3S Market 探討報導
本文受眾：安控設備製造商、系統整合商、工程商、通路商、保全業者，以及正在規劃 AI 安控布局的中高階經營者。
AI 幾乎已經成為今天安控產業最熱門的名詞。攝影機有 AI、門禁有 AI、NVR 有 AI、VMS 有 AI，展覽、論壇、產品發表會也幾乎言必稱 AI。站在市場氣氛來看，AI 安控似乎已經全面啟動；但如果換一個角度問：這些 AI 到底有多少真正進入日常安全管理？多少設備只是具備 AI 運算能力，功能卻沒有被啟用？多少製造商真正知道自己未來五年的 AI 產品要往哪裡走？答案立刻就沒有那麼簡單。
所以，現在投入 AI 安控究竟太早還是太晚？這篇文章不準備給一個標準答案。市場本來就是動態的，不同國家、不同場域、不同產品線，成熟速度都不一樣。真正值得觀察的，不是「AI 安控什麼時候會到來」，而是：它已經在哪些地方開始，又有哪些地方可能接著跟進？
要看清楚這個問題，不能只看 AI 攝影機出貨，更不能只看市場聲量，而要同時觀察 技術、產品、產業、市場 四個面向的成熟度。
技術先跑了多遠？「有 AI」與「AI 能做多少」不是同一件事
AI 安控首先要面對一個很基本的問題：今天市場上標示 Deep Learning、AI Analytics 的設備，能力到底是不是同一個水平？
答案顯然不是。
一台攝影機能做人、車分類，降低樹影、動物造成的誤報，也可以被稱為 AI；另一套系統則可能進一步辨識人員屬性、車種車色、PPE 穿戴、異常滯留、群聚、周界入侵，甚至透過跨鏡搜尋、自然語言或 VLM 找到特定事件。兩者同樣貼著 AI 標籤，實際能解決的安全問題卻完全不同。
目前市場研究普遍可以看到一個方向：具備 Deep Learning Video Analytics 的 Embedded Chipset 正快速進入新出貨的 Network Camera，AI 運算能力往 Mainstream 產品下放的趨勢已相當明顯。但這類數字比較適合說明「設備正在 AI 化」，不能直接等同「AI 安控已大量使用」。
今天真正需要看的，是 AI SoC/NPU 能跑多大的模型、能不能同時執行多項分析、在高解析度與低照度下還能維持多少效能，以及 Model、Firmware、SDK、VMS、Edge Server、Cloud 之間能不能協同。也因此，未來 AI 安控產品不能再只用解析度、FPS、WDR 或儲存容量來比較。
AI 技術已經不是「能不能做」，而是「能做到多深、能不能長期可靠地做」。
產品成熟度：買到 AI 晶片，不等於做出了 AI 安控產品
這可能是製造商接下來最大的分水嶺。
過去安控產品的開發，硬體平台、影像品質、韌體、功能、成本控制與量產能力非常重要；進入 AI 之後，這些能力仍然需要，但後面又多出模型、資料、AI Framework、運算資源配置、演算法調校、應用軟體、模型更新、第三方 API、資安與隱私等長期工作。
所以真正值得問製造商的，不是「你有沒有 AI Camera？」而是：
你知不知道未來準備用 AI 解決哪些安全問題？
如果只是挑一顆具 NPU 的晶片，套入 Chip Vendor 提供的人車分類模型，再把產品名稱加上 AI，這當然也是商品化的一種方式；但長期來看，可能只是把傳統組裝能力延伸成「AI 組裝」。
真正的 AI 產品規劃應該從應用往回走：哪一個場域？哪一種事件？需要辨識什麼？多少誤報可以接受？事件之後要觸發什麼程序？再決定 Camera、Access Control、Sensor、Edge Box、Server、Cloud 與 Software 如何搭配。
因此，AI 安控產品的成熟度不能只看「有多少 AI SKU」，而要看產品能不能持續升級、能不能整合、能不能被複製到不同專案，更重要的是，AI 功能能不能真的變成業主願意付錢的應用。
產業像一座彈珠台：產品推出去之後，才真正開始考驗
如果把產品比喻成一顆彈珠，產業就像一座大型彈珠台。
一顆 AI 產品被打出去之後，不會筆直從製造商滾到業主手上。中間會碰到原廠、代理商、盤商、經銷商、工程商、SI、VMS、門禁平台、IT 部門、OT 部門、顧問、標案規格、法規、資安要求、預算與既有系統。
