Oct 07, 2026 / Avik Arefin

Works-Like vs Looks-Like Prototype: What Investors Need

Hardware accelerators and seed investors ask first for a rough works-like prototype that proves the core function, crowdfunding also needs photos of a real unit, and an AI feature counts as proven only when it runs on the target hardware.

Author

Avik Arefin

Published

October 07, 2026

Works-Like vs Looks-Like Prototype: What Investors Need

Hardware accelerators and seed investors who publish their criteria ask first for a works-like prototype: a rough unit that proves the product's core function. HAX, the hardware accelerator run by the venture firm SOSV, says it does not expect a prototype "to be overly polished." A looks-like model shows the intended form but does not work, so it answers a different question: whether customers want that shape. Crowdfunding asks for more: Kickstarter requires working prototypes and bans photorealistic renders, so campaign photos show a real unit. If your product's value is an AI feature, only a works-like prototype running the model on the target hardware shows that the feature exists.

Works-like, looks-like, and works-like-looks-like: the definitions

A works-like prototype does what the product will do but looks nothing like it. In a 2010 post, Pete Warden wrote that works-like prototypes "demonstrate the functionality that the designers are after." A looks-like is "an empty shell that doesn't work but reflects the intended form." A works-like-looks-like combines both in one unit.

Bolt, a seed fund for hardware startups, set out the same split in its 2015 product development guide. As of October 2026, Bolt's site says it no longer makes new investments under that name. Its guide defines a looks-like as "a model that represents the final product but doesn't function." In Bolt's guide, industrial designers do this work, making models "quickly and cheaply" from foam core, clay, or cardboard.

Bolt's works-like answers engineering questions about "core function, component selection, PCB, mechanics, feel, and assembly." A PCB (printed circuit board) is the board that carries the chips. Bolt calls it "common to build at least 3 or 4" full functional prototypes before moving on. It also runs engineering "simultaneously with the design phase," so different people build the two prototypes in parallel.

For a founder, that means two jobs to staff. An industrial designer builds the looks-like. Electronics, firmware, and software engineers build the works-like; firmware is the code that runs on the device's own chip.

EVT, DVT, and PVT: where the prototype stage ends

The prototype stage ends when the works-like and the looks-like merge into one unit. Bolt calls that unit the engineering prototype and expects "5 or fewer units (often 1 is enough)." It is built without tooling, the custom molds used for mass production. Its unit cost is "often >10x of the production version," and Bolt warns that the merge "often requires sacrifices on both sides."

Radio range shows why the merge is hard. Espressif, which makes the ESP32 family of Wi-Fi chips, tells designers to "consider the impact of the housing on the antenna." It recommends at least 15 mm of clearance around the antenna in all directions, then testing range in the final product. A works-like tested on an open bench and a looks-like drawn without that clearance can each pass alone and fail together.

After the merge, the product moves to a factory for three validation builds:

  • EVT (engineering validation test): the first build by the contract manufacturer, the factory that assembles the product. Bolt puts it at 20–50 units on first-shot tooling, with basic power, thermal, and EMI (electromagnetic interference) tests.
  • DVT (design validation test): Bolt puts it at 50–200 units, focused on production yield and environmental testing.
  • PVT (production validation test): Bolt puts it at 500 or more units, or 5–10% of the first production run.

Accelerators fund the step from prototype to factory. HAX says its program runs 6 to 18 months and aims to "reach Engineering Validation Testing (EVT)" and prepare for DVT. At HAX, a team applies with a prototype, and the investment carries it toward the factory builds. Our hardware timelines by stage show how long each step takes.

Investor and accelerator expectations at pre-seed and seed

Published criteria from accelerators and seed investors put proof of function ahead of polish. HAX lists "a working prototype" as one of three application criteria, next to a validated problem and a strong founding team. Among its reasons for rejection, it lists a company "still at concept stage" whose team "hasn't submitted a demo of a working prototype."

HAX asks for polish at the end of its program. By then, it expects founders to "finish an industrial version" of their hardware that accounts for manufacturability and human factors. So industrial design comes after the works-like has got the team in.

