AI Video Analytics Software: Nine Vendors, One Honest Comparison

Most comparison articles on AI video analytics are advertorials with footnotes.

The genre is well established. A vendor pays for a slot, the slot is dressed as editorial, the verdict is preordained. What follows here is the opposite exercise. Nine vendors that an operator in construction, industry or logistics will actually encounter when issuing a serious tender in 2026, judged against the same eight criteria, by an author who manufactures competing hardware and therefore has every reason to be precise about where competing software ends and begins. The conflict of interest is declared. The method below is designed to make that declaration survivable: the criteria are operator criteria, not vendor criteria, and the verdicts are qualitative ranges, not flattering precision.

The reader who expects a winner will be disappointed. There is no winner in this market because the market is not yet mature enough to produce one. There are credible vendors with distinct strengths, weak spots that they would rather not discuss, and a small number of cases in which an operator should walk away regardless of price. The pages below name the strengths, the weak spots and the walk-away cases. The book BOSWAU + KNAUER. From Building to Security Technology develops the broader argument in chapters nine and eighteen, and the present article extends it into the vendor landscape.

The eight criteria, defined before any vendor is named

A comparison is only as honest as the criteria it applies before the names of the candidates appear. Once the names appear, the criteria bend. The list below is therefore fixed first, in the order of weight that an industrial operator would apply on a Tuesday morning when something has actually gone wrong.

The first criterion is false-positive rate under field conditions, not laboratory rate. Vendors quote rates from controlled environments with calibrated cameras and clean backgrounds. The operator sees rain, fog, vibration, dust and the occasional bird. A vendor that cannot produce field data from sites comparable to the operator's own does not pass the first hurdle.

The second is detection latency from event entry to operator alert. Sub-second is the target. Two to four seconds is acceptable for non-critical zones. Anything above five seconds is a system that documents incidents rather than preventing them, and should be priced accordingly.

The third is the breadth of supported models, in the sense of what the system can recognise out of the box and what requires custom training. A vendor that requires custom training for person-versus-vehicle classification has not done its homework. A vendor that claims to recognise everything with no training has not done its testing.

The fourth is edge versus cloud architecture, and specifically the behaviour of the system when the cloud link drops. Industrial sites lose connectivity. The system must remain useful while the link is down, both for detection and for local storage, in line with the segmentation logic that IEC 62443 expects of operational technology.

The fifth is integration depth with adjacent systems: VMS, PSIM, access control, alarm receiving centres, leitsystems in industrial plants. A vendor with strong models and no integration is a research project. A vendor with weak models and deep integration is a middleware company. The operator needs both.

The sixth is data governance and the regime under which footage and metadata are processed. For European operators this means GDPR-compliant processing, documented data residency, and a clear position on whether training data leaves the customer environment. Vendors who cannot answer the residency question in one sentence are not ready for regulated sectors.

The seventh is cyber posture of the software itself, mapped against NIST CSF 2.0 and, for industrial deployments, IEC 62443-4-1 and 4-2 for the secure development lifecycle of the components. CISA advisories for the vendor's products over the past twenty-four months are part of the assessment. Zero advisories is not necessarily a good sign. It often means nobody has looked.

The eighth is commercial transparency: list prices, the structure of per-camera, per-stream or per-site licensing, the inflation clauses, the exit conditions. A vendor that treats pricing as a state secret is a vendor whose pricing the operator will regret.

The nine vendors, and why these nine

The criteria above were applied to the visible market in late 2025 and early 2026. The list that survived has nine names. Three of them are European, four are North American, two are from Asia. They are: Avigilon (now part of Motorola Solutions), Axis Communications with its analytics stack, BriefCam (Milestone), Eagle Eye Networks, Genetec with KiwiVision and Security Center analytics, Hanwha Vision with Wisenet AI, Hikvision with its AcuSense and DeepinView lines, Milestone Systems with XProtect plus the integrated analytics partners, and Verkada.

