It is 2:17 in the morning and console four at the regional 9-1-1 communications center is running three calls simultaneously. The telecommunicator, Dispatcher Ramos, has one caller who is crying and switching between English and Spanish, a second caller who is shouting over a barking dog, and a third caller on hold who called about a possible gas leak. The computer-aided dispatch (CAD) system needs a clean, concise entry for the crying caller's incident within seconds so the patrol unit can read it on the mobile data terminal while rolling. What goes into that CAD entry is how the officer will approach the scene.

What the CAD Entry Actually Does

Every call that enters the 9-1-1 system produces a record in the computer-aided dispatch (CAD) system, which is the platform that creates incident records, routes calls to dispatchers, and transmits information to responding units. The CAD entry is not an administrative footnote. It is a real-time operational document that officers, deputies, firefighters, and paramedics read on their mobile data terminals while driving to the scene. The words in the remarks field shape how they approach, what resources they request, and what information they announce over the radio before arriving.

A clean CAD entry for a domestic disturbance call might read: "Female caller states male subject threw a lamp, left the apartment on foot, wearing red hoodie. No weapons mentioned. Children in home, ages 5 and 8. Caller is mother, locked in bedroom." That entry takes fifteen seconds to read at 60 miles per hour and tells the responding officer the nature of the contact, the direction of travel, the description, the presence of vulnerable parties, and the caller's location. Those six pieces of information can mean the difference between a safe approach and a surprise.

The challenge is producing that entry in real time while managing an active call from a distressed caller who may be whispering, crying, intoxicated, limited-English proficient, or panicking and whose account of the incident is nonlinear, fragmented, and emotionally driven. The dispatcher simultaneously has to gather information, verify the address through the automatic number identification / automatic location identification (ANI/ALI) system, assess the call type, assign priority, coordinate units, and monitor the radio. The cognitive load is enormous, and the margin for error is measured in seconds.

This is where AI-assisted call summarization enters the dispatch workflow. The premise is straightforward: an AI system that can listen to a call in real time, transcribe it, identify key facts, and surface a structured draft summary reduces the cognitive burden on the telecommunicator and speeds the production of an accurate, usable CAD entry. The technology is in active deployment in some centers in 2026. The value is real. So is the risk.

How AI Call Summarization Actually Works

AI-assisted call summarization in a dispatch context involves several capabilities that are distinct but often marketed together. Understanding each one separately is important for a telecommunicator who needs to know what the tool can do, what it cannot do, and where it is most likely to fail.

Real-Time Transcription

The first capability is transcription: converting the audio of the call to text. Modern automatic speech recognition (ASR) systems trained on broad audio datasets perform well in clean, clear audio environments. They perform less well when audio quality is poor, when the caller has a strong regional accent or a non-native-English accent, when there is background noise such as a dog barking, traffic, wind, or a television, when the caller is crying or shouting, or when the call involves code-switching between languages. In a 9-1-1 context, those conditions are not edge cases. They are common. A transcription system that works at 96% accuracy in a quiet audio environment may drop to 82% accuracy when there is a screaming child in the background, and those errors cluster around the most emotionally charged moments in the call, which are often also the most informationally important moments.

This matters because everything downstream of transcription inherits its errors. If the transcription converts "he has a gun" to "he has a bun," or "she's not breathing" to "she's not leaving," the summarization model that reads that transcript may produce a fundamentally different call type and priority recommendation than the actual call warranted.

Key Fact Extraction

The second capability is extraction: identifying from the transcript the key facts that belong in a CAD entry. Well-defined facts include the address or location, the nature of the incident, the number and description of parties involved, the presence of weapons, injuries, or medical emergencies, and any special hazards. Extraction models trained on dispatch data can identify these categories with reasonable accuracy in well-structured calls where the caller provides information in an expected order.

Dispatch calls rarely follow that structure. A caller describing a traffic accident will give the address, then loop back to describe what they saw, then mention that the driver looked "real bad" before specifying that they are not moving, then remember to mention a child in the back seat. The extraction model must piece together the narrative from a fragmented, out-of-order account and decide which statements are actionable facts and which are emotional expression. When the model is uncertain, it fills in. That gap-fill behavior, producing a plausible-sounding detail that the caller did not actually provide, is the hallucination risk in call summarization.

Draft Summary Generation

The third capability is draft generation: producing a structured summary in the format used by the agency's CAD system. Some systems produce a plain-text remarks field; others produce structured fields for location, call type, narrative, and unit disposition. The AI generates text in the expected format based on the extracted facts and its training on prior dispatch entries for similar call types. The draft is a starting point for the telecommunicator, not a completed entry.

The failure modes at the generation stage mirror those in any generative AI context. The model may generate a call type based on the pattern of the call rather than its specific content. A caller describing a neighbor who is "acting crazy and throwing things" may produce a draft that classifies the call as a mental health welfare check when it should be a domestic disturbance, or vice versa, depending on what patterns the model has learned. The classification matters because call type determines priority tier, which determines response time, which determines whether one unit goes or two, whether fire is dispatched simultaneously, and how the CAD record is coded for statistical tracking and after-action review.

