Sell human and vehicle classification first, then line-crossing and area alerts, counting where a client wants numbers, and plate recognition for gates and car parks only; keep face recognition out of standard offers. Before any sale, run a 14-day false-alert test and finish the privacy checks: purpose, signage, retention, access and, where it applies, an impact assessment.
Every accuracy claim a vendor makes is unverified until you test it on the client's own scene. Hikvision's own AcuSense page concedes there is no one-size-fits-all solution for video analytics, and it gives no detection rates or test method. That is your business case: the product you sell is tuning and proof, not the algorithm.
Features and prices below were read from vendor pages in October 2026. Anything a vendor says about accuracy is marked as a claim, and the sample numbers (alert logs, walk tests, margins) are illustrative, built to show the method.
The analytics you can sell, and what each one does
Names differ by brand, but the useful analytics fall into a few families. The right-hand column is the one clients rarely ask about and installers pay for later.
| Analytic | What it does | Where it fits | Main watch-out |
|---|---|---|---|
| Human and vehicle classification | Sorts moving objects into humans, vehicles and other, so animals, leaves and rain can be ignored. Axis, Hikvision AcuSense and UniFi Protect all list it | Every site; it is the base layer for alerts | Reduction in false alerts is a vendor claim until tested |
| Object in area, line crossing | Alerts when a human or vehicle enters a drawn zone or crosses a line (Axis lists both) | Yards, perimeters, roofs, loading areas | Zones drawn carelessly trigger on roads and hedges |
| Time in area (loitering) | Alerts when someone stays in a zone longer than a set time (Axis lists it) | Shop fronts, car parks, loading bays | Customers waiting for a lift or taxi look like loiterers |
| Counting and occupancy | Counts people crossing a line or present in an area (Axis lists crossline counting and occupancy in area) | Shops, venues, offices | Needs a privacy notice; doors and groups skew counts |
| Number plate recognition | Reads plates at entrances (UniFi lists plate recognition and says it needs an AI or G6 series camera) | Gates and car parks | Plates can be personal data; needs the right angle, light and a retention rule |
| Face recognition | Matches faces against a list (UniFi lists it, on AI or G6 series cameras) | Rarely suitable for a small client | Biometric data; generally prohibited unless an exception applies |
| Audio classification | Flags sounds such as glass breaking (UniFi lists it) | Specific, justified sites | Audio recording has stricter rules; leave it off by default |
| Tailgating and hard-hat detection | Axis lists tailgating detection (marked new) and hard-hat detection (marked beta) | Controlled entrances, building sites | Beta features, and they involve monitoring staff |
The wording in the table follows each vendor's own page and is not a ranking. An analytic being listed is not evidence that it works on a given scene; that is what the tests below are for.
From "I want AI alerts" to a written scope
A client's first request is rarely specific enough to quote. A typical one: "I want AI that tells me when someone is in my yard." Turn it into answers before you price anything.
| Question to ask | Typical answer | What it decides |
|---|---|---|
| When does it matter? | After 7 pm and at weekends | Schedule: analytics armed out of hours only |
| Who receives the alert? | The owner's phone, and nobody else | Whether a monitoring centre is needed |
| What happens when it fires? | "I look at the clip, then phone a neighbour" | Clip quality and speed matter more than extra analytics |
| What counts as an intruder? | A person; not a van delivering | Human-only alerts, or a zone for deliveries |
| What are you worried about? | Theft of pallets from the far fence | Camera positions and the verified distance |
| Who gets filmed? | Staff by day, a public lane at the edge | Privacy masks, signage and the purpose statement |
The scope that comes out of that table is short: human-only detection with line crossing on the far fence, armed out of hours, alerts to the owner's phone, tested for 14 nights, with the lane masked. Notice what is absent. There is no face matching and no plate reading, because nothing the client said needs them. Most analytics quotes get smaller and cheaper, and easier to defend, after this conversation.