These nine are not a ranking. They are the names that an operator with a credible budget and a serious tender will receive bids from. Smaller specialists exist, some of them excellent in narrow niches such as wrong-way detection on highways or PPE compliance on construction sites. They are absent here because the article is about platforms that an operator can build a multi-site programme on, not about point solutions. ASIS International's surveys of the integrator community broadly confirm this shortlist for the segment in question.

Two omissions deserve a sentence. BriefCam is included as part of Milestone after the 2018 acquisition, but its forensic search capability is treated separately in the integration section because it is still sold as a distinct module. Bosch Security has been omitted as a standalone entry because its analytics are tightly coupled to its cameras and the relevant comparison is at the camera level, not at the software level. The decision is debatable. An operator who is already committed to Bosch hardware should add it as a tenth.

The vendors that an operator should be cautious about, and which therefore did not make this list, share two traits. They claim per-camera analytics performance that exceeds what their published latency figures permit, and they decline to provide reference deployments at the operator's scale. The combination is diagnostic.

How each vendor performs against the eight criteria

The most useful way to read what follows is by criterion, not by vendor. Vendors are described in clusters where they behave similarly, and singled out where they diverge sharply.

On false-positive rate, the cleanest field performance comes from Avigilon, Axis and Hanwha, in that order, for outdoor industrial scenes with mixed lighting. Genetec and Milestone, because they rely on a combination of native and partner analytics, are bimodal: the native models are competitive, the partner models range from excellent to embarrassing depending on which partner is configured. Hikvision's AcuSense performs well in stable installations and degrades faster than the others when scene conditions drift, in part because retraining is less accessible to the integrator. Verkada delivers a strikingly low false-positive rate in tidy retail and office environments and a less impressive one in heavy industrial scenes, where its camera-bound analytics meet conditions for which they were not primarily designed. Eagle Eye sits in the middle, with strong cloud-side filtering that partially compensates for variability at the edge. BriefCam, used in its intended role of forensic search rather than real-time alerting, does not really compete on this axis at all.

On detection latency, Verkada, Avigilon and Hanwha lead for edge-resident detection, typically below one second from event to alert when the operator and the camera are on the same network. Genetec and Milestone introduce additional latency through the VMS layer, often in the one to three second range, which is acceptable for most use cases and unacceptable for perimeter triggers that must drive a deterrent response. Eagle Eye, by virtue of its cloud architecture, runs longer at the high end of the acceptable range and depends on the customer's uplink. Hikvision is fast on its own cameras and slower when its analytics are forced through third-party VMS.

On model breadth, Genetec, Milestone and Avigilon are the most complete out of the box, because they have either invested heavily in native models or curated their partner ecosystems aggressively. Verkada has narrower model coverage but the coverage it has is well tuned. Hanwha and Axis cover the standard industrial classifications competently. Hikvision's breadth on paper is large; the breadth that survives independent testing is more modest. BriefCam is in a category of its own, optimised for retrospective search across long video archives rather than for real-time classification, and within that category it has no real peer.

On edge versus cloud, Avigilon, Hanwha, Axis and Hikvision are edge-first by design and behave correctly when the link drops. Genetec and Milestone are server-based on-premise with optional cloud, which for European operational technology environments is the architecture that maps most cleanly onto IEC 62443 segmentation. Eagle Eye and Verkada are cloud-native, with local buffering, and the operator must accept that architecture or move on. Both are honest about it, which is more than can be said for some others.

On integration depth, Genetec and Milestone are unmatched, because integration is their business model. Avigilon integrates cleanly inside the Motorola Solutions stack and competently outside it. Axis and Hanwha integrate through ONVIF and through their own SDKs with the major VMS platforms. Hikvision integrates broadly but the quality of the integrations varies. Verkada integrates with a deliberately limited set of partners, which is a feature for some buyers and a defect for others. Eagle Eye integrates with what its API supports, which is growing but still narrower than the on-premise incumbents.

