Tony Fadell: Why AI Gadgets Flopped, and Where the Money Is

By Lazar Milicevic · Published October 7, 2026 · 9 min read

You're tempted to bolt AI onto your offer, but you're not sure whether it will earn trust or just borrow hype. The people who built the first wave of AI gadgets made the same bet, and most of them lost. Here's what Tony Fadell said about why, and the four-question check I run before building any automation for a client.

What Fadell actually said

Fadell led the original iPod and helped create the iPhone and the Nest thermostat. In comments published by TechCrunch on October 7, 2026, he described the first generation of AI gadgets as "Gen 1" devices that met no real need and were interesting technology for geeks. The next wave, he argues, has to earn consumers' trust.

The examples came from a talk at MIT Future Fest. According to a recap, he showed a slide of three discontinued devices: the Rabbit R1, the Humane Ai Pin and the Limitless pendant.

One argument I found useful on why assistants are a hard sell: he said fewer than 0.01% of the world's population has ever had a human assistant, and that most consumers don't know what an assistant is (source). If the buyer has never used the thing you're imitating, you can't sell them on the concept. You have to sell them on a result.

That's the claim. I'm not going to put words in his mouth beyond it, and I haven't found a full transcript of the talk, so I'm sticking to what TechCrunch and the reporting around it support.

The failure was visible because you could hold it

Most commentary on this story is about hardware. That's the wrong lens. The failure mode is identical in software, and gadgets just made it physical.

The numbers on the hardware side are blunt:

• Humane: the $699 AI Pin launched in April 2024 and was poorly received by reviewers. On February 18, 2025, Humane announced it was shutting down and selling assets to HP Inc. for $116 million, per Axios, well below the billion-dollar valuation it had reportedly been aiming for (How-To Geek).
• The shutdown detail that matters: existing Pins kept working only until 3 PM ET on February 28, 2025. After that, calling, messaging, AI queries and cloud access stopped. Customers who paid for a device found out it was a thin client for a server someone else controlled.
• Rabbit R1: $199 at launch with no subscription. Marques Brownlee called it "barely reviewable". In September 2024, The Verge reported, citing Rabbit CEO Jesse Lyu, that only about 5,000 of roughly 100,000 purchasers were using the device on any given day (summary).
• Limitless: Meta announced its acquisition on December 5, 2025 and stopped selling the Pendant to new customers that day, promising support for existing users for at least another year.

I'm not going to pick a single sales figure or price for these products, because the reported numbers conflict. The pattern is consistent enough without them: a shiny demo, a buyer who has to change behavior, a miss that happens confidently in front of the customer, and trust gone.

Rabbit's story also shows the afterlife. Its OS3 platform, an agentic OS that works without the R1, became generally available on September 22, 2026. It has no monthly fee, uses a bring-your-own-key model with keys from providers such as OpenAI or Anthropic, and one account can link up to five computers. Rabbit has said it won't make an R2. Whether the R1 itself counts as discontinued is reported inconsistently, so I'll leave that alone.

The same failure in a small business

I see this often with small teams. The owner says, "We tried ChatGPT, but it didn't actually integrate with anything." That's the gadget failure in software form. The AI was a toy sitting next to the work, not inside the work. Nobody trusted it because it couldn't act, and nobody could verify what it did.

Here's the arithmetic that makes this concrete. Many founders lose a significant chunk of every day to inbox triage, invoicing and lead follow-up. A chatbot that gives advice about those tasks doesn't remove a single minute. An automation that does the task and shows its work removes hours.

That's the gap. Advice is a feature. Completed, verifiable work is the product.

The four-question check I run before building anything

This is my own framework, not Fadell's. I built it because it forces the idea to solve a real problem and builds the trust mechanism in from day one, which is exactly what the first wave skipped.

1. What is the one specific job, and how many minutes does it cost you today? Write the number down. If you can't, you don't have a problem yet, you have a vibe.

2. What triggers it, and what is the finished artifact? Not "helps with email." I mean: an email arrives, and a draft reply plus a tagged lead record is created in your CRM. A trigger and a deliverable.

3. Where does a human approve? This is the trust question Fadell is pointing at. Separate what the system suggests from what it completes. Let it act on the boring, reversible stuff and route exceptions to you.

4. How will you know it's working? Log every run. Count errors, count minutes saved, review it weekly. Trust is earned with receipts, not promises.

