AI on your phone: what a 14MB agent changes for your business

Your field team loses signal in the middle of a visit and the "smart" app turns into a brick. It happens on construction sites, on farms, in warehouses, in building basements. And every time it happens, someone writes it down on paper to type in later. This week a piece of news came out that touches this problem: a 14MB AI agent that runs right on the phone, with no server and no internet. It's worth understanding what AI on your phone actually changes for a business like yours.

What was announced, in plain English

A company called Cactus Compute released Needle2. It's a language model (the same family of technology as ChatGPT) that weighs 14MB. For comparison: a good phone photo is about 4MB. The WhatsApp app is over 100MB.

The promise is that it runs on phones, watches, smart home devices and robots. And the word "agentic" in the title means it doesn't just answer questions. It can carry out tasks: call a function, fill out a form, trigger a command.

I want to be honest with you: a model this size can't write a contract or analyze a balance sheet. It doesn't compete with GPT or Claude on knowledge. It's small on purpose. What it does well is understand a simple instruction and turn it into action.

Think of it as an intern who knows nothing about the world, but perfectly understands "book John's visit for Tuesday at 2pm".

Why size matters more than it seems

When you use AI today, this is almost always what happens: your text leaves the device, travels to a server in another country, gets processed there and the answer comes back. That has three practical consequences:

• You pay per use. Every question costs a few cents. At scale, it becomes a bill.
• You depend on the internet. No signal, no AI.
• Your data leaves your hands. In some industries that's a legal problem, not just a trust issue.

A 14MB model running on the device itself eliminates all three at once. Per-use cost: zero. Internet: not needed. Data: never leaves the phone.

I wrote here on the blog about AI agents running locally and the conclusion back then was "worth it, but it needs a good machine". What changed now is the scale. We're no longer talking about a server at your company. We're talking about any 200-dollar Android phone.

Cheap AI isn't the one with the lowest price per token. It's the one with no tokens at all.

Where this solves a real problem

Here are examples of things clients have asked me for that today are expensive or fragile because they depend on the cloud.

Maintenance technician in the field. He finishes the job and dictates: "replaced the compressor, customer asked for a cleaning quote, back on the 15th". The agent on the phone turns that into a filled-out work order, creates the quote task and schedules the return visit. No signal. When the phone gets a network, it syncs.

Store with a scale or kiosk. A clerk says "two kilos of picanha and half a kilo of sausage" and the system builds the order. No API key needed, no two-second latency, no stopping when the store's internet goes down (and it does).

Clinic or doctor's office. Reception takes a voice message and the agent classifies it: it's a reschedule, it's an insurance question, it's an emergency. None of that goes to a third-party server. For health data, that makes life a lot simpler with privacy law.

Equipment on the production line. A sensor detects abnormal vibration and the agent decides whether to alert the operator, log an incident or stop the machine. Decision in milliseconds, on the equipment itself.

Notice the pattern. In none of these cases does the AI need to "know a lot". It needs to understand a sentence, pick one action from a few options and execute. That's exactly what a small model is for.

What this model doesn't do (and where they fool you)

Here's the strong opinion: most AI vendors will push big models for everything, because that's where their margin is. And a good part of business problems don't need that.

But the opposite is a trap too. A 14MB model:

• Doesn't write a good email or report.
• Doesn't answer open questions about your market.
• Makes more mistakes when the instruction is ambiguous.
• Needs to be properly configured with the actions it can take. Without that, it's a parrot.

The architecture that works in practice is hybrid. The small model sits on the device and handles what's fast, repetitive and sensitive. When something comes up that it can't handle, it hands off to a big model in the cloud. Eighty percent of interactions solved for free and on-site. Twenty percent with the heavy intelligence. The math works and the system doesn't freeze when the Wi-Fi drops.

How to test this without spending a fortune

You don't need to buy a robot or replace your fleet of phones. A path I use with clients:

1. Pick a boring, repetitive task that today depends on someone typing. Filling out work orders is the champion.
2. List the 5 to 10 possible actions. "Create work order", "schedule return", "flag pending part". If the list goes past 20, the task is too big to start with.
3. Run a pilot on 2 or 3 devices for two weeks. Measure how much time it saved and how many times the agent got it wrong.
4. Only then decide whether to expand, switch models or give up. Giving up cheap is a win, not a failure.

The cost of a pilot like this is days of work, not months. And the result is concrete: either it saved team hours or it didn't.

What I think is coming

A 14MB model today is Hacker News material. Two years from now it will ship from the factory in every phone, watch and home appliance. Whoever already has their processes mapped into "actions an AI can execute" will plug this in over an afternoon. Whoever still has everything on paper and WhatsApp will start from zero.

It's not urgent. But it's the kind of thing that gets cheaper for those who prepare early.

If you have a process like that in mind, the one someone on the team complains about every week, tell me. I like to look at the problem first and only then talk about technology. See how I work.

Your "smart" app turns into a brick every time the field team loses signal. On a construction site, a farm, a warehouse. And someone writes it down on paper to type in later.

This week a 14MB AI agent came out that runs right on the phone. No server, no internet, no per-use cost. Smaller than WhatsApp, fits on any 200-dollar Android.

It doesn't write contracts or analyze balance sheets. But it understands "replaced the compressor, back on the 15th" and turns that into a filled-out work order. Offline.

Cheap AI isn't the one with the lowest price per token. It's the one with no tokens at all.

Two years from now this will ship from the factory in every device. Whoever already has their processes mapped into simple actions plugs it in over an afternoon. Whoever is still on paper and WhatsApp starts from zero.

Got a process like that on your team, the one someone complains about every week? Tell me in the comments.

#ArtificialIntelligence #OnDeviceAI #Automation #SmallBusiness #Technology