suggestion the telecommunicator accepted without active assessment is a CAD entry the telecommunicator owns. If the classification was wrong and the response was inadequate, the question in the post-incident review is not "what did the AI suggest?" It is "what did the telecommunicator enter, and why?"

This framing is not meant to create anxiety about AI use in dispatch. It is meant to establish the professional standard clearly: AI assistance in dispatch is a tool the telecommunicator uses, not a system that makes the decisions. That clarity is what allows the tool to be used well and the professional to be accountable for the outcomes. The telecommunicator who can explain every classification decision is the one who is using the tool correctly.

AI-Assisted Post-Call Documentation and Handoff

Beyond real-time assistance, AI is also beginning to assist with the documentation that follows call closure: the CAD narrative, the incident summary, and the information handed off to responding units and detectives. These downstream applications carry the same evidentiary weight and the same verification requirements as the real-time tools.

AI-assisted CAD narratives, summaries of what the caller reported and what was dispatched, are beginning to appear in some platforms. The same failure modes as in police report drafting apply: the model may soften the caller's reported threat level, may fill a gap in the call transcript with an inferred detail, or may produce a summary that does not capture the full urgency of the caller's communication. A unit responding to an incident that was summarized as "unknown disturbance, caller was disconnected" rather than "caller stated 'he's here, I can't' and was disconnected mid-call" is operating with critically different information about the threat environment.

The verification standard for post-call documentation is: the telecommunicator or the reviewing supervisor confirms that the AI-generated narrative or summary accurately represents the call before it is finalized in the CAD record. For complex, high-stakes, or ambiguous calls, the standard should be heightened: the record of the actual call audio and the record in the CAD system should be checked against each other before the incident is marked closed.

Looking ahead within the program: the next lesson covers AI in records, redaction, and public-records response. The CAD record itself, including call logs, transcripts, and incident summaries, is subject to public-records statutes and may be released in response to a public-records request. The way AI-assisted documentation is created determines the quality and the privacy implications of what gets released. That connection between dispatch documentation and public-records response is a thread worth holding as you move through this chapter.

Key Takeaways

• AI assists 911 telecommunicators in three areas: real-time transcription, language translation support, and call classification suggestion. In each case, the AI provides a starting point or a data point. The telecommunicator confirms, overrides, and owns the decision.
• Call classification is the application where human ownership of the final decision is most critical. A misclassified call can mean the wrong resource type, the wrong priority, and inadequate preparation for the threat environment. The cost of an automated misclassification in dispatch is measured in safety outcomes, not just record accuracy.
• AI classification models fail predictably in three patterns: underrepresented incident types, ambiguous or partial information at the start of the call, and the gap between the caller's word choice and the actual incident category. Knowing these failure patterns makes the override reflex faster and more accurate.
• Automation bias, the tendency to defer to an automated system even when human judgment is correct, is a documented failure mode in human-automation interaction. The antidote in dispatch is an explicit standing protocol: every AI suggestion is a starting point, confirmed against the telecommunicator's own assessment before it enters the CAD.
• Everything entered in the CAD is a timestamped, subpoena-able record. The telecommunicator owns every entry, regardless of what AI suggested. The professional standard is the same as for report writing: AI assists, human is accountable.
• AI-assisted post-call documentation, including incident summaries and narrative handoffs, carries the same evidentiary weight and the same verification requirement as real-time call handling. A summary that softens the caller's reported threat level is a record that may produce an underprepared response.
• The CAD record, including AI-assisted call logs and transcripts, is subject to public-records statutes. The quality and privacy implications of AI-assisted documentation in dispatch flow directly into the records and public-records release processes covered in the next lesson.