Norfolk AI Non-Emergency Calls Pass 40%, but Humans Still Control 911

Norfolk AI non-emergency calls now account for more than 40% of inquiries reaching the Virginia city's public-safety communications center. The assistant answers routine questions, routes residents to city departments, and transfers callers who need a human. However, Norfolk has drawn a firm line around emergency calls. Every 911 call still goes directly to a trained telecommunicator.

That separation defines the real significance of Norfolk's deployment. The city is not asking an automated voice to make emergency decisions. It is using artificial intelligence to remove routine work from employees who also handle shootings, fires, cardiac arrests, and CPR instructions.

The city says the system could eventually process 50,000 to 60,000 calls if its current resolution rate continues. That would make the project more than a limited chatbot trial. It would turn conversational AI into a working layer between residents and a public-safety agency.

Norfolk is also testing a larger promise facing local governments. Voice AI can answer many questions immediately, but public access cannot depend only on whether software understands a caller. The central contest is therefore automation versus dependable human access, not artificial intelligence versus public employees.

Norfolk AI Non-Emergency Calls Are Already Exceeding the Initial Target

Norfolk's assistant has moved from a projected call-diversion tool to a system handling more than 40% of non-emergency inquiries.

The Norfolk Police Department announced the service on October 5, shortly after its late-September launch. Its official deployment notice said the Amazon Web Services system was expected to divert roughly 35% to 40% of non-emergency calls.

By October 18, E911 Manager Steven Summerlin said the assistant was averaging more than 40%. That early result exceeded the upper end of the original estimate, although Norfolk has not released a detailed performance report.

The distinction between call handling and full resolution matters. A successful interaction can involve answering a question, transferring the caller to another department, or moving the call to a telecommunicator. Those outcomes do not provide the same level of service.

Norfolk's non-emergency number operates around the clock. Callers use it for police matters that do not involve an immediate threat, along with questions that sometimes belong with other city departments.

The city's human telecommunicators have handled more than 100,000 non-emergency calls during 2026, according to Summerlin. They have also taken more than 130,000 emergency calls during the same period.

The city's earlier announcement presented a broader annual picture. It said E911 employees answer more than 250,000 police, fire, and medical calls, plus approximately 150,000 administrative calls each year.

That mixed workload explains why apparently simple inquiries create a real operational problem. Emergency calls always receive priority, so a resident with a routine question can remain on hold for several minutes. A telecommunicator may then discover that the caller needs another department.

The AI assistant tries to eliminate that wait. It can answer questions about police reports, city services, storm preparations, evacuation guidance, shelters, code enforcement, and other documented procedures.

Residents do not have to navigate a conventional menu by pressing numbers. They can describe their request in ordinary language and receive a spoken response. Automatic speech recognition converts the caller's words into text so the system can identify the request.

The assistant then searches a controlled collection of city-approved information. Norfolk built that collection from municipal webpages and procedures documented by city employees. It does not search the open internet for an answer.

This design resembles a constrained AI knowledge base. The model retrieves information from an approved collection instead of treating the wider web as an authority.

City employees receive daily reports showing call volume and the share of interactions completed without a transfer. They can also inspect recordings and transcripts to find questions the system failed to answer.

Summerlin said officials can clarify the knowledge base quickly when a caller exposes a gap. That creates a practical improvement loop: review failures, add approved information, and test the revised response.

The system is not currently taking police reports or dispatching responders independently. Those functions remain possible future extensions, not capabilities residents should assume are available today.

Callers can also request a person. Saying, "I want to speak to an agent," moves the call into the human non-emergency queue. The transfer preserves a path to assistance, though it does not guarantee an immediate human answer.

That escape route is essential because non-emergency does not mean unimportant. Norfolk lists auto theft, burglaries not in progress, parking problems, loud music, and other police matters among appropriate reasons to call.

The first result is therefore narrower than the headline number suggests. Norfolk has shown that conversational software can absorb a substantial share of informational demand. It has not yet shown that the system improves every caller's outcome.

Why Norfolk Put AI Between Routine Questions and Dispatchers

The immediate goal is to reduce mental switching and queue pressure, not to eliminate public-safety jobs.