I write the answers into a spec file before I touch any code. For an invoicing job, it looks like this:

job: "Turn supplier PDFs into draft invoices"
baseline_minutes_per_week: 600        # measured, not guessed
trigger: "PDF lands in billing@ inbox"
artifact: "Draft invoice in accounting tool + row in review queue"
auto_actions:                          # reversible, low-risk
  - extract_fields
  - create_draft
needs_human:                           # exceptions routed to a person
  - amount_over_threshold
  - unknown_vendor
  - field_confidence_below: 0.9
metrics:
  - errors_per_week
  - minutes_saved_per_week
  - approvals_vs_edits
review_cadence: weekly
kill_rule: "If minutes_saved is flat after 2 weeks, stop"

If any field in that file is blank, I don't build yet.

Approval gates: where trust gets engineered

The third question deserves its own section, because it's the one most automation projects skip and the one the gadget flop makes obvious.

A small invoicing business hand-typing 200 invoices a week doesn't need an AI that guesses. It needs one that reads the source, fills the fields, flags the five weird ones, and lets a person click approve. The system does the 195 boring ones. The human spends their attention on five.

In code, the gate is unglamorous:

def route(invoice, confidence, rules):
    """Auto-complete only the reversible, high-confidence cases."""
    if invoice.vendor not in rules.known_vendors:
        return "human_review"
    if invoice.total > rules.auto_approve_limit:
        return "human_review"
    if confidence < rules.min_confidence:
        return "human_review"
    return "create_draft"  # a draft, not a sent invoice

Notice the auto path creates a draft, not a sent invoice. Drafts are reversible. Sent invoices aren't. Design every automation so the default action is the one you can undo.

The security community says the same thing about agents that act on a user's behalf: start with least privilege, separate read from write and spending authority, approve consequential actions, review logs, and keep a fast way to revoke access. That's the same four questions from the other direction.

Why this matters more as agents get access

The trust question gets sharper as agents get deeper access to your accounts. Meta's Muse agent launched in the US on September 8, 2026 with deep access to users' apps and accounts. 404 Media reported on October 5, 2026 that Meta engineers found several security vulnerabilities in Muse in the weeks before launch, at least one of which could have let an ordinary user reach sensitive internal Meta databases. Meta rushed to fix them.

I want to be careful here. Meta says Muse is safe, and it hasn't said whether any flaw was exploited. One anonymous source's belief that a breach is inevitable is opinion, not fact, so I'm not claiming one happened.

The part that is documented: Meta's bug bounty for a Muse VM escape pays up to $300,000, and Meta says that because a Muse agent holds a user's most sensitive data and can act for them, a compromise is a first-class security risk. When the vendor itself says that, a small business should take the hint. Grant the agent the narrowest access that gets the job done.

Rabbit's own terms, as reported, call OS3 a public technical preview or beta and say it isn't intended for production, enterprise, regulated, safety-critical or unattended use. I respect that honesty, and it's a useful signal: a vendor is telling you exactly what not to trust it with. Read those terms before you wire anything into your billing.

My take: the integration layer is the moat

The gadget flop was never about the hardware being bad. It was a category error: companies treated AI as the product when the product is the outcome.

On-device AI may matter here. Fadell predicts future agents will have to run on-device rather than depend on data centers, and he sees Apple as best placed on trust, hardware and privacy, even though its new Siri runs on a custom Gemini-based model (source). I'd treat that as his prediction, not settled fact.

My bet for a small business is more mundane. Models get commoditized every few months. The winners won't have the best model. They'll wire a model into the systems people already use, Gmail, billing, the CRM, chat, so work disappears without anyone learning a new interface. The boring integration layer is the moat, and the demo is just marketing.

A simple test follows from that. If your AI idea needs the customer to behave differently, you're building the next flop. If it quietly removes a task they hate, you're building something people will pay for and defend.

What "working" looks like

Run the check, build the smallest version, log everything. Give it two weeks. If the numbers don't show real hours reclaimed, kill it. The aim is to see the savings in your logs within the first few weeks, not to end up with a drawer full of abandoned subscriptions.

For the sake of a worked example (round numbers, illustrative only): if your baseline is 10 hours a week on invoice entry and the automation handles most of it while a human approves the exceptions, you should be able to see the drop in your logs within the first two weeks. If you can't see it, the automation isn't doing what you think.

Where this fits in my work

I build practical AI automations for solopreneurs and small teams, and the four questions above are the check I run before building anything. The principle carries over to any tool you evaluate or build: keep suggesting and completing separate, make the default action reversible, and log what happened so you can see what it saved. You can find more of my work at bizflowai.io.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.