# AI Parts Finder — what it costs, and what it can honestly do **Audience:** Anyone evaluating the AI Parts Finder concept — per-search economics and vision-model accuracy for photo + description parts search. **Companion to:** [Parts Finder prototype (legacy)](../legacy/homepage-parts-finder.html), [search-experience-concepts.md](search-experience-concepts.md), [model-landing-concepts.md](model-landing-concepts.md), [decisions.md](decisions.md) **Date:** August 2026 **Styled version:** [parts-finder-feasibility.html](parts-finder-feasibility.html) > **Cost is pennies per search and is not the constraint. What a photo can > honestly tell you is.** The prototype promises "find any part from a photo or > description." This brief prices that promise at current model rates and grades > it against what vision models reliably do: **identifying the machine and the > part is real today; judging remaining wear from a photo is not.** The design > should lean on the first and never promise the second. --- ## 1. The short version | At a glance | | | --- | --- | | **~2¢** | per photo search on the mid-tier model, including AI analysis and candidate matching | | **~$200/mo** | at 10,000 searches per month — less than the card-processing fee on the orders it produces | | **+⅓¢** | is all the photo itself adds over a typed search — nameplate, worn part, or whole machine | Three findings drive everything below: - **Per-search cost is noise** against parts-order value, because the 140,000-SKU catalog never passes through the model (section 2). - **Accuracy splits into three honest tiers** — identification is shippable, damage-and-diagnosis works with the right framing, and quantitative wear judgment should stay out of the promise (section 4). - **The right product framing is the parts-counter expert, not the inspector** — recommend and say how to verify; never claim to measure remaining life (section 5). --- ## 2. Why it is cheap The economics work because of one architectural rule: **the catalog never passes through the AI.** The model does the small job only a model can do — look at the photo or description and emit brand, model, serial range, part type, and a confidence score as structured data. That result drives an ordinary catalog search, which is free. Optionally, a second small call ranks the top candidates the search returns. So each search is one or two small API calls — roughly 5,000 input tokens including the photo and about 500 out — not a big-context call over catalog data. The photo itself is nearly free: the API downscales every image to about 1.2 megapixels, so even a full phone photo bills around 1,600 tokens, about a third of a cent. A serial plate costs the same as a walk-around shot. ### Calibration — UC Assist UC Assist, already live on this demo, runs about **18¢ per warm question** because it deliberately reads its entire ~295K-token corpus every time (the right call for its job). The Parts Finder is roughly **10× cheaper per interaction** precisely because the catalog stays behind search instead of inside the model's context. --- ## 3. What a search costs Per-search estimates at current Claude API list pricing (August 2026). "Full flow" is the identification call plus a second candidate-matching call. | Model tier | Typed search | Photo search | Full flow | | --- | ---: | ---: | ---: | | **Economy** — Claude Haiku 4.5 | ~0.5¢ | ~0.8¢ | ~1.3¢ | | **Mid** — Claude Sonnet 5 **(recommended)** | ~1¢ | ~1.5¢ | ~2.5¢ | | **Top** — Claude Opus 5 | ~3¢ | ~4¢ | ~6.5¢ | A back-and-forth session with clarifying questions runs roughly 3–8¢ end to end on the mid tier; prompt caching keeps follow-up turns from re-billing the setup. **At volume:** 10,000 searches a month on the mid tier is roughly $150–250; 100,000 is $1,500–2,500. Even top-tier quality at 10,000 searches is about $500–650 a month. Against a parts-order average order value, cost per search is smaller than the payment-processing fee on the resulting order. --- ## 4. What the AI can and cannot see Three tiers of honesty, from shippable to off-limits. ### Identification — ship with confidence **Data plates and labels** are the strongest leg: brand, model, serial, voltage, and part numbers read reliably, with a graceful failure of "can't read it — retake the photo." **Machine category** is near-certain from any photo and **brand** is strong when trade dress is visible. **Part type** ("that's a rear squeegee blade") is solid. One honest limit inside the strong tier: the exact model from a walk-around photo (T500 vs. T500e) is weaker, because the differences often are not visible. Design it as a