A Norfolk telecommunicator can handle between 85 and more than 100 calls during a 12.5-hour shift, according to Summerlin. Those calls can move abruptly between life-threatening emergencies and questions about garbage collection, overgrown lawns, or missing pets.

That switching carries an operational cost. An employee who has just guided a caller through CPR may next explain which office handles a routine complaint. Minutes later, the same employee must be ready for another crisis.

Norfolk's system removes some low-risk informational contacts from that sequence. The expected benefit is not simply fewer calls. It is more sustained attention for work requiring judgment, careful questioning, and emotional control.

The staffing context makes that claim especially important. Summerlin said the center had a vacancy rate near 50% about two years earlier. He described the current rate as approaching single digits.

However, a separate city update showed that staffing pressure had not disappeared. Police Chief Mark Talbot told an August 25 City Council work session that the E911 division lacked 16 employees across 83 positions.

Those figures are not necessarily contradictory. Vacancy counts can change, and a center may operate adequately on a typical shift while still carrying open positions. They show why Norfolk must measure workload more carefully than a single vacancy rate allows.

Summerlin has repeatedly said the assistant is not intended to replace telecommunicators. In local public-media coverage, he described it as a tool for repositioning available staff toward the center's highest-priority mission.

The operational boundary supports that position. Human employees still answer all 911 calls. The new assistant sits only on the non-emergency line and transfers matters requiring police or fire intervention.

Norfolk already uses another form of AI behind emergency calls. RapidSOS Communicator produces real-time transcriptions while a human speaks with the caller.

The software can flag words such as "gunshot," "not breathing," and "unconscious." Supervisors can then monitor a developing emergency while the original call taker remains engaged.

That background system illustrates a different automation model. AI listens, transcribes, summarizes, and signals. A trained person retains the conversation and the authority to interpret it.

The non-emergency assistant goes further because it speaks directly with the public. That makes the quality of recognition, retrieval, and escalation part of the resident's experience.

The technology is also inexpensive to expand compared with adding a staffed position for every increase in routine volume. Norfolk pays according to usage, but its reported commercial terms are not necessary to evaluate the public-service result.

The better question is what the city receives for that spending. Useful measurements include reduced hold times, fewer abandoned calls, faster emergency pickup, correct routing, and successful transfers to humans.

Norfolk has supplied an early automation rate, but it has not published those supporting measures. Summerlin said the system could eventually contribute to faster emergency responses, yet he did not claim that response times had already improved.

That distinction should remain visible. Removing routine calls creates capacity. Whether the center converts that capacity into faster or more accurate service requires separate evidence.

Norfolk's workload also varies during storms and other major events. Residents can flood city lines with questions about evacuations, shelters, pets, road conditions, and public facilities.

A conversational assistant can scale across many simultaneous calls during those peaks. Human staffing cannot expand as quickly, especially when the same event also generates real emergencies.

This is where the deployment has its strongest case. A current, city-controlled knowledge base can distribute consistent storm information without consuming the employees needed for urgent calls.

The same scenario also raises the stakes of an error. Outdated shelter details or misunderstood evacuation questions could send residents to the wrong place. Norfolk must therefore treat content maintenance as an operational responsibility, not a one-time setup task.

The Mechanism Is Controlled Retrieval, Routing, and Human Escalation

Norfolk's approach limits the assistant to approved information and predefined handoffs, reducing the freedom that makes general-purpose AI unpredictable.

The system begins by recognizing what the caller says. It then identifies the likely intent, searches the city's approved material, and produces a spoken answer or routing decision.

That sequence differs from a traditional interactive voice response system. A conventional system asks callers to choose from a fixed menu. Norfolk's assistant accepts conversational requests that may not use the city's preferred terminology.

A caller can ask where to report an abandoned vehicle without knowing which department owns that process. The assistant can interpret the request, provide instructions, or route the call.

The technology provider is Amazon Web Services. AWS documentation describes voice self-service systems that answer questions, maintain conversational context, and escalate interactions when a person is needed.

Its self-service documentation includes escalation as a standard action. That matters because a public-safety assistant must recognize both explicit requests for a person and situations it should not handle alone.

