One in Five Calls Now Ends Without an Officer: Drone-as-First-Responder and the Calls-for-Service Data Layer

More than a thousand US agencies now hold FAA waivers to fly a drone to a 911 call before a patrol car leaves the lot. In the longest-running program, roughly one mission in five ends with no ground unit responding at all. That is a real operational result. It is also a change to the machinery that manufactures crime statistics — and almost nobody is measuring it.

The Waiver Curve Went Vertical

Drone-as-first-responder (DFR) is the practice of launching an uncrewed aircraft from a fixed rooftop dock in response to a dispatched call, so that a live camera feed reaches the incident before officers do. The first US program launched in Chula Vista, California, in 2018. For most of the following seven years it stayed a niche practice, because the required FAA authorization — permission to operate beyond visual line of sight (BVLOS) — was slow and case-by-case.

That constraint broke in 2025. According to the Electronic Frontier Foundation, 976 DFR-enabling waivers were issued between 2018 and April 2025. After the FAA streamlined its review process in April 2025, it issued more waivers in the following ten months than in the prior seven years combined. By February 2026, more than 1,000 public safety agencies held the authorization.

Holding a waiver is not the same as running a program, and a program covering a two-mile radius around one dock is not the same as citywide coverage. The waiver count is a ceiling on adoption, not a measurement of it. But the direction and the slope are not ambiguous, and the operational case is strong enough that adoption is unlikely to stall on its own.

What the Operational Numbers Say

Chula Vista PD passed 25,000 DFR missions in May 2026, the largest published record of any program. The department reports the drone arrived before ground units on 17,170 of those missions, with an average lead-arrival response time of 96.98 seconds. It reports the aircraft played a role in 4,138 calls resulting in arrests. And it reports 4,629 calls — about 18.5% of all missions — where no patrol unit responded at all, because the drone either resolved the situation or confirmed there was nothing to respond to.

Other programs report figures in the same family, with wide spread:

• Fairfax County, VA— launched fall 2025. In a February 24, 2026 report to the Board of Supervisors, the department said drones arrived first on more than 70% of the program's first 100 missions, averaging under 90 seconds. Phase two plans 18 dock locations sited from calls-for-service data.
• Miami Beach, FL— reported handling over 1,200 calls for service in its deployment area across 22 weeks, clearing 41% of them without officers.
• Casa Grande, AZ— reported a July 2026 average drone response of 1:30 against 4:19 for ground units.

Treat the spread between 18.5% and 41% as a methodological finding rather than a performance ranking. These figures are self-reported by the agencies deploying the technology, frequently in budget or procurement contexts, and often reproduced in vendor marketing. More importantly, the denominators are not the same quantity. Chula Vista's rate is over all DFR missions launched. Miami Beach's is over calls for service in a defined deployment area. A department that launches selectively — only on call types where a camera is likely to settle the question — will post a much higher clear-without-officer rate than one that launches broadly, without being better at anything.

None of these are peer-reviewed evaluations, and we are not aware of a published quasi-experimental study that isolates the effect of DFR deployment on recorded crime volume. The honest summary is: the response-time improvement is large, consistent, and mechanically plausible; the share of calls resolved without an officer is somewhere in the high teens to low forties depending on how you count; and the downstream effect on the crime record has not been measured by anyone.

Two Datasets, One Event

Understanding why this matters requires separating two things that are routinely conflated. A police department produces at least two distinct data streams from the same underlying event:

Calls for service (CAD). A dispatcher opens a record when someone calls 911 or an officer self-initiates. It carries a call type, a location, a timestamp, a priority, and eventually a disposition. It reflects what someone reported or what a dispatcher coded. It is high volume and fast — often available same-day.

Incident reports (RMS). If a responding officer concludes a crime occurred, they write a report. That report is classified, entered into records management, and eventually flows to NIBRS and to public incident feeds. It reflects what an officer documented. It is lower volume and slower.

Almost every crime map, safety score, and crime data API in production is built on the second stream. The first is the funnel that feeds it. We have written before about the dark figure of crime — the gap between crimes that occur and crimes that are reported. The DFR question sits one stage further down: the gap between calls that are dispatched and reports that get written.