Y Combinator's hardware advice points the same way. In 2018, Pebble founder Eric Migicovsky, then a YC partner, told hardware founders to be "as hacky and prototype-y as possible." He wrote that as a seed investor, he looks "first to see if the startup has shipped product to a paid customer." A rough unit sold to a paying user meets that test, and a model that cannot run cannot be shipped.

EU funders score readiness on a formal scale. The European Commission defines nine Technology Readiness Levels (TRL), from TRL 1, "basic principles observed," to TRL 9, "actual system proven in operational environment." Every level describes the technology's maturity, and none describes appearance. The European Innovation Council's EIC Accelerator, which pairs grants with equity, targets TRL 6–8 as of October 2026. TRL 6 means "technology demonstrated in relevant environment," a level no looks-like model can reach.

If your next audience is an accelerator or a seed investor, put the first prototype budget into the works-like. Build the looks-like when you need customer reactions to the form, or before a demo day or a crowdfunding launch.

Crowdfunding prototype rules: Kickstarter and Indiegogo

Crowdfunding platforms ask for a working prototype and photos of a real unit. Kickstarter's hardware rules, updated August 6, 2026, say creators "are required to show working prototypes" and "We do not allow photorealistic renderings." Kickstarter also asks creators to explain how they will produce the product and whether they have made anything like it before.

Kickstarter's 2020 launch guide says a prototype "that shows off all your product's core features" is required by its rules. Indiegogo's hardware team gave similar advice in a 2020 webinar deck: launch with "a working prototype," an engaged audience, and manufacturing secured. The same deck asks for product shots of a "final industrial designed prototype."

The Zano drone shows what these rules guard against. Zano drew more than £2.3 million in pledges from about 12,000 Kickstarter backers in 2014. A report Kickstarter commissioned found "convincing evidence" that the pitch video was "misleading as to the existing capabilities and readiness level" of the drone. The project failed to deliver the autonomous flight it promised, and the few drones that shipped lacked key features.

For a crowdfunding launch, plan a works-like that shows every core feature on camera, inside a physical form you can photograph. In practice, that is a works-like-looks-like.

Works-like vs looks-like risk: what each prototype proves

Each prototype retires one kind of risk. Warden described technology risk as the risk "that your technology won't work." Market risk is the risk that customers do not want the product, and Warden noted that many recent startup techniques target it. A works-like retires technology risk, and a looks-like retires the part of market risk that concerns form.

Prototype Question it answers Risk it retires What it cannot show Who builds it
Works-like Does the core function work on the chosen parts? Technical feasibility: components, circuit board, firmware, radio, app Whether customers want the form Electronics, firmware, and software engineers
Looks-like Do customers want this size, shape, and interaction? Desirability of the form Function, battery life, radio range Industrial designers
Works-like-looks-like Do the function and the form fit in one unit? Integration: antenna clearance, fit, assembly Whether a factory can build it at volume Both teams together
EVT, DVT, PVT Can a factory build it reliably? Manufacturing: tooling, yield, power, thermal, EMI Customer demand Contract manufacturer

The costs differ. Bolt's looks-like models are made "quickly and cheaply" from craft materials. Its works-like takes at least three or four full builds, and the merged engineering prototype can cost more than ten times the production unit. For dollar figures, see how much a smart-device prototype costs in 2026.

Build first the prototype that retires the risk most likely to end the project. If customers already buy products that do the same job and your difference is the form, start with the looks-like. If your difference is a function nobody has shipped at your size, price, or battery budget, start with the works-like.

Unsure which risk is largest in your device? Send us the idea in three sentences through cortextech.dev, and we will reply with the part we would prove first.

AIoT prototypes: why an AI feature is a works-like risk

AIoT means AI running on an Internet of Things device, such as a camera that recognizes a person without sending video to a server. The AI feature is a function, so only a works-like prototype can prove it. A looks-like model can carry a lens and a status light, but no model runs inside it.

Four on-device limits decide whether the feature works, and a laptop demo hides all four.

Memory. The ESP32-S3, a Wi-Fi chip from Espressif with extra instructions for AI math, has 512 KB of on-chip SRAM, its fast working memory. Some versions add 2, 8, or 16 MB of PSRAM, an extra memory chip inside the package. A laptop has gigabytes, so a model that runs on a laptop says little about the chip.