Norfolk has not publicly disclosed every service, model, prompt, confidence threshold, or contact-flow rule behind its implementation. Calling the system "Amazon AWS" identifies the platform provider, but not the complete technical configuration.

The city has explained the information boundary more clearly. The assistant relies on municipal webpages and institutional procedures gathered from employees. Summerlin said it does not browse the open web.

That restriction addresses one common failure mode. A general model can generate an answer that sounds credible without grounding it in an approved source. A controlled retrieval system narrows the material available for each response.

Narrowing the source material does not eliminate mistakes. The underlying documents can be incomplete, outdated, inconsistent, or difficult to match with a caller's language.

Speech recognition introduces another source of error. Accents, background noise, emotional speech, uncommon names, poor connections, disabilities, and code-switching can all affect a voice system's interpretation.

The assistant currently speaks English. People needing another language can transfer to a human and use the city's interpretation service.

Norfolk delayed multilingual automation because officials did not want callers to complete a long AI conversation before reaching an employee who could not continue in that language. That is a reasonable rollout constraint, but it leaves an access gap.

A resident who speaks limited English receives a different service path from an English-speaking caller. The city should measure whether that path involves longer waits, repeated explanations, or abandoned calls.

Human escalation also needs more than a recognized phrase. A nervous caller may not know to say, "I want to speak to an agent." The system should respond to varied requests, confusion, silence, repeated failures, and signs that an apparently routine situation is becoming urgent.

Amazon's architecture can preserve conversational context during a handoff. Norfolk has not publicly detailed how much context reaches its telecommunicators or whether callers must repeat their information.

That handoff quality will shape public acceptance. A transfer that carries the transcript and stated need can save time. A transfer that restarts the conversation adds another barrier.

The city can improve performance through its daily reports. Officials can identify unfamiliar questions, low-confidence interactions, transfer patterns, and calls that end unexpectedly.

Reviewing recordings creates its own governance questions. Residents should know when a voice is artificial, what data is stored, who can access transcripts, and how long recordings remain available.

Norfolk's public statements explain the human option but provide limited detail about disclosure and retention. The city should make those policies easy to find because the system sits at the entrance to public safety.

A clear disclosure also prevents a basic trust failure. Residents should not have to ask whether the person speaking is software.

The mechanism succeeds when it stays within a narrow role. It should answer documented questions, route known requests, and transfer uncertainty. It should not improvise about safety or persuade a caller that human help is unnecessary.

This division of labor gives Norfolk a defensible operating model. Automation provides speed and scale for repeatable information. People retain judgment, emergency questioning, and responsibility for decisions affecting responders.

The Real Test Is Access, Not the Call-Diversion Percentage

A 40% automation rate is useful only if callers receive correct information and can reach a person without unnecessary friction.

Norfolk's early number measures activity, not equity or accuracy. It does not reveal which callers the assistant failed to understand, how many abandoned the interaction, or whether transfers reached the correct destination.

Comparable deployments show why those details matter. Richmond, Virginia, launched a customized Amazon Connect assistant for non-emergency and city-service calls before Norfolk.

Richmond reported that its system completely handled about 10% to 15% of non-emergency calls. It routed another 5% directly to other agencies.

An independent Richmond field test found mixed results. The assistant correctly addressed some infrastructure questions, but one transfer appeared to disconnect. A Spanish-language interaction also ended with an English-speaking dispatcher.

Richmond's experience does not prove that Norfolk has the same problems. It shows that a high-level resolution rate cannot substitute for testing the full journey.

Renée Cummings, a University of Virginia assistant professor who studies ethical AI, framed the access problem clearly in the Richmond reporting. A system does not improve access if it works only for people it easily understands.

That standard is especially relevant in public safety. Commercial support systems can frustrate customers without affecting their legal rights or physical security. A government phone line carries a different obligation.

Norfolk's callers may include older residents, people with speech disabilities, residents using low-quality connections, and people speaking under stress. The system must accommodate them rather than define the easiest caller as normal.

Disclosure represents another requirement. The National Emergency Number Association has advised agencies to tell callers when they are interacting with AI, according to the Richmond reporting.