有的彈珠技術很好，卻卡在通路根本不知道怎麼賣；有的 AI 功能很強，SI 卻不知道如何驗證 Accuracy；有的專案 PoC 很漂亮，但到了正式部署，誤報、模型維護與現場網路條件卻讓系統卡住。相反地，也有一些技術並不特別新，卻因政策、法規、重大事故或大型業主要求，突然快速往下滾。
這就是為什麼「產品成熟」從來不等於「產業成熟」。
AI 安控真正要進入規模市場，SI、工程商與通路也必須知道如何賣 Solution，而不是繼續用「幾支攝影機、幾台 NVR、多少 TB」來做報價。Proposal 要怎麼寫、AI 成效怎麼驗證、誤報由誰處理、模型誰來調、維護費如何收、Software License 或 Subscription 是否能成立，全部都是產業成熟度的一部分。
市場愈熱，產業愈容易被題材帶著跑；但真正決定這顆彈珠能不能滾到底的，不是聲量，而是整個產業有沒有能力把它接住。
市場真的開始了多少？安全專業人員的回饋，可能比出貨數字更值得看
如果只從設備供應端看，AI 安控似乎已經跑得很快；但是從 End User 的實際使用看，畫面就不一樣。
ASIS International 2025 年針對安全專業人士的研究中，約 26% 的受訪者表示其組織已把進階生物辨識或臉部辨識加入 Access Control，23% 使用 AI 做監控物件偵測或移動追蹤，19% 使用 AI 偵測異常狀況並發出告警；另一方面，仍有 43% 表示沒有在 Security Function 使用 AI。ASIS 也直接指出，目前沒有任何單一 AI Security Application 擁有特別高的 Adoption Rate。
這組數字值得注意，不只是因為它比較保守，而是受訪者主要來自實際從事安全管理工作的專業族群。它讓我們看到一個很有意思的落差：AI 能力可能已經進入設備，但真正用進安全程序的比例，仍遠低於設備 AI 化的速度。
而 ASIS 調查中，目前較高的實際應用不是最炫目的 Generative AI，而是生物辨識與門禁。這其實很合理。因為 Access Control 本來就建立在 Identity、Authority、Event 與 Traceability 上，AI 並不是憑空創造一套流程，而是在既有安全程序上提升辨識與判斷能力。
這也提醒我們：AI 安控不能只從影像監控看。
門禁、生物辨識、智慧鎖、入侵偵測、周界、LPR、Radar、Thermal、Sensor、VMS、Command Center，以及 Edge/Cloud AI，全部都可能是 AI 安控的一部分。
哪些場域已經開始？先看高要求市場，而不是只看「量大」
要知道 AI 安控真正走到哪裡，高安全、高要求場域比一般普銷市場更具有指標性。
政府、國防、警政、交通、醫療，是第一組值得觀察的市場。這些場域本來就有較高的安全風險、較複雜的事件量，也比較有能力建立專業 Security Operation。但真正要看的不是有沒有 AI 標案，而是到底有多少 AI 功能真正被啟用，而且進入日常 SOP。
以警政為例，AI 已經可以從單純車牌辨識延伸到車輛軌跡、特定目標即時告警、車型車色分析、影像濃縮，甚至進一步和案件調查、巡防流程連結。交通也類似，從車流統計延伸到異常停留、道路事件、軌道障礙與即時告警。
但即使在這些高階場域，也不能因為找到幾個成功案例，就宣布整個市場已經成熟。真正的指標應該是：部署是否持續擴充、功能是否重複採購、AI 是否由單點辨識逐漸進入跨系統安全流程。
台灣另一個值得盯住的指標：半導體、科技廠與公用事業
如果要判斷台灣 AI 安控下一步，我認為半導體與科技廠房尤其值得持續觀察。
不是因為它們一定已經大量部署最先進 AI，而是它們具備最完整的壓力測試條件：高價值資產、營業秘密、嚴格人員權限、Cleanroom、Visitor、Contractor、物流、物料、工安、環境安全，以及 IT/OT 管理。
如果 AI 安控能在這類場域從單一 PPE、車牌或人車辨識，進一步串聯 Identity、Authority、Asset、Behavior、Perimeter、Traceability 與 Incident Response，就代表市場真的開始跨過「AI Feature」走向「AI Security Solution」。
