AI Product Innovation

I use the same sequence whether the brief is a hackathon prompt or an internal opportunity. The order matters: the insight comes before the concept, and the prototype comes before the pitch.

Device protection is the one product customers pay for every month and hope never to use. Carriers resell substantially the same insurer-administered product on substantially the same terms, so it competes on price. Competing on experience has barely been tried.

Where it breaks, from the member’s side:

• The deductible is a surprise. The most important number in the product is the one customers learn last, at the counter or midway through a claim.
• Loss claims run on trust alone. An administrator that doesn’t operate a network can’t check whether a phone really went dark. It believes the customer or asks for an affidavit.
• Coverage is bound to a handset. The policy names a device, and the backup belongs to whichever cloud account the customer set up months earlier.
• Nothing happens between claims. There is no reason to open the product in the months when nothing breaks.

AT&T runs the network. For every line it already knows when a device last reached a tower, whether it went dark abruptly or wound down, and whether the SIM has since appeared in a different handset. An administrator without a network has no equivalent and no path to one.

That is a structural advantage, not a head start, and it reframes the product: a claim can be corroborated by the network instead of taken on trust.

DeviceFlex turns protection into a membership that does something every month, and rebuilds the claim around signals only the operator holds. It is not a competing product: it layers onto the protection program AT&T already runs, inside the existing myAT&T experience, with no separate app and no new brand.

• A tiered membership (Basic, Plus, Family) offered at the add-ons step of a normal purchase, with deductibles visible before a plan is chosen.
• A household Protection Score that sets fraud-review sensitivity, inspection needs and whether a spare is pre-positioned.
• A Data Vault, so data continuity is guaranteed even when a spare handset isn’t.

The obvious design wires a vision model’s confidence straight into approve or deny. DeviceFlex does the opposite. Models are confined to perception: turning messy input, like a photo of a cracked screen or a disconnection pattern, into structured facts. Every consequential decision (approve, price, route, blocklist) is made by a pure, replayable function over those facts and the account state.

That split is the governance model. A model cannot approve, deny or price anything; it can only describe what it observed. The constraint is architectural rather than procedural, so it holds under pressure and can be demonstrated instead of attested.

The same insurer sits at the end of the path. DeviceFlex changes what reaches its intake, not who receives it.

The protectable core is the mechanism, not the membership. Pricing tiers are an economic practice; these are technical improvements to how the system works.

The levers, without internal figures:

• Attach. Answering the deductible question up front narrows the gap between online and in-store attach.
• Revenue per account. A tier ladder, with the Family tier growing revenue per household faster than per device.
• Claims cost. Network corroboration gates loss and theft fraud before a person reviews it.
• Replacement time. Pre-staging and line-anchored provisioning are designed for a roughly 15-minute in-store swap instead of a shipped replacement.
• Retention. A shared household pool and vault create switching costs.
• Cost to serve. Inference scales with claims, not traffic: dashboards, scores and eligibility checks are pure functions over state the system already holds.

A working, deployed prototype, built solo between 19 August and 5 September 2026: 17 routes covering the purchase flow, enrolment, claims, replacement, the vault, the household pool, and an impact screen where any decision can be replayed live. All five mechanisms run in it. The interface mirrors the existing myAT&T experience because the product is designed to live inside it.

• No acquisition problem. The audience is already inside myAT&T.
• An evidence-first rollout. Deductible transparency ships first as a standalone change. Network corroboration then runs in shadow mode, logged but not enforced, until its agreement with manual review is proven.
• A platform, not a feature. Corroboration and attestation are general capabilities that fraud, care and retail teams can all use.
• A portable build. The same build targets Vercel, Cloudflare Workers or Node, and the vision provider is three environment variables.

Honest risks. Rules on customer network data (CPNI) are the long pole, and shadow mode exists to build that case with evidence. Corroboration can be wrong, because a phone in a drawer looks like a lost phone, which is why the system flags rather than refuses. Pre-staging ties up inventory, so the threshold starts conservative.

• Protect the mechanism, not the business model. A tiered membership is an economic practice anyone can copy. Network corroboration, line-anchored restore and a replayable ledger are technical, and hard to replicate without a network.
• Cutting a feature can strengthen a product. Removing the loaner made the promise smaller, fully deliverable and more defensible.
• A verifier that always agrees verifies nothing. The demo includes a device whose telemetry contradicts its loss report, because the system has to be seen disagreeing.
• Show the loop. A feedback loop nobody can see isn’t a demo, so the score’s effect on fraud sensitivity is visible in the interface.