Speed. Espressif lists its own pedestrian detector at 118.3 ms of model time per frame on the ESP32-S3, plus 11.6 ms of pre- and post-processing. That is about 7–8 frames per second. The model sees a 224 × 224-pixel copy of each frame, about 2.4% of the pixels in a 1080p (1920 × 1080) frame. A person far from the camera covers few of those pixels, so the mounting position changes what the model can detect. The same model takes 55.6 ms on Espressif's ESP32-P4, a chip without built-in Wi-Fi. Espressif says an ESP32-C or ESP32-S chip can serve as its "wireless companion chip," so the faster option adds a second chip to the board. Our ESP32-S3 vs Jetson Orin Nano vs Hailo-8L comparison covers the larger options.

Accuracy after conversion. Espressif's ESP-DL library runs quantized models, whose numbers are stored as 8-bit or 16-bit integers instead of 32-bit floating-point values. Google's LiteRT documentation says full-integer quantization makes a model 4x smaller and over 3x faster, with "little degradation in model accuracy." The same page tells developers "to check the accuracy of the quantized model." The accuracy measured on a laptop becomes the device's accuracy only after that check.

Real conditions. Google Health's model for diabetic eye disease reached over 90% accuracy, which its team called "human specialist level." In Thai clinics, nurses often took the photos in poor lighting, and the system rejected more than a fifth of them. The CHI 2020 study of that deployment found "tensions between the model's thresholds for data quality" and the data a real clinic produced. A camera at the install site meets lighting and angles that a training dataset may not contain.

For an AIoT pitch, run the model on the target chip, with the real camera, on data recorded where the device will be used. Report the time per frame and the error rates measured there, which no looks-like or laptop demo can produce.

CortexTech's method: testing an AI feature on field data

We have built a connected home-entry device with video calling, on-device AI, and a companion app for a non-technical founder. The method below is for a different product.

Consider a founder's ceiling sensor that counts people in a meeting room with on-device AI. With a device like this, we first test the model's accuracy in the real room. The founder cannot see this risk from the outside, because a model's published accuracy comes from its own test images. Espressif's page for its pedestrian detector lists timing but no accuracy figure.

In a 1–2 week prototype sprint, we mount a dev board and camera where the product will go, at the planned height and angle. We record frames through a working day, including a full room, an empty room. A person counts the people in a sample of those frames, and that count becomes the reference.

Then we feed the same frames to two versions: the original 32-bit model on a laptop and the quantized model on the device. The laptop result shows how the model handles the room. The gap between laptop and device shows what quantization cost, which is the check Google's LiteRT documentation asks for.

The obvious shortcut is to trust the model's score on a public dataset. That score comes from other cameras, angles, and lighting. The Google Health case shows that a lab score does not predict field results.

The founder then decides with measured miss and false-count rates. If the room causes the errors, the fix is more training frames from real rooms or a different mount position. If quantization causes them, the fix is a 16-bit model, which ESP-DL supports, or a larger chip. If no fix fits the schedule, version one can report whether the room is in use and add the count later.

The method applies to any connected device. Find the part most likely to end the product, prove it first in a works-like, and design the form around the measurements.

For the looks-like, we bring industrial designers into the project and give them the works-like's measurements to design around.

Does your device's pitch depend on an AI feature? Describe the feature and where the device will be used at cortextech.dev. We will reply with how we would test it in a 1–2 week sprint.

Works-like or looks-like first: the answer by audience

Your next audience What they ask for What to build first
Accelerator application (HAX) A working prototype, not "overly polished" Works-like
Seed investor (Migicovsky, YC, 2018) Product shipped to a paying customer Works-like units you can sell
End of an accelerator program (HAX) "An industrial version" of the hardware Works-like, then merge with a looks-like
Kickstarter or Indiegogo Working prototype with core features; real photos, no renders Works-like-looks-like
EU EIC Accelerator TRL 6–8 A works-like demonstrated in a relevant environment
Customer interviews about form Reactions to size, shape, and use Looks-like

Five of the six audiences above ask for a unit that works. For an AIoT device, the works-like is also the only prototype that can show the AI feature exists.

Sources

Want to read more?

More engineering deep dives from the Cortex R&D team.

All Articles