funnel — show the top two or three candidates to confirm, or route the customer to the data plate. ### Damage and diagnosis — assist, don't decide **Obvious damage** — a torn blade, shredded bristles, a cracked hose — is spotted reliably. **Symptom triage** is the quiet win: "streaking on the floor" maps to a ranked list of likely causes (worn rear blade, debris under it, blade angle, vacuum issue) because those chains are textbook service-manual content. What the photo cannot do is confirm *which* cause — so the product presents ranked likelihoods with a check step, never a single verdict. ### Wear measurement — keep out of the promise "Worn enough to replace" is not a photo judgment. Squeegee wear is assessed by edge rounding you feel with a finger; brush wear is measured against a built-in indicator. A phone photo at arbitrary angle and lighting does not capture either — and the model will still produce a confident-sounding answer, which is worse than declining. The same goes for naming an exact SKU from an unlabeled part's shape: photos carry no scale, and dimensions are usually what decide the SKU. --- ## 5. The design consequence: build the parts counter, not the inspector The right mental model is the expert at the parts counter. You show him a photo and he says: "That's the rear blade for a T500 — looks chewed up. Have you flipped it through its other edges? Here's the kit." He recommends and tells you how to verify; he never claims to measure remaining life. Three design moves deliver that: - **Guided capture.** Ask for the data plate, the blade edge-on, or the brush's wear indicator. Photographing the indicator turns wear judgment — the weak tier — into label reading, the strong one. - **Ranked causes with a confirm step.** Every diagnosis ships as "most likely, plus how to check," not a verdict with a percentage. - **Bias toward the cheap consumable.** Blades and brushes are scheduled replacements, so an early "replace it" costs the customer little. The one answer the system should never volunteer is "it's fine." ### Prototype copy note The current prototype copy says the AI "detects what needs replacing" with confidence scores — that is the wear-measurement tier writing a check the model can't cash. Soften it to "identifies your machine and the likely part." The honest claim is still a strong one. --- ## 6. What settles it The accuracy tiers above are calibrated from known vision-model behavior, not yet from USA Clean's machines. A **half-day evaluation** converts them to measured numbers: 30–50 real photos from the service team (data plates, worn parts, floor symptoms), run through the actual prompt, scored by someone who knows the equipment. The same test settles whether the economy model holds up — which is the difference between the cost table's rows. Two dependencies to plan around: - **Compatibility tagging.** The catalog side of the lookup — "here are the parts that fit your T500e" — rides on tagging parts with `compatible_models[]`, the same data effort already scoped for [model landing pages](model-landing-concepts.md) Phase 2 (PD-008). The AI finder makes that tagging more valuable, not more optional. - **Scope of this brief.** This is a feasibility read on cost and accuracy, not a full feature brief. If the client green-lights exploration, the next artifact is a `parts-finder-concepts` brief on the house template — Stencil path, app-vs-custom shortlist, phased recommendation with triggers — and the decision of record gets the next `PD-NNN`. --- ## 7. Assumptions & sources **Cost assumptions.** Roughly 5,000 input tokens per identification call (prompt, taxonomy hints, and one photo) and 500 output; the two-call flow about doubles it. Anthropic list pricing as of August 28, 2026: Claude Haiku 4.5 at $1 in / $5 out, Claude Sonnet 5 at $2 / $10, Claude Opus 5 at $5 / $25 per million tokens. Images bill as input tokens at roughly (width × height) / 750 after downscaling. Estimates are conservative; even off by 2× they do not change the conclusion. **Sources:** - [Anthropic API pricing](https://docs.claude.com/en/docs/about-claude/pricing) - [Anthropic vision documentation — image token math](https://docs.claude.com/en/docs/build-with-claude/vision) - [UC Assist worker README — the 18¢/question calibration point](https://github.com/saturdaysort/usaclean-research/blob/main/tools/ucassist-worker/README.md) **In this repo:** [Parts Finder prototype (legacy)](../legacy/homepage-parts-finder.html) · [search experience brief](search-experience-concepts.md) · [model landing brief](model-landing-concepts.md) · [decision log](decisions.md).