Edge AI Business Risks: What Goes Wrong and Who Should Skip It

The main risks of selling local AI installs are the support burden, models and hardware that change under you, licence traps, liability for wrong answers and scope creep. None of them is a reason to avoid the work, but each needs a line in your contract and your price. Skip this if you want passive income, cannot take a client's call during the working day, or do not want to read licence terms and privacy law.

This is the last guide in the edge AI series.

What happens when the box stops working?

The client calls you. A box in a clinic that stops answering at 9am on Monday is your problem until it works, whatever the cause: a power cut, a Windows update on the network, an expired certificate, a cleaner who unplugged it.

Plan for it before the first install:

• Monitoring. A simple uptime check that messages you before the client notices.
• A UPS. A small uninterruptible power supply prevents most corrupted-disk calls.
• A spare. Your demo box can stand in while a fault is fixed.
• Response times in writing. "Next business day" is honest for a side business. "Within the hour" is a promise you will break.

If you have a full-time job, be honest with yourself about whether you can take that Monday call. Count support hours as part of the job when you set the retainer.

What changes under you?

Models. Open models move fast. Alibaba released Qwen3.5 in February 2026, Qwen3.6 in April and Qwen3.8 in August, according to its GitHub repository, and Google released Gemma 4 in April. Clients will read about each one. Test updates on a copy and switch only with sign-off, as part of the retainer.

Hardware prices. NVIDIA raised Jetson prices by 33% to 101% in July 2026, and Raspberry Pi raised board prices four times in under a year. Date every quote and do not hold stock.

Cloud prices. The other side of the comparison moves too, often downwards. Anthropic's newest small model, Haiku 5.5, is listed at $0.10 per million input tokens for shorter prompts, a tenth of Haiku 4.5's $1. A box sold mainly to save on API costs can lose that argument a year later. Sell on privacy, control and offline working first, and cost second.

Services. Cloud services close as well: Amazon shut down Lookout for Vision on October 31, 2025. That works in your favour when you pitch a box the client owns, as long as the software on it is open and maintained.

Which licences can catch you out?

• Ultralytics YOLO is AGPL-3.0. Closed commercial use, including internal tools and edge devices, needs a paid Enterprise License, priced by quote. Find out before you quote a vision job.
• Open WebUI forbids removing its branding above 50 users in a rolling 30 days without an enterprise licence or permission. Do not rebrand it for a client.
• Model licences differ. gpt-oss, Qwen and Gemma 4 are Apache 2.0; Gemma 3 and Llama have their own terms. Keep a record of which model each client runs and under which licence.

Details are in the local LLM stack.

Who is liable when the AI gets it wrong?

Possibly you, depending on your contract and local law. A model that misquotes a lease clause or a dosage protocol can cause real harm, and the client will look for someone to blame.

Protect yourself and the client:

1. Say what it is. The contract and the handover note state that the system can make mistakes and that a qualified person checks anything relied on.
2. Design for checking. Answers show their sources. Staff are trained to click through.
3. Cap your liability in the contract, usually to the fees paid. Have a lawyer draft your standard terms once.
4. Insure. Insureon reports a median of $75 a month for technology errors and omissions cover for IT consultants and $128 a month for cyber insurance.
5. Sign the data agreements. A data processing agreement under GDPR or the PDPA, or a business associate agreement under HIPAA, if you can reach personal or patient data. See private AI for offices, clinics and law firms.

How does scope creep start?

With a favour. "While you're here, could you look at the printer?" Then "could it also read our emails?" Then a retainer that covers three hours a month is costing you ten.

• Write what the retainer includes, as a short list, and what it does not.
• Track your hours from the first month, even if you never bill by the hour.
• Price new work as a small project with its own fee.
• Review the retainer every 12 months against the hours you actually spent.

Managed IT providers know this squeeze. Kaseya's 2026 survey found 71% named winning new customers as their biggest challenge, and the share reporting customer spend above $25,000 a year fell from 75% to 41%. Margins in IT services are thin, and unpaid hours are where they go.

What about the camera and privacy risks?

Camera projects carry the most legal risk in this series. Face recognition in particular can bring damages of $1,000 to $5,000 per violation under Illinois's BIPA, and some uses are banned under the EU AI Act. Stick to counting and detection that stores no faces, and refer anything else to a lawyer. The rules by country are in AI cameras for shops and restaurants.

Who should not start an edge AI business?

Skip it if any of these is true:

• You want passive income. This is a service business. Every client is a multi-year support relationship.
• You cannot answer calls during the working day, and cannot arrange for someone who can.
• You dislike networks and hardware. Half the problems are not the AI. They are Wi-Fi, power, updates and permissions.
• You will not read licences and privacy rules. The client relies on you to know which model, which licence and which law.
• You need income next month. Landing the first paying install can take months of demos and pilots.

It suits people who already fix computers for others, work in IT or operations, or know one industry well, and who want their own clients badly enough to build slowly and support them properly. If that is you, the hub has the full series in order and the local AI vs cloud calculator works out what one install could pay.