At 2:47 on a Tuesday morning, telecommunicator Adriana Reyes was twelve calls deep into a shift that had never fully slowed down. The PSAP (public safety answering point, the communications center that receives emergency calls) had deployed AI-assisted transcription and triage support three months earlier, and Adriana had become good at using it. She had learned the rhythm: the AI transcribed incoming audio in near real-time, surfaced a suggested call type and priority tier alongside the transcript, and auto-populated the incident fields in the CAD (computer-aided dispatch system, the platform that routes calls to units and tracks every action from first ring to final close). She had learned to scan the suggestion, trust her ear, and override when something felt wrong. On most calls, the AI saved her six to eight seconds of keying. On a busy shift, those seconds were real. Then came call 247. A woman's voice, low and controlled, said she needed the police at her address. The AI classified it: non-emergency civil disturbance, Priority 4. Adriana had heard something different. The flatness of the woman's voice. The way she said "I need the police" without the word "please," without affect, without the agitation that usually accompanied a domestic noise complaint. Adriana reclassified it: Priority 1, domestic violence with possible weapon. Units arrived in four minutes. They found exactly what Adriana's ear had heard in the flatness.

The Workflow in Full: What AI-Assisted Call Handling Actually Looks Like

This lesson maps the end-to-end workflow of AI-assisted call handling: from the moment a call enters the PSAP through transcription, triage classification, CAD entry, unit assignment, and call closure documentation. Each stage has an AI role, and each stage has a human role. Understanding where they split, and why that split is non-negotiable at specific points, is the core competency this lesson builds.

The workflow is not complicated in outline. A call arrives. Audio is captured and transcribed. The transcript and audio together feed an AI classification engine that produces a suggested call type and priority tier. The telecommunicator reviews the suggestion against what they heard and what the CAD history for that address shows, then accepts or overrides. A CAD entry is generated with the accepted call type. Units are dispatched. The call is documented as it develops, with the AI summarizing relevant new information at each update. At closure, the AI produces a draft incident summary. The telecommunicator reviews and adopts the summary. The record closes.

In that outline, the AI is doing real work: transcription, classification suggestion, CAD field population, update summarization, and closure drafting. It is also making recommendations at every step that, if accepted without review, can be wrong in ways that cost lives. Adriana Reyes's call is not the exception. It is the test that every AI-assisted workflow must be designed to pass.

AI assists the workflow at every stage. The telecommunicator commands every decision that affects human safety. Those are not the same role, and the workflow must keep them separate.

Why the End-to-End Frame Matters

Public safety agencies often evaluate AI tools stage by stage: "Does the transcription work? Does the classification suggest the right call type most of the time? Does the CAD auto-population save keystrokes?" These are legitimate questions. But they are the wrong frame for evaluating whether the workflow is safe, because a workflow is a chain. A problem at one stage propagates through every downstream stage. A misclassified call at intake does not stay a misclassified CAD entry. It becomes a misclassified dispatch priority. It becomes units responding at the wrong speed. It becomes time lost that cannot be recovered.

The end-to-end frame asks a different question: at every handoff in this chain, is the human in command of the decision that matters? Not "did the AI get it right most of the time" but "when the AI gets it wrong, does the workflow surface that error before it becomes a dispatch?" That is the standard this lesson applies.

Stage One: Intake and Transcription

The first stage of the workflow is intake: the call comes in, the telecommunicator answers, the audio is captured, and transcription begins. AI-assisted transcription converts the caller's speech to text in near real-time, giving the telecommunicator a running text record of the call alongside the audio they are actively monitoring.

Transcription is the stage where AI is doing the most technically reliable work. Transcription accuracy on clear audio in modern systems is high, and errors, when they occur, tend to be words rather than meaning. A caller who says "he has a knife" is not likely to be transcribed as saying something that changes the call type. But transcription is not perfect, and it has known failure modes in the 911 environment.