The AI draft summary is a starting position, not a finished entry. The telecommunicator reads it, corrects it, and owns what goes into the CAD system. Everything the officer reads on that mobile terminal comes from a human decision to submit.

What Clean CAD Entries Look Like and Where AI Helps Produce Them

A clean CAD entry, for purposes of this lesson, means an entry that is accurate, sufficient, and usable. Accurate means it reflects what the caller actually said, not what a similar call usually contains. Sufficient means it includes all the information the responding unit needs to approach safely and deliver appropriate service. Usable means it is formatted in a way that can be read quickly at speed.

AI-assisted summarization can help with the sufficiency and usability dimensions. A well-functioning summarization tool can identify that the caller mentioned a child in the apartment and surface that fact in the draft even if the telecommunicator was focused on obtaining the address at the moment the caller mentioned it. It can produce consistent formatting across entries for similar call types, reducing the variation that comes from different dispatchers writing entries in different styles. It can flag that a caller mentioned a specific location detail, such as "apartment on the second floor, unit 2B, the door is unlocked," that might be buried in a long transcript and easy to miss under time pressure.

These are genuine productivity and quality benefits. A center that processes 800 calls per shift and relies on manual entry for every CAD record is asking a lot of each telecommunicator across an eight-hour stretch where cognitive fatigue sets in by hour three. A tool that pre-fills a draft that the dispatcher reviews and corrects rather than types from scratch can meaningfully reduce that load.

The accuracy dimension is where the human judgment is irreplaceable. The AI draft may be accurate, or it may contain a gap-fill detail, a misclassification, or a transcription error that carried through to the summary. Only the telecommunicator who heard the call can evaluate whether the draft reflects what the caller actually said. The tool helps with production; the human provides the truthfulness check.

The Gap-Fill Risk: What the AI Adds That the Caller Did Not Say

The most operationally dangerous failure mode in AI-assisted call summarization is not a transcription error or an incorrect call-type classification. It is the gap-fill: a detail added to the summary that sounds plausible for the call type but that the caller never actually provided.

This failure mode occurs because the AI models that generate dispatch summaries are trained on large datasets of prior calls and CAD entries. The model learns patterns: when someone calls about a domestic incident at 2 a.m., the entry often includes a description of the subject, a disposition note, and a weapons status. When a caller in this call has provided the address and the nature of the incident but has not mentioned weapons status, the model may generate "no weapons mentioned" as a default. That is a gap-fill that may or may not reflect reality.

In some cases, the gap-fill is neutral. "No weapons mentioned" is a reasonable default if the caller truly gave no indication of weapons. In other cases, the gap-fill is dangerous. If the caller's fragmented account contained an ambiguous statement, such as "he's got his thing," which the telecommunicator heard as a potential reference to a weapon, but the AI transcription did not capture the ambiguity and the model defaulted to "no weapons mentioned," the CAD entry now actively contradicts the dispatcher's professional judgment about what the call may have contained. The officer reading the entry may approach the scene with different caution than the one the dispatcher believed was warranted.

A related gap-fill risk occurs with location information. ANI/ALI provides the callback number and location of record for the calling device, but cell phone location accuracy varies widely. A caller who says "I'm at the park on Jefferson" while the ANI/ALI shows an address two blocks away creates a discrepancy that the AI may resolve by defaulting to the ANI/ALI address. That default may be correct or incorrect depending on whether the caller is at the registered location of their carrier or at their actual position. The telecommunicator's judgment about which location to use, based on the caller's description and any clarifying questions, must be what the CAD entry reflects, not the model's default.

Gap-fill is not malicious. The model is not trying to deceive anyone. It is producing the most statistically probable completion of the entry given its training. That is the mechanism, and that mechanism does not know the difference between a detail the caller actually provided and a detail that would typically appear in this call type. The telecommunicator does know the difference, which is exactly why the review step cannot be skipped.

The Review Discipline: Owning Every Word in the CAD Entry

The verification habit for AI-assisted call summarization is straightforward in concept and requires discipline in practice: before submitting a CAD entry, the telecommunicator reviews the AI draft against the call itself and corrects every detail that does not accurately reflect what the caller said.

This requires answering three questions for each key fact in the draft:

Did the caller actually say this? Every factual claim in the draft should trace back to something the caller said. If the draft says "subject is a white male, approximately 6 feet tall, wearing a red jacket," the telecommunicator should be able to confirm that each element, the race, the gender, the height, the description, came from the caller's account. If the caller said "some guy" and nothing more, the draft should not contain a physical description. A description that was not provided is a gap-fill, and a gap-fill in a subject description can send an officer looking for the wrong person.

