Ori
The assistant that works with no signal.
Every other support assistant in this category has the same design flaw. It needs the internet to explain to you why you have no internet.
The moment this is built for
You have landed. The plane door opened twelve minutes ago. Your phone says No Service, or it says the carrier name but nothing loads. There is a queue at passport control, no wifi you trust, and a person waiting for you at an address you cannot look up.
This is the exact moment that every support chat in this industry stops working. The widget needs to load. The bot needs to reach a model. The help centre is a website. All of them are on the other side of the connection you do not have, so the advice arrives after the problem has solved itself or after you have paid a carrier $12 a day out of desperation.
Ori was built backwards from that moment. The first question we asked was not what the model should say. It was what has to be true for anything to be said at all with the radio off.
Two layers, and only one of them needs a network
Layer one is retrieval, and it does the work. Every article in the help centre, every device specific install path, every setting name in every supported operating system version, is compiled into a compact search index of about 4 MB that ships inside the app and updates when you happen to be online. Ask Ori a question with the phone in airplane mode and it searches that index on the device, matches your handset and operating system version, and shows you the article with the exact taps, in order.
Layer two is the model, and it is the polish. When the network is reachable, the model takes the article the index found, plus your account state and your device diagnostics, and writes an answer fitted to your situation rather than making you read a general article and work out which half applies to you. It also handles the questions the index cannot phrase match, the ones where you describe a symptom rather than name a feature.
The order matters and it is the opposite of how these things are usually built. Retrieval first, model second. Most assistants put a model in front and treat retrieval as a source it may consult, which means no model, no answer. Here the answer exists before the model does. The model improves the wording. It is not load bearing.
4 MB
The help centre, compiled onto your phone
0
Bytes of network needed to get an answer
Under 60 s
Target to a human reply, 24 hours a day
100 MB
Free rescue data a month, if you need to get back online
What works offline, and what does not
| No connection at all | Connected | |
|---|---|---|
| Install steps for your handset | Yes | Yes |
| The data roaming toggle, with the exact path | Yes | Yes |
| APN settings for the network you are on | Yes | Yes |
| Reset or reinstall a profile | Yes | Yes |
| Your balance and plan at last sync | Yes, with the timestamp | Yes, live |
| Emergency and embassy numbers | Yes | Yes |
| Buy or top up data | Queued | Yes |
| Live coverage and current speeds | No | Yes |
| Refunds and account changes | No | Handed to a human |
| An answer written for your exact situation | Article, not written prose | Yes |
Queued means Ori takes the instruction offline and runs it the moment any connection appears, including the free 100 MB a month of Lifeline data every account carries in 145 countries.
The guardrail, written down
There are three categories of fact Ori is not permitted to generate, because being wrong about them costs you money or strands you.
- Prices. Every figure Ori shows is read from the live catalogue, the same source as the pricing page. It cannot estimate, round or infer a price, and it will not tell you what something probably costs.
- Coverage and speed. Numbers come from the measured dataset with the sample count attached, never from a guess about a country. If we have fewer than 25 samples for a place, Ori says the sample is too small rather than inventing a median.
- Refund outcomes. Ori can tell you the refund policy and it can open the case. It cannot tell you that you will be refunded. A person decides that, and a bot promising a refund it cannot authorise is the fastest way to turn one annoyed customer into two.
When it does not have the fact, it says so in one sentence and offers the handover. It does not produce a confident paragraph shaped like an answer, which is the failure mode people have learned to expect and the reason most of them skip the bot entirely.
The handover to a person
Target is under 60 seconds to a human reply, 24 hours a day, in every timezone, because being stuck happens at 4am somewhere. What makes that fast is not staffing alone. It is that the person opening your message already has the diagnostics attached: the handset, the operating system version, the profile state, the last network attached, the last three error codes and your balance.
That removes the eight message opening exchange where somebody asks you what phone you have and whether you have tried turning it off and on again, while you are standing in an arrivals hall with 4 percent battery. The first reply you get is about your actual problem.