目前台灣智慧工安的公開案例已經可以看到 PPE、火焰、煙霧、高溫、氣體洩漏、門禁身分驗證與危險區域偵測，也出現既有 Camera 加 AI Platform、Edge AI、熱成像以及 Cloud Subscription 等不同導入方式。這些訊號值得觀察，但離「整個製造業全面 AI 化」顯然還有距離。
因此，半導體、科技製造、公用事業不是用來證明 AI 安控「已經到了」，而是很適合拿來觀察下一個成熟度拐點會不會從這些高要求場域長出來。
學校、商辦、社區何時跟進，才是另一個重要訊號
高階市場能買得起 AI，不代表 AI 已經走向普及。
真正值得注意的下一層，是學校、一般商辦、社區建築。這些市場沒有龐大的 Security Team，也不可能為每一套 AI 建立專門 Data Scientist 或模型維護團隊，因此產品必須變得夠便宜、夠穩、夠簡單，而且解決的問題非常直接。
陌生人進入、尾隨、異常開門、停車場事件、公共區域異常、跌倒、緊急求救、設備與空間管理，如果開始大量透過標準化 AI 模組部署，而不是每一案都重新做 PoC，這才會是一個很重要的 Growth Signal。
零售則需要另外看。Walmart、Costco、Amazon 這種大型零售與一般商店都叫 Retail，但 AI 使用深度完全不同。大型連鎖可能同時做 Loss Prevention、POS 關聯、倉儲、停車、員工安全與營運分析；小型店可能只有基本人流或人車辨識。因此零售可以是很大的「量」，卻不一定是判斷整體 AI 安控成熟度最好的指標。
第一桶金從哪裡來？也許不必急著猜場域
談到最後，「第一桶金」很容易讓人期待一個答案：是警政？交通？半導體？醫療？還是零售？
但市場可能沒有這麼整齊。
比「哪個場域先爆發」更值得觀察的，是哪一種安全問題已經成熟到可以讓 AI 穩定創造價值。
Identity/Biometrics、LPR、Perimeter、PPE、False Alarm Reduction、Object Search、Alarm Verification，現在之所以比較容易找到實際應用，不一定是因為它們最先進，而是因為問題明確、結果容易驗證，而且原本就存在安全預算與作業程序。
換句話說，第一桶金不一定埋在最熱門的 AI，而可能埋在最成熟、最容易證明價值的安全問題裡。
對製造商來說，問題是要不要繼續賣一顆「AI 彈珠」，還是開始理解這顆彈珠進入產業之後會撞到哪些擋板；對 SI 與工程商來說，機會則可能出現在如何把 AI 從 Feature 接到 Workflow；對通路與服務業者而言，真正的新生意也許不是再多賣一台設備，而是模型、軟體、維運與持續服務。
結語：真正值得看的，是哪些地方已經開始，又有哪些地方可能接著跟進
十年前談監控雲端化時，也有人一直問：「Cloud 到底什麼時候真正到來？」
今天回頭看，Cloud 早就在某些市場大量發生，但 On-premise、NVR、Local Server 同樣沒有消失。市場從來不是在某一天跨過一條線，然後所有人同時換代。
AI 安控也一樣。
所以現在投入 AI 安控究竟太早還是太晚？也許真正值得產業持續追蹤的，不是一個 Yes/No，而是四條線正在怎麼移動：
技術做到哪裡、產品跟到哪裡、產業接得住多少、市場又真正用了多少。
市場一定會有風向，也一定會有炒作；但是最後決定這個題材能炒幾番、能不能形成真正長期市場的，仍然是技術、產品與產業成熟度。
AI 安控已經在哪些地方開始，又有哪些地方可能接著跟進？
這個問題，恐怕比預測「哪一年全面進入成長期」更值得我們持續追下去。
English text version
Is It Too Early or Too Late to Invest in AI Security? Where Will the First Real Profits Come From?
Target readers: security equipment manufacturers, system integrators, contractors, distributors, security service providers, and mid- to senior-level executives planning their AI security strategies.