That transparency is more than etiquette. It helps callers interpret pauses, repeated questions, and unexpected responses. It also tells them when requesting a human may be appropriate.

Accuracy needs a published denominator. Norfolk says the assistant handles more than 40% of calls, but the public needs to know what "handles" means.

A rigorous report would separate questions answered, calls routed to departments, transfers to E911 staff, caller-requested escalations, disconnected calls, and unresolved sessions. It would also track whether the final destination was correct.

The city should compare those measures across language needs, call types, times of day, and recognition difficulties. Aggregated reporting can reveal service gaps without exposing individual callers.

Norfolk should also audit whether automation changes caller behavior. Some residents may use 911 when the non-emergency assistant frustrates them. Others may abandon reports that still matter to neighborhood safety.

These spillover effects would weaken the deployment even if the assistant's internal completion rate looked strong. The city should monitor 911 misclassification, repeat calls, and online-reporting volume.

Data governance presents another uncertainty. Voice recordings and transcripts can contain names, addresses, medical details, allegations, and information about people who never called the city themselves.

Norfolk has said staff can review transcripts and recordings to improve the system. It has not widely published the retention schedule, vendor access rules, redaction process, or use of call data for model training.

Those policies should be explicit before the assistant takes on more sensitive work. Collecting preliminary information for police reports would carry greater risk than answering a shelter question.

The city must also protect its knowledge base from stale instructions. A correct answer today can become wrong after a department changes its hours, phone number, eligibility rules, or emergency plan.

Daily technical reports do not solve content ownership. Each participating department needs someone responsible for approving changes and checking high-impact information.

External testing would strengthen Norfolk's claims. Staff who built the system know the expected questions and may not reproduce the unpredictable language used by real callers.

Testing should include accented speech, interrupted sentences, ambiguous emergencies, background noise, disability-related speech patterns, and requests that cross department boundaries. Reviewers should also test indirect demands for human help.

The goal is not perfect recognition. No human or automated system achieves that standard. The goal is fast recovery when recognition fails.

Norfolk's decision to preserve a live-agent request is therefore one of the project's strongest safeguards. Its value depends on whether the assistant consistently recognizes that request and completes the transfer.

The city should publish performance evidence before expanding the assistant into report intake or more consequential workflows. A large call-diversion number alone does not justify a larger decision-making role.

Charleston and Richmond Show This Is Becoming a Municipal Playbook

Norfolk is part of a growing shift from experimental chatbots toward voice automation embedded in emergency communications operations.

Norfolk learned about the approach from Charleston County, South Carolina. That county had already placed machine learning on its non-emergency line and became a reference point for other public-safety centers.

Charleston announced a fully automated non-emergency workflow in November 2023. The county said its center received about 210,000 non-emergency calls each year.

Its automation announcement described common requests involving barking dogs, stolen bicycles, and noisy neighbors. Those matters still needed reporting, but they did not always require a call taker.

Charleston paired phone automation with online tools for reports, agency contacts, duty officers, and street-name reservations. That broader approach matters because voice AI should not become the only route to service.

Virginia Beach also adopted similar technology before Norfolk, according to Summerlin. After Norfolk's launch, two additional centers contacted the city to discuss the project.

This pattern suggests that municipal adoption will spread through peer networks. Emergency communications leaders attend the same conferences, compare staffing problems, and examine systems already operating in similar jurisdictions.

The vendor platform also lowers the implementation barrier. Amazon Connect supplies telephony, contact flows, speech services, analytics, and knowledge-based self-service within one cloud environment.

Local agencies still need to configure the workflow, approve information, and define escalation rules. They do not have to build a voice network or speech engine from the beginning.

That combination makes the model replicable. A city can start with frequently asked questions, monitor transfer rates, and add approved processes later.

The competitive pressure is not primarily between Norfolk and another city. It falls on the traditional operating model in which trained emergency staff also serve as general municipal information agents.

Once one center reports that AI removes a third or more of its non-emergency burden, neighboring agencies must explain why they are maintaining long hold times. Vendors then gain a credible public-sector reference.

Yet Charleston and Richmond also show that deployments differ substantially. Charleston emphasized automation and online reporting. Richmond uses AI as a city-services navigator. Norfolk emphasizes conversational answers from a controlled knowledge base.