DFR inserts itself precisely there. It changes what happens between dispatch and documentation.

Three Mechanisms, Two Directions

The tempting conclusion — drones clear calls, so fewer reports get written, so recorded crime falls artificially — is too fast. There are at least three distinct mechanisms and they do not all point the same way.

1. Disposition substitution (suppresses records)

A call is dispatched. The drone arrives, the operator sees an empty parking lot, and the call is closed as unfounded or gone-on-arrival without a ground unit. Under the previous workflow an officer would have driven there, made contact, and — depending on department policy, the complainant's wishes, and what they found — possibly written a report. Some fraction of those reports now do not get written. In categories where the report was discretionary in the first place (minor property damage, disturbances, suspicious circumstances, low-value theft), this is where the effect should concentrate.

2. Verification (works both directions)

A drone overhead at 90 seconds sees things an officer arriving at four minutes does not. That can confirm a call was unfounded — suppressing a report that would have been written on weaker evidence. It can also confirm an offense in progress that would otherwise have been logged as unfounded, and Chula Vista's 4,138 arrest-assisted calls suggest this is not rare. Better evidence at the scene changes classifications in both directions, and the net sign is an empirical question nobody has answered.

3. Reallocation (redistributes records)

Thousands of redirected patrol hours do not evaporate. They go somewhere — proactive patrol, traffic enforcement, follow-up investigation, other calls. Self-initiated police activity generates records. If freed capacity is spent on activity that produces reports, total recorded volume can rise even as clear-without-officer counts rise, and the composition of recorded crime shifts toward police-initiated categories. This is the same dynamic that makes drug and weapons offense counts primarily a measure of enforcement intensity rather than underlying behavior.

The net effect of these three on any given city's published incident count is genuinely unknown. What is not unknown is that they represent a change to the data-generating process, arriving on a steep adoption curve, at different times in different jurisdictions, with no flag in the output.

A Bounding Exercise

It is worth putting a rough ceiling on this before treating it as a crisis. Assume a mid-size city where a mature DFR program covers a deployment area accounting for 25% of citywide call volume, and assume it resolves 20% of the calls it launches on without a ground unit. Assume further — generously — that half of those calls would previously have produced an incident report. That yields a ceiling of roughly 2.5% of citywide dispatched calls losing their report, concentrated almost entirely in low-severity, high-discretion categories.

Every one of those assumptions is a guess, and the middle one is doing the most work. A 2.5% shift is not visible against national aggregates. It is very much visible in a year-over-year comparison for a single beat, in a hotspot map of a deployment radius, or in a block-level safety score computed on a 12-month window that straddles a program launch. The finer the geography, the larger the artifact relative to the signal.

This Does Not Explain the National Decline

It should be said plainly, because the temptation to over-apply a mechanism like this is strong. The crime decline currently underway is far too large, too broad, and too early to be a DFR artifact.

The Real-Time Crime Index, covering 586 agencies and 122.4 million people, shows January–May 2026 against the same period in 2025: violent crime −5.6% (215,636 vs. 228,316), murder −18.1% (2,519 vs. 3,076), robbery −14.5%, property crime −10.7%, motor vehicle theft −19.7%. This extends a decline that USAFacts tracks back through 2001, with the 2024 US violent crime rate at 359 per 100,000 and property crime at 1,760 per 100,000 — overall crime down 49.1% since 2001.

Three reasons DFR cannot be behind this. The declines are steepest in murder and motor vehicle theft — the two categories where a drone clearing a call and no report being written is least plausible; nobody closes a homicide by camera. The decline predates the 2025 waiver acceleration by several years. And DFR coverage across the RTCI sample is far too thin to move an aggregate of that size.

The DFR data effect, if it exists, is a local measurement artifact in specific deployment areas over specific windows. It is not a national trend story. Conflating the two would be exactly the error this post is arguing against.

Calls for Service Deserve to Be a First-Class Layer

The constructive read is that DFR raises the value of a dataset most crime products ignore. If the interesting action is increasingly happening at the dispatch layer, the dispatch layer needs to be visible.