AI has become one of the hottest topics in today’s security industry. Cameras have AI. Access control has AI. NVRs have AI. VMS platforms have AI. Trade shows, conferences, product launches, and marketing campaigns are now filled with AI messaging.
From the market’s point of view, it can easily look as if AI security has already taken off.
But if we ask a different set of questions, the picture becomes much less straightforward.
How much of this AI has actually entered day-to-day security operations? How many devices are merely AI-capable, while their AI functions remain unused? How many manufacturers truly know where their AI product strategies are heading over the next five years?
The answers are far less obvious.
So is it too early or too late to invest in AI security?
This article does not attempt to provide a standard answer. Markets are dynamic. Different countries, vertical markets, applications, and product categories mature at different speeds.
What is really worth observing is not “When will AI security arrive?”
The more useful question is:
Where has it already started, and which markets are likely to follow next?
To understand that, we cannot look only at AI camera shipments, and we certainly cannot rely only on market hype.
We need to examine AI security through four dimensions of maturity:
technology, products, industry, and market adoption.
How Far Has the Technology Really Come? “Having AI” Is Not the Same as “How Much AI Can Actually Do”
AI security first faces a very basic question:
Are all products labeled “Deep Learning” or “AI Analytics” operating at the same level?
Clearly, they are not.
A camera that distinguishes humans and vehicles from trees, shadows, or animals can already be marketed as an AI camera.
Another system may go much further: identifying personal attributes, vehicle type and color, PPE compliance, loitering, crowding, perimeter intrusion, or even enabling cross-camera search, natural-language search, and VLM-based contextual understanding.
Both may carry the same AI label, but the security problems they can solve are completely different.
Current market research generally points in one direction: embedded chipsets capable of Deep Learning Video Analytics are rapidly moving into newly shipped network cameras, and AI computing capability is moving from premium products toward the mainstream.
But this kind of data is more useful for proving that devices are becoming AI-capable.
It does not automatically prove that AI security itself is being widely used.
What matters today is whether an AI SoC or NPU can run sufficiently sophisticated models, whether it can execute multiple analytics simultaneously, whether it can maintain performance under high resolution and low-light conditions, and whether the model, firmware, SDK, VMS, edge server, and cloud architecture can work together effectively.
This means AI security products can no longer be compared only by resolution, FPS, WDR, or storage capacity.
The technology question is no longer simply “Can AI do it?”
It is increasingly becoming:
“How deeply can AI do it, and can it do it reliably over time?”
Product Maturity: Buying an AI Chip Does Not Mean You Have Built an AI Security Product
This may become one of the biggest dividing lines among security manufacturers.
In the traditional security product era, hardware platforms, image quality, firmware, features, cost control, and mass-production capability were critical.
They still are.
But AI adds another long-term layer of work: models, data, AI frameworks, computing-resource allocation, algorithm optimization, application software, model updates, third-party APIs, cybersecurity, and privacy.
So the real question for a manufacturer is no longer:
“Do you have an AI camera?”
The more important question is:
“Do you know which security problems you intend to solve with AI?”
If a company simply selects an NPU-enabled chipset, uses the chip vendor’s ready-made human and vehicle classification models, adds “AI” to the product name, and launches it, that is certainly one way to commercialize AI.
But over the long term, it may simply extend traditional assembly capability into AI assembly.
A true AI product strategy should work backward from the application.
Which environment?
What type of incident?
What needs to be identified?
What level of false alarms is acceptable?
What action should be triggered after the event is detected?
Only then should the manufacturer decide how cameras, access control, sensors, edge boxes, servers, cloud platforms, and software should be combined.
Therefore, AI security product maturity should not be measured simply by the number of AI SKUs in a product portfolio.
The real questions are whether the product can continue to evolve, be upgraded, be integrated, and be replicated across different projects.
Most importantly:
Can the AI capability become an application that an end user is willing to pay for?
The Industry Is Like a Pinball Machine: The Real Test Begins After the Product Is Launched
Imagine the product as a pinball.
The industry is the pinball machine.
Once an AI product is launched, it does not move directly from the manufacturer to the end user.
It hits distributors, resellers, contractors, system integrators, VMS platforms, access control platforms, IT departments, OT departments, consultants, project specifications, regulations, cybersecurity requirements, budgets, and legacy systems along the way.