On data governance, Axis, Genetec and Milestone are the cleanest for European deployments, with documented residency, mature data processing agreements and clear answers on training data. Hanwha is solid. Avigilon is acceptable inside the Motorola structure. Verkada and Eagle Eye, being cloud-native and US-headquartered, require careful contractual work for European regulated sectors, and the work is doable but should not be underestimated. Hikvision raises issues that go beyond data governance into procurement policy in several jurisdictions, and operators in critical infrastructure should consult their national authorities. BSI guidance in Germany on Chinese-origin surveillance components in critical contexts is the relevant reference.

On cyber posture, the picture in 2025 and 2026, drawing on published CISA advisories and the vendors' own disclosures, places Axis, Genetec, Milestone and Avigilon in the top tier, with mature secure development practices aligned to IEC 62443-4-1. Hanwha is improving. Verkada has rebuilt its programme after the 2021 breach and now publishes credible material on its practices. Eagle Eye is competent. Hikvision has been the subject of multiple advisories over the years and operators should weigh them against their own risk tolerance. NIST 800-53 controls mapped onto these products differentiate them less than the secure development lifecycle differentiates them, and the latter is where the operator should look.

On commercial transparency, the European vendors are generally clearer than the American ones, with the exception of Verkada which publishes structured pricing tiers. Hikvision is opaque in Western markets and aggressive on price. Eagle Eye is transparent within its subscription model. Genetec and Milestone are channel-driven and the operator's experience depends on the integrator, which is a polite way of saying that pricing transparency is variable.

The false-positive question, treated honestly

False-positive rate is the criterion vendors most often misrepresent and the criterion that most often determines whether a system survives in operation. A system with a false-positive rate of one alert per camera per day is workable. A system at ten alerts per camera per day is not, because the operators will mute the channel within a fortnight, and the GDV's loss data on alarm-system effectiveness confirms what every site manager already knows: a muted system is a system that does not exist.

The honest range for the nine vendors in industrial outdoor conditions, with mixed lighting and weather, is between roughly one and roughly fifteen false positives per camera per day before tuning. After two to four weeks of tuning by a competent integrator, the same systems settle into a range of zero point three to three false positives per camera per day. The vendors that publish numbers below zero point one in marketing material are quoting indoor, well-lit, low-motion conditions. The numbers are not wrong. They are just not relevant to the construction yard at three in the morning.

The NICB and equivalent European bodies have begun to publish loss data that, indirectly, validates these ranges, by showing that sites with tuned analytics see substantially reduced reported incidents while sites with untuned analytics show no measurable improvement. The lesson is not that analytics do not work. The lesson is that the integrator's tuning effort is more determinative of outcome than the vendor's marketing rate.

Industrial deployments, where the picture sharpens

For industrial deployments, the criteria that matter most are edge resilience, IEC 62443 alignment, false-positive rate under harsh conditions, and integration with the plant's operational technology. The list of vendors that meet all four well is shorter than the list of nine. Axis, Hanwha, Genetec and Milestone form the core of credible choices. Avigilon is credible inside the Motorola ecosystem, which for plants already using Motorola radios is often the path of least resistance. The cloud-native vendors are credible for sites where the OT environment is air-gapped from the surveillance environment and the surveillance environment can be allowed to depend on a wide-area uplink, which is a smaller set of sites than it appears.

For construction sites, where the deployment is temporary and the environment is the harshest, the relevant criteria shift. Speed of deployment, ruggedness of the hardware that carries the analytics, and the ability to run autonomously without site IT become decisive. Here the analytics question merges with the hardware question, and the operator is buying a system rather than a piece of software. This is the territory in which BOSWAU + KNAUER builds its own platforms, and the article is conscious of that overlap. The honest disclosure is that for permanent industrial sites the software vendors above are the relevant choices, and for temporary construction sites the question is different and is treated in chapter eight of the book referenced earlier.

What an operator should do with this comparison

The nine vendors above will, in any realistic tender, produce three or four credible bids. The operator's task is not to pick the best vendor in the abstract but to pick the vendor whose strengths align with the operator's worst risks. A logistics yard with a steady connection and a tolerance for cloud architecture will choose differently from a chemical plant with strict OT segmentation and a low tolerance for foreign vendors in safety-relevant zones. The comparison above gives the operator the vocabulary to articulate the difference. It does not make the choice.

