A while ago, we sat in a meeting with the rights holder of a well-known Japanese IP. They were interested in an AI version of their character, and they asked one question that stopped the room:
“If a user asks it about a sensitive chapter of history between our two countries, what should it say?”
It's a fair question. A Japanese user and a Chinese user would expect different answers, and once a character can talk, an answer given in one market doesn't stay there. A silent character can be sold worldwide. A talking one can close a market.
Nobody had planned for it. We see the same pattern with many IP owners.
The demo works. The character greets you, remembers your name and tells a decent story. Everyone in the room is excited. Then the project moves toward production, and problems that were invisible in the demo surface one by one. Most have little to do with how smart the model is.
IP-based AI toy projects tend to stall in five places. None of them can be fixed at the last minute, and all of them are far cheaper to decide at the start. Below is each one, and what we do to help IP owners avoid it.
Choose one scenario and one audience before any prototype work; “a talking version of our character” is not a product brief. Many projects begin with exactly that sentence. Ask what the character should say, in which situation, and why a user would talk to it instead of watching the cartoon. The answers are usually missing.
The same goes for interaction beyond speech: touch, vibration, a camera, NFC. These are what make a character feel alive. A shell that only talks tends to lose its novelty within days.
How Joinet approaches this
We start by narrowing the project to one core scenario and one audience, then define the interaction around it. This happens before any module is selected.
A persona that passes a demo will not necessarily hold across thousands of units; it has to be built as an asset you can test. The common shortcut is a general-purpose model with one paragraph describing the character. In a demo, a few scripted questions get convincing answers. In the field, the toy forgets what a child told it yesterday, breaks character the moment a question falls outside the script, and sounds playful in one reply and cold in the next.
A standard solution adds a second problem. The shell is yours, but the dialogue templates, knowledge base and character settings are the same ones every other client received, and you cannot change them.
A character's identity is its fixed story, its way of speaking, its emotional limits and the topics it must never touch. A single prompt cannot hold all of that steady.
How Joinet approaches this
We don't sell standard solutions. Each IP gets its own character corpus, persona guardrails, memory management and emotional voice, packaged so they can be regression-tested and version-locked. The target is the same behavior on every unit, in the same situation.
Write down what your character may and may not say before the first prototype; boundaries are requirements, not a compliance check at the end. This is the pitfall we see least discussed, and it comes in three forms.
Your character is a ten-year-old from an invented world. A user asks it about trigonometry or the Big Bang. If it answers, it steps outside the world you built. If it refuses, it looks unintelligent. Both are failures, and users will test exactly this.
Suppose a user asks for a story in which your character fights a monster alongside another company's famous hero, one you hold no license for. Should the toy generate it? If it does, the story also has to respect each character's standing: who wins, who gets captured, whose role carries more weight. Without that, the output can contradict your own IP.
The opening question belongs here. Any answer touching politics or history can be acceptable in one country and damaging in another, and a talking character carries that answer everywhere it is sold.
How Joinet approaches this
During requirements, we work with you to produce a short boundary brief. It covers what the character knows, what it declines, how it declines while staying in character, how it handles other IPs, and where answers must differ by market. It becomes the reference for every later step.
Plush structure, acoustics, certification and connectivity are production constraints; if you design only for the prototype, they come back as rework.
Electronics and plush are made by different industries. Once the main board goes in, the body can bulge and distort the character's silhouette. Stuffing can muffle the microphone, so the toy misses its wake word in a noisy room, and the fabric can dull the speaker. A battery compartment that was never planned leaves assembly workers with no room to work. Each pattern change restarts sampling.
An AI plush toy is both a children's toy and a connected electronic product, so the requirements stack. A change to firmware or the module can trigger retesting. Some module certificates cover only sample units and cannot be used for mass shipment or relabeling, which teams often learn only when the first large order is already accepted.
A plush toy has no screen. If every conversation depends on the cloud, a weak home Wi-Fi signal or a child who can't complete network setup leaves the toy silent, and silent toys get returned.
How Joinet approaches this
We bring module, firmware and acoustic design together with the plush pattern work early, so the structure is settled before tooling. We map the certificate chain before you commit to a module, including whether it supports relabeling. We use a local-plus-cloud setup so basic conversation still works offline and network setup stays simple.
An AI toy is a service you keep running; plan for content, updates and cloud costs from the first day. A project that budgets only for development and manufacturing treats launch day as the finish line. For an AI toy it is where decline begins.
Users exhaust the existing conversations and need new stories, seasonal content and updated knowledge. Mistakes in persona or answers have to be fixed in devices already in homes. Cloud and moderation costs grow with every unit sold. There have been overseas cases where server costs spiraled, services shut down and devices already in homes stopped working.
How Joinet approaches this
The solution ships with an OTA update channel and a content management platform, so you can update devices in batches, refresh the IP knowledge base in one place and keep long-term operation as a standard process rather than a rescue task.
If you can't answer all five, your project is likely to stall at one of the places above.
The cheapest time to fix any of these is before the first prototype.
Considering an AI version of your IP?
Bring your answers, or your gaps, and we'll go through them with you before you commit to a prototype.
Talk to our team