Some products may be technically excellent but become stuck because the channel does not know how to sell them.
Some AI functions may be very powerful, yet the SI does not know how to validate accuracy.
Some PoCs look impressive, but after full deployment begins, false alarms, model maintenance, network conditions, and operational complexity create serious problems.
On the other hand, some technologies that are not especially new may suddenly move quickly because of regulation, policy, major accidents, or the requirements of a large end user.
That is why product maturity never automatically equals industry maturity.
For AI security to scale, system integrators, contractors, and channels also need to know how to sell a solution, instead of continuing to quote projects simply as:
“how many cameras, how many NVRs, and how many terabytes of storage.”
How should the proposal be written?
How should AI effectiveness be validated?
Who is responsible for false alarms?
Who tunes the models?
How should maintenance fees be charged?
Can software licensing or subscription models be sustained?
All of these are part of industry maturity.
The hotter a market becomes, the easier it is for the industry to be pulled along by the topic.
But what ultimately determines whether the pinball reaches the bottom is not market noise.
It is whether the industry has the ability to catch it, integrate it, and move it forward.
How Much AI Security Is Actually Being Used? Feedback from Security Professionals May Tell Us More Than Shipment Numbers
If we only look at the supply side, AI security seems to be moving very quickly.
But the picture changes when we look at actual use by end users.
In the ASIS International 2025 study of security professionals, approximately 26% of respondents said their organizations had added advanced biometrics or facial recognition to access control.
Another 23% were using AI for object detection or movement tracking.
Around 19% were using AI to detect anomalous situations and generate alerts.
At the same time, 43% said they were not using AI within their security function.
ASIS also pointed out that no single AI security application had yet reached a particularly high adoption rate.
These figures are important not only because they are relatively conservative.
They matter because the respondents are closer to the people actually responsible for security management and operations.
This reveals an important gap:
AI capability may already be entering the equipment, but the proportion actually embedded into security procedures remains far lower than the pace of device-level AI adoption.
Another interesting point is that one of the more widely adopted physical security AI applications is not the most glamorous form of Generative AI.
It is biometrics and access control.
That makes sense.
Access control already operates around:
Identity, Authority, Event, and Traceability.
AI does not need to create an entirely new security process from scratch.
It can strengthen an existing process by improving identification and decision-making.
This also reminds us that AI security should never be viewed only through video surveillance.
Access control, biometrics, smart locks, intrusion detection, perimeter protection, LPR, radar, thermal imaging, sensors, VMS, command centers, edge AI, and cloud AI can all form part of AI security.
Where Has AI Security Already Started? Look First at High-Security, High-Demand Environments
If we want to understand where AI security really stands, high-security and high-demand environments may be better indicators than mass-market volume.
Government, defense, law enforcement, transportation, and healthcare are the first group worth watching.
These environments naturally involve higher security risk, more complex event volumes, and a greater ability to build professional security operations.
But the key question is not whether AI appears in procurement documents.
The real question is:
How many AI functions are actually enabled and embedded into daily SOPs?
In law enforcement, for example, AI may begin with license plate recognition and extend to vehicle trajectory analysis, real-time alerts for specific targets, vehicle type and color classification, video summarization, and eventually integration with investigation and patrol workflows.
Transportation follows a similar path: traffic counting may evolve into abnormal stopping detection, road event detection, rail obstacle detection, and real-time alerting.
But even in these high-end environments, a handful of successful projects does not prove that the entire market is mature.
The stronger indicators are:
Is deployment continuing to expand?
Are functions being purchased repeatedly?
Is AI moving from single-point detection into cross-system security workflows?
Another Key Indicator in Taiwan: Semiconductors, High-Tech Manufacturing, and Utilities
If we want to understand where Taiwan’s AI security market may go next, semiconductor fabs and high-tech manufacturing deserve especially close attention.
Not because they have necessarily deployed the most advanced AI everywhere.
But because they offer one of the toughest real-world stress tests.
These environments involve high-value assets, trade secrets, strict personnel authorization, cleanrooms, visitor management, contractors, logistics, materials, occupational safety, environmental safety, and complex IT/OT management.