Where the analytics come from: built in, licensed or cloud
How the analytics are priced decides where your margin sits, so work out the model before you quote. Four shapes turned up in the pages read for this tutorial.
| Model | Example, as published | What it means for your margin |
|---|---|---|
| Built into the camera | Axis says Object Analytics comes preinstalled on compatible Axis network cameras at no extra cost and cannot be downloaded separately | No licence to resell; earn from hardware, commissioning and service |
| Local console, no licence | UniFi Protect's page says licensing is still $0 and that AI processing runs on the local console; face and plate recognition need an AI or G6 series camera | No licence margin; the console and cameras are the product |
| Feature on select devices | Hikvision lists AcuSense on selected cameras, PTZ cameras and recorders; the page states no pricing or licensing | Ask your distributor what, if anything, is charged per channel |
| Cloud-managed, hardware plus licence | Rhombus lists MSRPs such as $399 of hardware plus a $149 licence for one mini dome, says cameras need an ongoing licence, and states that AI features and cloud archiving are not included in its Professional licence | Recurring revenue on licences, but check which licence tier carries the analytics you sold |
Two practical points follow. First, a "free" built-in feature still costs you time, because it must be set up and tuned per camera. Second, the headline licence price on a cloud page may not include the AI feature the client asked for, as the Rhombus page itself warns, so confirm the tier in writing before you put a price on a quote. Some manufacturers also face restrictions on certain public-sector or government-funded work in some markets. Check what your clients are allowed to buy before specifying a brand, and ask your distributor for current guidance.
Reading a vendor's accuracy claim without being fooled
Marketing pages talk about minimising false alarms triggered by animals, falling leaves and heavy rain. That describes the aim. It is not a result. Neither the Axis page nor the Hikvision page I read gives a detection rate, the lighting or weather conditions it applies to, or a test method. Hikvision's page adds that its Self-Learning Analytics lets users mark recurring false alarms so the system reduces similar ones, which is a polite way of saying the out-of-the-box result may not fit the scene.
Turn any claim into questions before you repeat it to a client:
- Compared with what? Fewer false alarms than plain motion detection is a low bar.
- Under what conditions? Lighting, rain, distance, mounting height and camera angle all change results.
- At what sensitivity? Any system can show a low false-alert rate if it is set to miss real events.
- Who ran the test, and on which firmware version?
- What counts as a hit? A person detected once in a walk across the scene, or every frame?
If the vendor cannot answer, treat the claim as unverified and say so in your quote. A line such as "analytics reduce, but do not eliminate, false alerts; performance is verified on site as described in the test sheet" protects you, and it is true.
A tuning routine to run on every camera
These steps are general practice, so check each manufacturer's own installation guidance for mounting heights and limits. The order matters: fix the physical set-up before touching software.
- Clean the lens and check for spider webs, glare from nearby lights and anything that moves in the wind within the view.
- Check the mounting height and angle against the manufacturer's guidance, and confirm the camera sees people from the front or side rather than straight down.
- Set the schedule so the analytic is armed only when the client wants alerts.
- Draw zones and lines to cover the real boundary, leaving out roads, public paths and trees.
- Choose the classes to alert on, normally humans and vehicles, and leave "other" off.
- Set the minimum object size so distant small movements do not trigger.
- Start sensitivity in the middle, run the first three nights, and read every alert.
- Adjust one setting at a time and note the change, so you know what fixed what.
- Where the system offers a way to mark false alarms, use it, as Hikvision's page describes for its self-learning feature.
- Photograph the final settings for the handover pack.
A 14-day false-alert test you can run before the sale
The test is the thing you sell. Put one representative camera on the scene, switch on the analytic you propose and record every alert for 14 nights. Mark each as true or false and note the cause. The log below shows how a typical outdoor yard might read before tuning.
| Cause of alert | Alerts in 14 nights | Fix to try |
|---|---|---|
| Spiders and insects near the lens at night | 9 | Clean the lens, reposition away from webs, add cleaning to the service plan |
| Headlights sweeping across the yard from the road | 4 | Exclude the road side with a detection zone |
| Wind-blown hedge in heavy rain | 3 | Mask the hedge; alert on humans only |
| A cat crossing the yard | 2 | Alert on humans and vehicles only |
| False alerts | 18 | |
| True events (two delivery drivers, one trespasser) | 3 | All three detected |
That is 18 false alerts over 14 nights, or 1.3 a night. After the fixes, a second 14-night run might show four false alerts, or 0.29 a night, with both staged intrusions (a colleague crossing the fence line on two of the nights) detected. The numbers matter less than the habit: you now have a before and after, a list of causes and a tuned configuration, which is more evidence than the vendor's own page offers.
Agree the pass mark with the client in writing before the test starts. A reasonable starting point, which you should adjust, is no more than one false alert per camera per night on average over 14 nights, with every staged event detected. A site that sends alerts to a monitoring centre needs a tighter figure, because each false alert costs someone time. Camera placement affects all of this, so planning camera placement from a floor plan belongs before the test, not after it.