Those differences complicate comparisons. One city's resolution percentage can reflect simpler calls, broader routing rules, or a different definition of success.

Public officials should avoid treating call diversion as a league table. A cautious system that transfers uncertainty may show a lower automation rate while delivering safer outcomes.

The most useful comparison is operational. Cities should ask how frequently the assistant recognizes intent, how quickly callers reach the correct destination, and how often humans must repair an automated interaction.

They should also compare service continuity during storms and major events. Voice automation has the greatest operational value when demand rises suddenly and information changes quickly.

Norfolk's coastal location gives that use case particular importance. Hurricane questions can arrive in large waves, including requests about evacuation zones, shelter openings, transportation, and animals.

A city-approved assistant can repeat current guidance across simultaneous calls. However, officials must update the source material immediately when conditions change.

This is not a "set and forget" chatbot project. It is a new public communications channel requiring staffing, monitoring, testing, and governance.

The deployment also illustrates a broader development in applied AI. Many useful systems are becoming narrower, not more autonomous.

Norfolk restricts the assistant's sources, keeps emergency calls human, and offers escalation. Those constraints are not weaknesses. They define where automation has a plausible public benefit.

Cities considering the same model should begin with that boundary. Routine information and deterministic routing are suitable starting points. Emergency interpretation and discretionary decisions are not.

Norfolk's rollout will become influential if it demonstrates more than impressive volume. Other agencies will want evidence that the assistant reduces queues without creating new access barriers.

What to Watch After Norfolk's 40% Milestone

The next phase should be judged by emergency performance, equitable access, and disciplined limits on expansion.

The first signal is whether Norfolk publishes outcome data beyond the share of calls handled by AI. Emergency pickup times provide the most direct test of the city's central claim.

A sustained reduction in 911 answer times during comparable periods would strengthen the case that removing routine inquiries benefits emergency operations. Flat results would suggest the recovered capacity is going elsewhere.

Non-emergency hold times also matter. Callers transferred by the assistant should not face longer queues than they experienced before automation.

The city should report abandoned calls and repeat contacts. Falling numbers would indicate that people are receiving useful answers. Rising numbers would expose frustration hidden behind the completion rate.

The second signal is the performance of human escalation and language access. Norfolk needs to show that callers can reach a person through ordinary speech, not just one memorized command.

That evaluation should include residents who speak with accents, use assistive technology, or require an interpreter. It should also cover noisy environments and calls made under emotional stress.

If Norfolk expands automated language support, the entire handoff must remain multilingual. A polished AI conversation followed by an English-only transfer would move the barrier rather than remove it.

The third signal is whether the city expands the assistant into report collection or dispatch preparation. Summerlin has discussed those functions as future possibilities, but they are not part of the present service.

Collecting structured information before a human joins could reduce repetitive questioning. It would also expose the system to more sensitive facts and greater consequences from misunderstood details.

Any expansion should follow a public impact review. Norfolk should explain what the system collects, how staff verify it, when data enters official records, and how residents correct errors.

Emergency dispatch should remain outside the assistant's autonomous authority. The city's current position supports that boundary, and Summerlin has said direct human interaction remains essential when someone is in danger.

Norfolk should also clarify its disclosure, retention, and vendor policies. Residents deserve to know when AI is speaking, how their calls are stored, and whether their data improves any external model.

These questions do not negate the early result. Handling more than 40% of non-emergency inquiries is a meaningful operational signal, especially when the same center manages large emergency volumes.

The figure shows that many calls reaching public-safety employees are informational and repeatable. Conversational software can address a substantial portion without forcing residents through rigid menus.

The unresolved question is whether Norfolk can preserve that convenience across the people and situations the software understands least easily.

For public agencies, the strongest AI deployment is not the one that removes the most humans from a workflow. It is the one that gives humans more time for judgment while keeping them reachable.

Norfolk AI non-emergency calls now provide a visible test of that principle. Watch the answer-time data, failed transfers, language performance, and scope of future expansion. Those measures will reveal whether the assistant is improving public access or merely producing an attractive automation percentage.