If AI security in these environments can move beyond isolated PPE detection, license plate recognition, or human/vehicle classification, and begin linking:
Identity, Authority, Asset, Behavior, Perimeter, Traceability, and Incident Response,
then the market is beginning to move from AI Feature toward AI Security Solution.
Publicly available smart industrial safety cases in Taiwan already show applications such as PPE detection, flame and smoke detection, high-temperature monitoring, gas-leak detection, identity verification, and hazardous-area monitoring.
Different deployment architectures are also emerging, including existing cameras combined with AI platforms, edge AI, thermal imaging, and cloud subscription models.
These are signals worth watching.
But they still do not mean the entire manufacturing sector has become AI-driven.
That is why semiconductors, high-tech manufacturing, and utilities should not be used to prove that AI security has “arrived.”
They are more useful as indicators of whether the next maturity inflection point may emerge from these demanding environments.
When Schools, Offices, and Residential Communities Begin to Follow, That Will Be Another Important Signal
The fact that high-end markets can afford AI does not mean AI has reached broad adoption.
The next layer worth watching is schools, ordinary office buildings, and residential communities.
These markets do not have large security teams.
They cannot afford to maintain a dedicated data science team or rebuild an AI PoC for every project.
So AI products must become affordable, stable, simple, and directly useful.
Unauthorized entry, tailgating, abnormal door opening, parking incidents, unusual activity in public spaces, falls, emergency calls, equipment management, and space management are all examples of problems that can potentially be standardized.
If these applications begin to be deployed as repeatable AI modules rather than requiring a fresh PoC every time, that will become a meaningful growth signal.
Retail needs to be viewed separately.
Walmart, Costco, Amazon, and a small local shop may all be classified as “retail,” but their depth of AI usage is completely different.
A large retailer may simultaneously deploy loss prevention, POS-linked analytics, warehouse monitoring, parking management, employee safety, and operational analytics.
A small store may only use basic people counting or human/vehicle classification.
So retail may represent a very large market in terms of volume.
But it is not necessarily the best indicator of overall AI security maturity.
Where Will the First Real Profits Come From? Maybe We Should Not Rush to Guess the Vertical Market
At this point, the phrase “first pot of gold” naturally makes people ask:
Will it come from law enforcement?
Transportation?
Semiconductors?
Healthcare?
Retail?
But markets rarely develop that neatly.
Rather than asking which vertical will explode first, it may be more useful to observe:
Which security problems have already matured enough for AI to create stable, measurable value?
Applications such as:
Identity and biometrics,
LPR,
perimeter protection,
PPE detection,
false-alarm reduction,
object search,
and alarm verification
are already easier to find in real deployments.
Not necessarily because they are the most advanced AI applications.
But because the problems are clear, the results are easier to verify, and the security budget and operating procedures already exist.
In other words:
The first real profits may not be hidden in the hottest AI technology.
They may be hidden in the most mature security problems with the clearest value proposition.
For manufacturers, the question is whether they want to keep selling a single “AI pinball,” or begin understanding which obstacles it will hit once it enters the industry.
For system integrators and contractors, the opportunity may lie in connecting AI features to actual workflows.
For distributors and service providers, the new business may no longer be simply selling one more device.
It may come from models, software, maintenance, and recurring services.
Conclusion: The Real Question Is Not When AI Security Will Arrive, but Where It Has Started—and Who Will Follow Next
About a decade ago, the security industry repeatedly asked:
“When will cloud surveillance really arrive?”
Looking back today, cloud surveillance has already become significant in some markets.
At the same time, on-premise systems, NVRs, and local servers have not disappeared.
Markets do not cross a single line on one particular day and then suddenly move forward together.
AI security is the same.
So is it too early or too late to invest in AI security?
Perhaps the more useful question is not a simple Yes or No.
What the industry should continue to observe is how four different lines are moving:
How far has the technology gone?
How far have the products followed?
How much can the industry absorb and deliver?
How much is the market actually using?
Markets will always have hype.
There will always be companies trying to shape the narrative.
But what ultimately determines how many times a market theme can be promoted—and whether it can become a sustainable long-term business—is still the maturity of the technology, products, and industry itself.
Where has AI security already started, and which markets are likely to follow next?
That may be far more valuable to keep watching than trying to predict the exact year in which AI security will “enter the growth stage.”
按此 ☞ 回今日3S Market新聞首頁
0 comments:
張貼留言