Walk tests: proving detection at the distances in the quote
False alerts are half the story; the other half is missed events. A walk test is a person crossing the scene on a set route while you note whether the analytic fires. Ten walks per cell is enough to see a pattern. Here are sample results from a yard camera.
| Distance from camera | Daylight, walks detected of 10 | Night, walks detected of 10 |
|---|---|---|
| 10 m | 10 | 10 |
| 20 m | 10 | 9 |
| 30 m | 8 | 5 |
The reading is plain. At 20 m the system passes a 9-of-10 target at night; at 30 m it does not. So the quote says "human detection verified to 20 m at night", and a second camera covers the far end of the yard. That sentence is worth more to a client than any brochure percentage, because it states what you tested and what you did not.
Resale margin per camera, with tuning and call-outs counted
Take a 16-camera warehouse. You sell analytics as a service at $9 a camera a month. Your cost is a licence at $5 a camera a month, which is an assumption to replace with your distributor's price.
| Line | Working | Per year |
|---|---|---|
| Client pays | 16 x $9 x 12 | $1,728 |
| Your licence cost | 16 x $5 x 12 | $960 |
| Margin before service | $1,728 - $960 | $768 (44%) |
| Two false-alert call-outs at $120 each | 2 x $120 | $240 |
| Margin after call-outs | $768 - $240 | $528 (31%) |
Now the two traps. Commissioning takes time: at 1.5 hours a camera and $60 an hour of internal cost, 16 cameras cost $1,440 to tune. Billed at $110 a camera, that is $1,760 and a small margin. Folded into the monthly fee instead, it eats nearly two years of the licence margin. The second trap is unlimited call-outs. On a difficult site with eight call-outs a year, the cost is $960, which is more than the $768 margin, so the contract loses $192. Include a set number of call-outs and charge for the rest.
The built-in alternative has no licence margin, but there is still money in service. A health-check plan at $5 a camera a month brings in $960 a year across 16 cameras. If a remote check takes 10 minutes a camera each quarter, that is 40 minutes, or about $40 at $60 an hour, per camera per year; across 16 cameras the cost is $640 and the margin $320, or 33%. Either way the margin comes from commissioning and care. CCTV maintenance contracts with remote health checks and renewals shows how to package it.
Privacy and data-protection checks for each analytic
The client is normally the controller of the footage, but your configuration decides much of what is collected. These checks are practical, not legal advice. The rules differ between the places your clients operate, so the client's data-protection adviser or solicitor should approve the paperwork.
| Check | Why it matters | Evidence to keep |
|---|---|---|
| Purpose written down | Analytics should serve a stated aim such as site security, not staff performance | One paragraph in the client file |
| Impact assessment | Where data-protection law requires impact assessments, large-scale systematic monitoring of a publicly accessible area is a standard trigger, and new technology such as AI analytics can be another | The assessment, or a note why it does not apply |
| Biometrics | Data-protection law commonly treats biometric data used to uniquely identify a person as a special category, prohibited unless a narrow exception applies | A decision not to enable face matching, or legal sign-off |
| Staff | Some AI laws ban systems that infer people's emotions at work, one of them since February 2025, and analytics aimed at staff raise consent questions too | A list of analytics switched off |
| Signage | People should know about cameras and what the system does, such as counting or reading plates | Photos of signs in place; wording checks with AI |
| Retention | Law sets no single figure; the client must be able to justify the period they choose | The setting, screenshotted |
| Access | Footage and cloud accounts need named users, strong passwords and two-step sign-in | The user list and who approved it |
| Audio | Recording sound is often treated more strictly than recording video | Microphones off, or a recorded justification |
| Privacy masks | Cameras should not film neighbouring windows or gardens | Mask screenshots at handover |
| Data location | Cloud vendors store footage and metadata somewhere; the client should know where | The vendor's processing terms |
| Subject access | People can ask for footage of themselves; you should know how to export and blur others | A one-page procedure |
For a retail client, footfall cameras and privacy in a retail store goes into counting in depth. For a client with staff on site, AI CCTV in a factory, with detection, cost and staff consent covers the consent side.
Which analytic suits which client
A corner shop. Person detection on the entrance and till cameras, with alerts to the owner's phone out of hours. Skip face matching. If the owner wants customer counts, add a notice and keep the counts anonymous.
A warehouse with a yard. Human and vehicle classification with line crossing on the perimeter, active only out of hours, plus a time-in-area rule at the loading gate. This is where the 14-night test earns its fee, because yards have lights, hedges, wildlife and weather.
A small office car park. Plate recognition at the barrier, if the client has a stated reason such as managing permits, with a short retention period and a signed privacy note. Without that reason, ordinary human and vehicle detection is enough.