Transcription Failure Modes at Intake

The first failure mode is accent and dialect variation. Transcription systems trained predominantly on one dialect will perform less accurately on others. A caller with a strong regional accent, a non-native English speaker under extreme stress, or a caller using community-specific vocabulary will generate more transcription errors. The telecommunicator's ear remains the primary reliability check for any call where the transcript and the audio do not match.

The second failure mode is emotional state degradation. A caller who is screaming, crying, whispering from fear, or hyperventilating produces audio that degrades transcription accuracy significantly. These are precisely the calls where accurate classification is most important. They are also the calls where the transcript is least reliable. The telecommunicator must weight the audio over the transcript whenever emotional state makes the transcript unreliable.

The third failure mode is environmental noise. A call placed from a running vehicle, a crowded environment, outdoors in wind, or near a TV or radio will produce noise that the transcription system may misinterpret as speech. The telecommunicator must listen through the noise to the caller's actual words.

The critical discipline at this stage is simple: the telecommunicator hears the call. The transcript is a tool to help organize and record what was said. It is not a substitute for listening, and it is never more reliable than the telecommunicator's own ear on the specific call in progress.

What the AI Captures in the Transcript

Beyond words, modern AI-assisted intake systems are beginning to analyze paralinguistic signals: the pace of speech, pitch variation, pauses, and vocal stress markers. These signals can inform urgency assessment independently of the words spoken. Adriana Reyes was using her own paralinguistic read when she heard the flatness of the caller's voice. Some AI systems now surface this as a separate urgency indicator alongside the transcript. When they do, the telecommunicator should treat it as additional signal, not as a classification decision. The decision remains theirs.

The transcript, once generated, flows directly into the classification engine and becomes the primary text input for the suggested call type. What the transcription gets wrong, the classification will also get wrong. A transcript error that changes the apparent content of a call will produce a classification that is wrong for the right reasons given the transcript, and right for the wrong call. That is the gap the telecommunicator closes by listening.

Stage Two: Triage and Classification

The second stage is triage and classification: the AI analyzes the transcript, the caller's address history in the CAD, the time of day, and the pattern of language against a trained model of call types and priority tiers, then produces a suggested classification. This is where the workflow reaches its most consequential AI contribution, and its most consequential human responsibility.

Classification suggestion is genuinely useful. On high-volume shifts in a busy PSAP, the volume of calls exceeds what any human can classify from scratch with zero assistance. The AI suggestion gives the telecommunicator a starting point, a pre-positioned hypothesis about what the call is, that they then confirm, adjust, or reject based on what they heard. In studies of AI-assisted dispatch systems, classification suggestions are accepted without modification on a significant majority of routine calls. That is not a problem. That is the efficiency gain. The problem would be if the acceptance rate were high across all calls, including the calls where the AI is wrong.

How the Classification Engine Can Fail

The classification engine fails in predictable patterns. Understanding them is the foundation of the telecommunicator's oversight role.

The first pattern is controlled-language misclassification. Callers who are speaking in controlled, measured language because they cannot say what they mean (because the threat is in the room, because they are afraid, because they have been coached not to escalate) produce transcripts that describe benign or ambiguous situations. The AI, reading the words, classifies the call at a lower priority. The telecommunicator, hearing the voice, may classify it far higher. Adriana Reyes's call was exactly this pattern. The words said "I need the police at my address." The voice said "I am in danger and cannot say more." The AI read the words. Adriana heard the voice.

The second pattern is familiarity bias. When a caller's address has a high volume of previous calls of a particular type, the classification engine may weight that history heavily in its suggestion. A household that has generated multiple noise complaints may have a genuine domestic violence call classified as another noise complaint because the history pulls the classification toward the established pattern. The telecommunicator must treat the current call on its own merits, not on the address's history.

The third pattern is incomplete call information. Many callers do not give the AI time to gather the signals it needs for an accurate classification. They state an address, say one sentence, and then stop talking to manage the situation they are in. The transcript at that point is thin. The AI classifies from thin information. The telecommunicator must classify from thin information plus everything else they can hear: tone, background, what is not said.