Does the call type match what I heard? Call type classification is one of the areas where AI pattern-matching is most likely to diverge from dispatcher judgment. The telecommunicator who heard the call has access to paralinguistic information, the tone of voice, the sounds in the background, the pacing of the caller's responses, that the transcription and model cannot fully capture. If the AI has classified a call as a disturbance when the dispatcher assesses it as a potential in-progress crime, the call type needs to be changed. The call type determines the response, and an incorrect call type is a priority error.

Is there anything from the call that is not in the draft? AI extraction models are designed to identify the most salient facts, but the telecommunicator who heard the full call may have information that the model did not flag: a specific threat made in a tone of voice the transcription missed, a mention of a specific location feature that the model did not flag as operationally relevant, a detail about vulnerability that emerged conversationally rather than as a direct statement. The review step is not just about correcting the draft. It is about completing it.

The physical workflow for this review depends on the interface the CAD system and the AI tool provide. In some implementations, the draft summary appears in a sidebar or preview pane that the dispatcher can edit before committing the entry. In others, the AI suggestions are surfaced as suggested text fields that the dispatcher accepts or modifies. In either case, the principle is the same: the dispatcher is the author of the entry. The AI is a drafting assistant. The entry that reaches the officer's mobile data terminal is the dispatcher's account of what the caller said, reviewed and submitted by a human who heard the call and vouches for its accuracy.

Under time pressure, the temptation is to accept the draft and move on. That temptation is understandable. It is also the failure mode. A CAD entry that goes to an officer without a review is not an AI-assisted entry. It is an AI-determined entry, and the telecommunicator is submitting an account they did not verify as their own. When the entry contains an error that affects the response, the question of who wrote the entry has a clear answer: it was submitted from the dispatcher's console, under their login, and carries their professional authority.

Keeping the Record Tied to What the Caller Said

The principle that runs through every aspect of AI-assisted call summarization is the same principle that runs through every other AI-assisted function in public safety: the AI tool assists production, and the human professional ties the record to what actually happened. In the dispatch context, "what actually happened" is what the caller said, because the dispatcher was not at the scene. Everything in the CAD entry is secondhand from a distressed caller on a phone, mediated by a transcription system and a summarization model. The dispatcher's job is to receive that account, clarify where needed, and produce a record that accurately represents what the caller reported.

The value of AI assistance in this process is that it helps produce a more complete, consistently formatted draft more quickly than a dispatcher working entirely from memory and manual entry. The risk is that the draft may contain errors, gaps, or misclassifications that require correction before the entry is usable. The discipline is reviewing the draft against the call, correcting what needs correction, and submitting an entry the dispatcher is willing to stand behind as their account of what the caller reported.

An agency that deploys AI-assisted call summarization should ensure that the system retains the original audio, the transcription, and the submitted CAD entry as a set, so that any question about the gap between what was said and what was entered can be investigated. That audit trail protects the telecommunicator who followed good practice, and it surfaces the cases where the AI introduced errors that went uncorrected. Both uses are legitimate, and both depend on keeping the record complete.

Telecommunicators should also be aware that AI summarization tools vary in their design and training. A tool trained primarily on urban call patterns may perform differently on rural calls with different vocabulary, address formats, and call types. A tool trained on English-only calls may handle code-switching less reliably. Understanding how the specific tool deployed at your center performs under your center's specific call conditions is part of the professional competence required to use it responsibly.

Finally, if you have identified a pattern in which the AI consistently produces a specific type of error for a specific call type or caller population, documenting and reporting that pattern is a professional responsibility. A tool that systematically misclassifies calls from a specific area of the service district, or consistently fails to capture information from calls involving non-native-English speakers, has an accuracy problem with equity implications. The telecommunicators who use the tool and see its outputs day by day are the most qualified people to identify those patterns. Reporting them is how the agency can address them.

Key Takeaways

• The CAD entry is an operational document that shapes how officers approach a scene. Errors in that entry carry real-world consequences for officer safety, public safety, and call documentation.
• AI-assisted call summarization works through transcription, key-fact extraction, and draft generation. Each stage has failure modes, and errors introduced at the transcription stage propagate through the entire pipeline.
• The most dangerous failure mode is the gap-fill: a detail added to the draft that sounds plausible for the call type but that the caller never actually provided. Gap-fills in subject descriptions, location information, and weapons status can directly affect response decisions.
• The review discipline requires three checks for each key fact: did the caller say this, does the call type match the dispatcher's assessment, and is there anything from the call that is missing from the draft.
• The telecommunicator is the author of every CAD entry submitted from their console. The AI draft is a starting position. The submitted entry is the dispatcher's account, reviewed and vouched for by a human who heard the call.
• ANI/ALI provides the number and location of record for the calling device. When the caller's stated location differs from the ANI/ALI data, the dispatcher's judgment about which location to use must be what the entry reflects, not the model's default.
• Agencies should retain the audio, the transcription, and the submitted CAD entry as a complete set to support quality assurance, investigation of AI errors, and documentation of the review process.
• Telecommunicators who identify consistent error patterns in AI call summarization tools, especially those affecting specific call types or caller populations, should document and report those patterns so the agency can address systematic accuracy problems.