A building site. Night-time intrusion with human-only alerts and the camera mounted away from webs and lights. Hard-hat detection is listed as beta by the vendor that offers it, so do not sell it as proven, and tell the client it involves monitoring workers.
A nursery or school. Keep analytics to out-of-hours intrusion. Anything involving children needs the client's data-protection adviser before you quote.
Seven mistakes installers make with analytics
- Selling AI as a replacement for a monitored alarm. It reduces alert noise. It does not guarantee a response.
- Leaving defaults on. Audio, face matching and "smart" extras left enabled because the camera has them.
- Maxing sensitivity. A system set to catch everything produces alert fatigue, and the client mutes it.
- Ignoring mounting. A camera too high or at the wrong angle misclassifies people; no setting fixes that.
- No call-out limit. The margin sum above shows how one bad site can wipe out a year.
- No retention policy. Footage stored indefinitely because nobody set the period.
- Accounts in the installer's name. Cloud and recorder logins should belong to the client, with you as an invited user. Otherwise a handover becomes a dispute.
Quote wording that states what you tested
A quote for the yard camera in the earlier examples could carry these clauses. Adapt them with the client's adviser.
- Scope. Human and vehicle detection with line crossing on the far fence, armed 7 pm to 7 am and at weekends, alerts to two named phones.
- Verified performance. Detection verified by walk test to 20 m at night on the date of commissioning. Beyond 20 m, detection is not guaranteed.
- False alerts. Tuned until the average is no more than one false alert per camera per night over 14 nights. Analytics reduce but do not eliminate false alerts and can miss events.
- Call-outs. Two false-alert call-outs a year are included; further visits are charged at the agreed rate.
- Client responsibilities. Keep lenses and the view clear, trim vegetation, and tell us about changes to lighting or layout.
- Exclusions. Face recognition, audio recording and any use for monitoring staff performance are not included.
- Data. Retention set at the client's chosen period; accounts held in the client's name with the installer as an invited user.
What goes in the handover pack
Hand the client a short record of what you delivered: the zone maps, sensitivity levels and schedules for every camera; the false-alert and walk-test results; a statement that analytics reduce but do not eliminate false alerts and can miss events; the call-out policy; the retention setting; the data location; the controller and processor roles; and photos of the signs. A CCTV commissioning, signage and handover checklist shows the format. A pack like this turns an analytics sale into a defensible one, and it is what the client's adviser will ask for first.
Questions installers ask about selling analytics
Should I sell face recognition as an add-on?
Not as a standard offer. Many data-protection laws treat biometric data used to identify a person as a special category that may not be processed unless a narrow exception applies, and a client will usually need an impact assessment and legal advice first. Most small sites need only person and vehicle detection.
Is built-in analytics better than a paid licence?
Neither is better by default. Built-in analytics, such as Axis Object Analytics on compatible cameras, costs nothing extra but gives you no licence margin. Licensed or cloud systems add recurring revenue and recurring support. Choose by what the client needs and what you can maintain.
How many false alerts are acceptable?
There is no standard figure, so agree one in writing. A workable starting point is no more than one false alert per camera per night, averaged over 14 nights, with every staged event detected. Monitored systems usually need a stricter target because each alert costs someone time.
Can I promise a client that AI will stop false alarms?
No. Vendors publish claims of fewer false alarms, but the ones read for this tutorial give no test method, and one concedes there is no one-size-fits-all solution. Promise a tested, tuned system with stated limits and a call-out policy, not a result.
Do analytics change what the signage must say?
Often yes. If the system counts people, reads plates or records audio, the notice and the client's privacy paperwork should say so, with the purpose and a contact. Have the client's data-protection adviser approve the wording before you fix the signs.
Further reads
- Convenience Store AI Cameras: Cost and the Privacy Checks First — Costs and privacy checks for a single small shop.
- CCTV Installation Quote Template: A Worked AI-Drafted Walk-Round Quote — A walk-round quote template with analytics as a line item.
- Alarm Company Software: Monitoring, Service and Billing, and Where AI Helps — Monitoring and billing software for security firms.
- AI Security Risks for Small Businesses and How to Close Them — Account and device risks to close on every install.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: AXIS Object Analytics page (axis.com), UniFi Protect software page (ui.com), Hikvision AcuSense page (hikvision.com), Rhombus pricing page (rhombus.com), the published text of data-protection rules on biometric data and impact assessments, and of AI rules on emotion recognition at work. Features and prices read in October 2026; vendor accuracy claims are unverified.