The fourth pattern is novel call types. AI classification systems are trained on historical call data. Calls that do not match historical patterns well, because they involve a new situation, an unusual combination of elements, or an event type the training data does not cover, will be classified less reliably. The telecommunicator's professional judgment fills this gap.

The Override Is Not a Failure

In many agencies that have deployed AI-assisted classification, there is an unspoken pressure on telecommunicators to accept the AI suggestion unless they have a clear, articulable reason to override. This pressure is backwards. The override is not a failure of the workflow. The override is the workflow doing exactly what it was designed to do: putting a trained human judgment over a statistical model at the moment when that judgment matters most.

An override should be logged with a brief rationale: "override: caller voice indicators inconsistent with Priority 4 classification." That logging is not bureaucratic. It is the data that the agency uses to evaluate whether the classification engine is performing well, where its failure modes cluster, and whether the training data needs adjustment. An override without documentation is a correction that the agency cannot learn from.

A good workflow makes overrides easy. The suggested classification appears as a default that the telecommunicator can accept with one action or change with two. The override is not a multi-step exception process. It is a standard workflow action, because the standard workflow must include human judgment at every classification decision, not only at the ones where the suggestion looks obviously wrong.

Stage Three: CAD Entry and Dispatch

Once the call type and priority are confirmed by the telecommunicator, the workflow moves to CAD entry and dispatch. This is where AI-assisted workflows deliver significant operational speed gains. Auto-population of CAD fields from the confirmed call type, address, and transcript means the telecommunicator is not typing while simultaneously listening and deciding. The CAD system can pre-position units based on the call type before the telecommunicator has finished the entry. In high-volume PSAPs, this compresses dispatch-to-response time meaningfully.

The CAD entry is also a legal record. Every entry in the CAD is timestamped, logged, and available for review in any subsequent investigation, lawsuit, administrative proceeding, or public records request. The auto-populated fields that the AI generates are part of that record, and the telecommunicator's confirmation or override is part of that record. This is the same principle that applies to AI-assisted report writing: the AI may draft the entry, but the telecommunicator adopts it by confirming, and adoption makes it their entry, not the AI's.

The CAD Entry as Evidence

Investigators reviewing a use-of-force incident, a delayed response, a missed call, or a tactical failure will pull the CAD record first. They will look at the initial call type, the priority assigned, the timestamp of dispatch, the unit response time, and every update entered during the call. They will also ask: was the initial call type correct? Was the priority appropriate? If an officer arrives at what was classified as a noise complaint and encounters a domestic violence situation, the CAD record will be examined for what the PSAP knew, when they knew it, and what classification was assigned.

An AI-assisted CAD entry that was accepted without review, and that turns out to have been wrong, is not a technical failure of the system. It is a documentation failure of the telecommunicator who adopted a wrong entry without reviewing it. That distinction matters in litigation, in administrative review, and in the telecommunicator's deposition. The question is not "what did the AI suggest." The question is "what did you enter, and did you verify it before entering it."

Real-Time Updates and the Evolving Call

Once a call is active, the AI-assisted workflow continues. As the incident develops, additional information comes in: officers on scene radio updates, additional callers, witnesses, security camera feeds. AI-assisted systems can synthesize these incoming data streams and propose CAD updates, flagging new information that changes the picture. The telecommunicator reviews proposed updates, accepts or revises them, and maintains the CAD record as the incident evolves.

The critical discipline here is that every proposed update is a proposal, not a decision. The telecommunicator decides what enters the official record. They decide whether the new information changes the tactical picture. They decide when to escalate the priority, when to add units, and when the information warrants notifying a supervisor. None of these decisions are AI decisions. They are the telecommunicator's professional judgment, informed by AI-synthesized information.