CAD data has real advantages: it is fast, it is high volume, and it captures demand for police service rather than police documentation of crime. It has equally real drawbacks. Call types are assigned by a dispatcher from a caller's account, so they are frequently wrong. Duplicate calls for one event are common. Disposition codes are non-standard across agencies to a degree that makes NIBRS look tidy. And calls for service are emphatically not crime — presenting them as such would repeat the over-claiming problem we described in incident taxonomy normalization.

Used carefully, though, CAD is the control series for exactly this problem. If dispatched calls for a category hold steady while incident reports for that category fall after a DFR launch, you are looking at a documentation change, not a crime change. That comparison is only possible if both series are published. For analysts building this kind of pipeline, Andrew Wheeler's practical guide to Python data science for crime analysts is a good starting point for the pandas-and-SQL end of the work.

The Transparency Ask

The disclosure needed here is modest and mostly already technically available, since every one of these events already exists as a CAD record. Four things:

• A DFR flag on the call record. Whether a drone was launched, and whether it arrived first. Without this, the affected population cannot be identified at all.
• Disposition detail. Specifically, whether a call was closed without ground response, and under what code.
• Program launch dates and deployment geography. The dock locations and effective radius, so an analyst can define a treated area and a comparison area.
• Report-rate reporting. The share of DFR-resolved calls that produced an incident report, against the same share for comparable non-DFR calls.

California's AB 481 already requires advance public notice of military and surveillance equipment acquisition, a use policy adopted before purchase, and annual deployment reporting — a template that covers acquisition well and data effects not at all. The broader asymmetry is familiar: agencies are deploying increasingly capable sensing infrastructure while the corresponding public data stream stays thin, a pattern we traced in automated license plate reader networks. EFF has separately raised the concern that DFR aircraft integrated with ALPR and camera networks function as flying sensors over backyards and windows, and that deployments skew toward low-risk calls involving unhoused people and mental health crises rather than the high-risk scenarios used to justify procurement. Those are civil liberties questions, distinct from the measurement question here, and both deserve their own scrutiny.

What Developers Should Actually Do

• Treat program launches as structural breaks. Maintain a list of DFR launch dates for cities you cover, and do not compute year-over-year change across one without a caveat. This is the same discipline that a NIBRS transition or a records system migration demands.
• Check for level shifts in low-severity categories. Disturbances, suspicious persons, vandalism, petty theft. If a discontinuity shows up in those and not in aggravated assault or motor vehicle theft, suspect documentation before behavior.
• Never build a safety score on a single window that straddles a launch. Temporal decay weighting makes a recent documentation change disproportionately influential in the current score.
• Do not present response-time improvements as safety improvements. Faster arrival is a service-delivery metric. It is not evidence that a place became safer.
• Keep automated triage out of the classification path. The CrimeRadar false alert in Mount Vernon, Missouri is the reference case — a routine radio transmission misread as a school shooting, triggering a lockdown. As that write-up put it, systems like this are built to be fast, and safety requires being right. Those are not the same thing. The point generalizes: model confidence is not calibrated probability, as Gio Circo demonstrates in his analysis of LLM classifier calibration on NEISS injury data, where token probabilities run systematically overconfident. As DFR programs add autonomous launch-decision logic, that caution moves upstream into who gets a drone sent to them.

Bottom Line

DFR programs appear to work as advertised on the metric they were bought for: drones arrive faster than cars, by a wide and consistent margin. That is a genuine result and the adoption curve reflects it.

The measurement consequence is a second-order effect nobody purchased and nobody is tracking. When a call is resolved from 200 feet and no officer makes contact, the event still happened, someone still called about it, and a record still exists — in CAD, not in the incident feed. Whether that matters depends entirely on the resolution at which you are working. At national scale it is noise against a decline that is large, broad, and years older than the technology. At the block level, over a 12-month window, in a deployment radius, it is exactly the kind of artifact that gets mistaken for a neighborhood getting safer.

The fix is not complicated: publish the flag, publish the disposition, publish the launch date. Until then, the correct posture toward any DFR city's recent incident counts is the one that has served crime data consumers well for a long time — assume the pipeline changed before assuming the world did.