When the power goes out, your AI must not need the cloud

The power goes out. The router dies. You open the solar battery app on your phone: nothing useful. The smart-home service cannot be reached. Your freezer, card terminal and alarm are on their own. Standing in the dark, you do not need a clever answer from a distant server. You need to know what still works, what must stop and how long your essentials can last.

My view is simple: an AI agent intended to help during a blackout must not need the cloud. For a small business, or the home its owner returns to, I would favour a modest local agent over an impressive online assistant that disappears with the connection. But I would trust it only after its limits had been set and tested.

I would start with the installer, not the software. Ask whether your solar system supplies power during a mains failure. Many grid-connected systems shut down for safety when the grid fails. Having panels does not automatically mean having backup power. Check whether you have a backup-capable inverter, a suitable battery and the wiring needed to supply your essential circuits. An agent cannot create a capability your electrical installation lacks.

Next, I would write down what must stay on, and in what order. Consider the freezer, router, card terminal, essential lights, alarm and any medical equipment. Your priorities will depend on your household and business; medical needs require an appropriate backup plan agreed with the relevant professional, not an AI judgement. This written order becomes the agent’s rules, including which loads it may disconnect first.

Keeping a card terminal powered does not guarantee that payments will work without its required connection. Likewise, keeping a router alive may preserve local Wi-Fi without restoring internet access. I would make those distinctions explicit. A useful blackout plan describes what remains possible, rather than treating a glowing power light as proof that everything works.

For the agent itself, I would choose a small computer supplied by the backup battery, running a local AI model. It would need local access to the battery information and approved device controls, without an internet login. The router and any equipment carrying those local commands would also need backup power. Otherwise, the supposedly independent system still has a missing link.

I would give that agent rules, not freedom. It could switch off approved loads in the agreed order and preserve a defined battery reserve. It must never operate breakers, interfere with safety devices or change the inverter’s protection settings. I would keep electrical protection outside its authority and insist on a straightforward human override. A model’s suggestion should never outrank a safety limit.

I would also require a readable log: what it switched off, when and why. That information should appear on a local screen or on phones connected to the local Wi-Fi, without needing an external service. In the dark, nobody should have to guess whether the freezer stopped because of a fault or an authorised decision.

Then I would rehearse with the installer, deliberately testing loss of mains through their safe procedure. Check the backup supply, local controls, display and override together. Between outages, I would let the same agent earn its keep by scheduling suitable devices when solar power is available and charging the battery at appropriate times, within the same limits. Daily usefulness is welcome; resilience is the requirement.

At DevMojos, we help businesses adopt AI agents safely. I believe the time to write your blackout rules is before the lights go out.