One specific risk in AI-assisted real-time updates is the summary compression error. As more information comes in, the AI summarizes the developing situation. In that summary, a detail that was uncertain may become stated as certain, a detail that was one possible interpretation may become the only interpretation, and a detail that was attributed to an unverified source may become stripped of its attribution. The telecommunicator reading the AI summary must maintain awareness that it is a summary, that summaries compress and occasionally distort, and that the underlying communications are the record, not the summary.

Stage Four: Closure Documentation

When the incident closes, the AI workflow produces a draft closure summary from the CAD record, the radio traffic, and any transcribed communications during the call. This summary is the AI-assisted equivalent of the AI-drafted police report: a draft that needs review and adoption before it becomes the official record.

Closure documentation in a PSAP has its own evidentiary weight. It is the record of what happened on the call, what the PSAP's role was, what decisions were made, and what the outcome was. In any subsequent review of a serious incident, the PSAP closure documentation will be examined alongside the officer's report, the body-camera footage, and the CAD audit log.

Adopting the Closure Summary

The telecommunicator who adopts the closure summary is doing what the patrol officer does when they adopt the AI-drafted narrative. They are taking authorship of that record. The summary they sign off on is their account, not the AI's draft. If the summary mischaracterizes the call, overstates what was communicated, undersells the indicators they flagged, or omits the override they made and why, those errors are in the official record under their adoption.

A good closure review checks four things. First, does the summary accurately reflect the initial classification and any overrides, including the rationale? Second, does it accurately reflect the timestamps, particularly the time from call receipt to dispatch? Third, does it accurately represent what information the PSAP had and when they had it? Fourth, does it accurately capture any significant developments, escalations, or communications that affected the tactical response?

If the answer to any of these is no, the summary needs correction before adoption. The same principle that governs the officer's report governs the telecommunicator's closure summary: the official record must accurately reflect what actually happened. The draft is a starting point. The adopted summary is the account.

The Disclosure Thread in Dispatch Documentation

As AI-assisted dispatch documentation becomes standard, agencies are beginning to encounter discovery requests that include PSAP documentation. Defense counsel investigating a case will request not only the CAD record but any AI-generated summaries, any classification suggestions, and any documentation of overrides. This is the Brady disclosure obligation (from Brady v. Maryland, requiring disclosure of exculpatory evidence) extended into the dispatch record. If the PSAP's AI system classified a call in a way that was later overridden, and that override is relevant to the case, it may be discoverable.

Agencies should document AI involvement in the CAD record itself, not as a separate file, but as a field or notation in the standard record. "Initial AI classification: Priority 4. Telecommunicator override to Priority 1. Rationale: voice indicators inconsistent with Priority 4." That notation does two things. It creates a transparent record that can be disclosed. It also creates the data the agency needs to evaluate whether its AI tool is performing as expected, and whether there are patterns in the overrides that indicate a systemic classification problem.

The Speed-Accuracy Tradeoff: What AI Gives and What It Cannot Replace

The efficiency argument for AI-assisted call handling is real. In PSAPs operating at high volume, the AI's contribution to transcription speed, classification starting-point, and CAD auto-population represents meaningful operational capacity. The telecommunicator who spends fewer seconds on routine keystrokes has more seconds for the call itself: listening, assessing, responding to caller needs, coordinating units.

But the efficiency argument has a boundary. The seconds saved by AI-assisted transcription do not justify a reduction in the time the telecommunicator spends listening to the call. The seconds saved by AI-assisted CAD population do not justify accepting a classification the telecommunicator has not verified. The efficiency gain is additive: it is time returned to the work the AI cannot do, not time subtracted from the oversight that prevents the AI from doing harm.

The Life Cost of a Misclassified Call

The most important thing to understand about dispatch classification is the cost of an error. In report writing, an AI error that goes through verification produces an inaccurate document. That is a serious problem with serious legal consequences. In dispatch, an AI classification error that goes through without override can produce a delayed response to a life-threatening emergency. The consequences are not legal and documentary. They are immediate and physical.