The choice itself is best made on the basis of a structured audit of the operator's existing footprint, the documented incidents of the past twenty-four months, and an honest mapping of the criteria above against the operator's own risk profile. ISO 27001 and the NIST CSF 2.0 functions of identify, protect, detect, respond and recover provide the scaffolding. The audit produces the answer. The audit cannot be replaced by a comparison article, including this one.

For operators who want to test that proposition without commitment, Path I from the book provides a sixty-minute confidential conversation in which the criteria above are applied to the operator's actual situation. For operators ready to commit to a structured assessment, Path II is a three to five day audit that produces a written report with a vendor recommendation among the nine, ranked against the operator's specific risk profile. Path III, the ninety-day pilot, takes the recommendation into operational reality on one defined site before any multi-site rollout is contemplated.

What holds

The AI video analytics market in 2026 contains nine vendors that an operator can reasonably consider, and a larger number of vendors that should be treated as research projects until they earn a place on the shortlist. Among the nine, none is best across all criteria, and the operator who waits for one to emerge will wait through several more incidents.

The decision that matters is not which vendor to choose. The decision is which two or three criteria, among the eight, are non-negotiable for the operator's specific environment. Once those are fixed, the field of nine narrows to two or three, and the choice becomes a matter of integrator quality and commercial terms rather than of software capability. The work, in other words, is upstream of the vendor list.

The honest closing observation is that the difference between a system that prevents losses and a system that documents them is rarely the software. It is the tuning, the integration and the operational discipline that surrounds the software. A mediocre vendor in capable hands outperforms a top vendor in indifferent hands every time. Capability is in the operator before it is in the platform.

Frequently asked questions

Which vendors lead in AI video analytics?

No single vendor leads across all relevant criteria. For European industrial deployments, Axis, Genetec, Milestone, Hanwha and Avigilon form the core of credible choices, each with distinct strengths. Verkada and Eagle Eye lead in cloud-native environments where their architecture fits the operator's risk profile. Hikvision retains a significant installed base but raises governance and procurement questions in critical sectors. BriefCam leads in forensic search rather than real-time analytics. Leadership is therefore a function of use case, not an absolute ranking, and the operator's worst risks should drive the choice.

How are they compared?

The comparison in this article uses eight operator-defined criteria, fixed before the vendor names appear: false-positive rate under field conditions, detection latency, breadth of supported models, edge versus cloud architecture, integration depth with adjacent systems, data governance, cyber posture mapped against NIST CSF 2.0 and IEC 62443, and commercial transparency. Each vendor is described against each criterion in qualitative ranges rather than precise figures, because precise figures from vendor marketing rarely survive field conditions. The criteria themselves are weighted by the operator according to the operator's specific environment.

What is the false-positive rate of each?

Honest field ranges for industrial outdoor conditions before tuning sit between roughly one and fifteen false positives per camera per day, depending on vendor and scene. After two to four weeks of competent integrator tuning, the same systems typically settle between zero point three and three false positives per camera per day. Verkada and Avigilon perform near the lower end in tidy environments; Axis, Hanwha and Genetec perform consistently across mixed conditions; Hikvision and partner-dependent stacks vary more widely. Vendors quoting rates below zero point one are describing controlled laboratory or indoor retail conditions, not industrial outdoor reality.

Who is best for industrial deployments?

For permanent industrial sites with serious OT segmentation requirements, Axis, Hanwha, Genetec and Milestone form the core, with Avigilon credible inside Motorola-aligned environments. The decisive criteria are edge resilience when connectivity drops, alignment with IEC 62443-4-1 and 4-2, false-positive performance under harsh conditions, and integration with plant leitsystems. Cloud-native vendors fit a narrower set of industrial scenarios. For temporary construction sites the question shifts toward integrated hardware-software platforms rather than software alone. The book BOSWAU + KNAUER. From Building to Security Technology develops that distinction in chapters seven through nine.