A Priority 1 call dispatched as Priority 4 means units respond at a non-emergency pace. It means the call sits in the queue while higher-priority calls are addressed. It means response time, the single variable most correlated with survival in cardiac events, fires, active threat situations, and domestic violence with weapons, is extended by the length of the queue. Every minute of that extension has a probability cost in human life.

This is why the human decision standard in dispatch is not "the AI is usually right." It is "the human owns the priority, every time, on every call." The classification engine can be right 98 percent of the time and still represent an unacceptable risk if the 2 percent where it is wrong includes calls like Adriana Reyes's call 247. The telecommunicator's ear is not a redundant backup to the AI. It is the primary instrument. The AI is the tool that helps organize and speed the process. The human is the decision-maker.

Building the Workflow for the Hard Cases

A well-designed AI-integrated call handling workflow is built around the hard cases, not the easy ones. The easy cases, the straightforward noise complaint, the clear traffic stop request, the medical call where the caller is describing symptoms clearly, will be classified correctly and dispatched efficiently with or without AI. The AI adds speed on these. The hard cases, the controlled-language domestic violence call, the caller who cannot speak freely, the ambiguous situation that does not fit a clear category, the caller in emotional collapse, are the cases where the workflow's design determines outcomes.

For the hard cases, the workflow must ensure three things. First, the AI suggestion is presented as a suggestion, not a decision. The visual and procedural design of the PSAP interface must make clear that the classification shown is a starting point the telecommunicator assesses, not a default they approve. Second, the override mechanism must be fast, easy, and logged automatically. Every override is a data point, not a deviation. Third, supervisor notification must be triggered automatically or easily triggered manually for any call where the telecommunicator overrides the AI suggestion by more than one priority tier. That override gap is a signal worth reviewing, both for the individual call and for what it reveals about the AI system's performance.

The workflow is not technology management. It is life safety management. The tool exists to help the telecommunicator do their job better. When the tool and the telecommunicator's professional judgment diverge, the professional judgment wins. That is not a policy preference. It is the irreducible principle on which every life in that PSAP depends.

Key Takeaways

• AI-assisted call handling spans the full incident lifecycle: intake and transcription, triage and classification, CAD (computer-aided dispatch) entry and dispatch, real-time update synthesis, and closure documentation. At every stage, the AI produces a draft or suggestion; the telecommunicator owns the decision.
• Transcription is the most technically reliable AI function at intake, but it has known failure modes in accent and dialect variation, emotional-state degradation of audio quality, and environmental noise. The telecommunicator's ear is always the primary reliability check and is never replaced by the transcript.
• The classification engine fails in predictable patterns: controlled-language calls where words understate danger, familiarity bias from address history, incomplete call information, and novel situations outside the training distribution. Professional judgment fills every one of these gaps.
• The override is not a workflow failure. It is the workflow's most important function: a trained human putting professional judgment over a statistical suggestion at the moment it matters. Overrides should be easy, fast, and automatically logged with a brief rationale so the agency can learn from them.
• Every CAD entry confirmed by the telecommunicator is their entry, not the AI's draft. Adoption makes the record theirs. The same authorship principle that governs AI-drafted police reports governs AI-assisted CAD documentation.
• A misclassified emergency call is not a documentation error. It is a delay in response that has a direct probability cost in human life. The telecommunicator owns the priority on every call, every time, because the AI's classification accuracy, however high, is never high enough to remove human ownership from a life-or-death decision.
• Disclosure obligations under Brady and agency transparency policies are beginning to reach PSAP documentation. Agencies should document AI involvement, classification suggestions, and overrides in the standard CAD record so that the AI's role is transparent, auditable, and disclosable in discovery.
• The efficiency gains from AI-assisted call handling are real and operationally significant. They represent time returned to the work only the telecommunicator can do: listening, assessing, and exercising the professional judgment that keeps the 2 percent of AI errors from becoming dispatched mistakes.