Spare Parts by Photo: How an AI Agent Identifies Broken Components and Fills the Cart

A maintenance technician stands in front of a stopped production line built in 1998. The broken valve no longer has a readable number, and the machine’s nameplate is covered in oil. He takes a photo, sends it to the manufacturer via WhatsApp and waits. In the back office, someone searches through exploded views, calls back, asks for more photos. Getting to an order often takes half a day, sometimes more.

At Bosch’s Blaichach plant, according to a report in the German trade magazine Produktion, finding a spare part took 9.5 minutes on average, and over two hours in difficult cases spent in SAP, bills of materials and drawings. With an app that recognizes parts by photo, it dropped to under two minutes. That was in 2022, and the technology has advanced considerably since.

Today the entire chain can be automated: an AI agent receives the photo via WhatsApp or web chat, narrows the search using the nameplate and bill of materials, identifies the part, checks price and stock in the ERP, and places it directly in the customer’s cart. The customer just confirms. This guide shows how it works, what data you need, and where the limits are.

Why the spare parts search costs money

For many machine and equipment manufacturers, the spare parts business is the most profitable division. McKinsey puts the average EBIT margin in aftermarket services at around 25%, compared with around 10% for new equipment. According to Bain, spare parts account for more than half of industrial manufacturers’ service revenue.

At the same time, identifying the right part is often the bottleneck:

• Older equipment: Machines run for 30 to 40 years; nameplates get dirty, part numbers wear off.
• Similar variants: Seals, valves or terminals often differ only in details that are barely visible in a photo.
• Knowledge in a few heads: Often only a handful of experienced back-office staff can recognize a part by its shape.
• Return visits: Missing or wrong spare parts are considered the most common reason for a second technician visit in field service.

Customers also expect self-service. According to Gartner (March 2026), 67% of B2B buyers prefer purchasing without a sales rep. In the 2026 B2B Market Monitor by the German research institutes ECC Köln and IFH, 67% of executives see AI-powered product search as an important lever, and 65% would welcome AI agents that answer price inquiries on their own.

The chain at a glance: from photo to cart

There are solutions on the market for individual steps: visual search providers, messaging platforms, B2B shops with spare parts catalogs. The AI agent is the link that connects these steps into one continuous workflow.

How ready is your data?

Whether and how quickly such an agent can be implemented depends mainly on your data. The check shows your readiness score and the best place to start in two minutes.

Step 1: The intake channel

WhatsApp

For many service technicians and maintenance staff, WhatsApp is the most natural way to send a photo. According to Bitkom, 87% of Germans aged 16 and over use messengers such as WhatsApp privately. Business use comes with its own rules, though:

• Business Platform, not the app: An automated agent requires the WhatsApp Business Platform (Cloud API), usually through a provider. The regular WhatsApp Business app isn’t suitable for this, either technically or for data protection. With the Cloud API, Meta acts as a data processor, and an addendum on data transfers applies to EU customers.
• Meta’s AI rules: Since January 15, 2026, Meta has excluded providers of general-purpose AI assistants from the platform. Businesses using AI for their own customer service, such as spare parts inquiries, are not affected. Meta’s policy does, however, require a prompt, clear path to a human, for example a handoff to an agent, a phone number or a form.
• 24-hour window: Free-form replies are only allowed within 24 hours of the customer’s last message; after that, only approved templates.
• Costs: Since July 2025, Meta charges per message. According to Meta’s pricing documentation, service messages have been free only up to a monthly allowance since fall 2026. Check current terms before launch.
• Images: Photos arrive as JPEG, PNG or WebP up to 5 MB and can be retrieved via the API.

Web chat in the shop

A chat window in your own shop or customer portal is the more robust alternative: no platform risk, your own data retention rules, and the customer is usually already logged in. That means customer number, price list and installed machines are known immediately. Many companies offer both: WhatsApp for the technician on site, web chat for the buyer at their desk.

Step 2: Narrowing via nameplate and BOM

This is the most important and most often overlooked step. Instead of matching a photo against the entire catalog of tens of thousands of parts, the agent first asks: Which machine is the part in?

Here’s how it works:

1. The agent asks for a photo of the nameplate, if it doesn’t have one yet.
2. It reads the serial or machine number using text recognition. Modern multimodal AI models are very good at this, even with angled or partly dirty plates. QR or Data Matrix codes are even more reliable.
3. It looks up the machine in the ERP or in the list of machines installed at the customer (installed base).
4. It loads the BOM for exactly this machine, ideally with an exploded view.

Tens of thousands of candidates become a few dozen. That’s what makes the difference in accuracy, as the next section shows. Festo already offers a photo search in its spare parts catalog that reads the nameplate and Data Matrix code, and vendors such as nyris describe narrowing via the BOM as a core building block.

Step 3: Recognizing the part

Visual search with embeddings

The proven technique for part recognition is visual search: a model converts every image into a numeric vector (an “embedding”). Similar-looking parts sit close together in vector space. The customer’s photo is converted too and compared with the vectors of all catalog images.

How good is it? General-purpose image models are surprisingly weak on industrial parts. In a benchmark published in March 2026 by the vendor nyris, a freely available top model (Meta DINOv3) found the right part in first place for clips and connectors in only 26% of cases, versus 63% for nyris’s specialized model. According to nyris, overall accuracy rises to over 90% only with text recognition, filters and narrowing via the BOM. An independent research benchmark (ILIAS, CVPR 2025) reaches the same basic conclusion: recognizing one specific object among many similar ones is “far from solved” for general models.

That confirms why step 2 matters so much: the smaller the search space, the more reliable the match.

Multimodal AI models

Models such as GPT, Claude or Gemini can understand images and talk about them. For spare parts, they are both strong and weak:

Studies on fine-grained image recognition show that multimodal language models lag behind specialized models with very similar objects. The documentation from OpenAI and Anthropic explicitly lists small, rotated or blurry images as weak spots.

The right architecture combines both: multimodal AI leads the dialog, reads nameplates and describes the part. Specialized visual search finds the exact part number in the narrowed catalog. The agent orchestrates both.

Training with CAD data instead of thousands of photos

Traditionally, visual search needs many photos per part. According to field reports, depending on the method, that’s around 100 to 500 shots from different angles. Bosch even built a dedicated “digitization machine” for 360-degree images and trained around 20,000 of 30,000 parts.

The alternative: generate training images from CAD models. Photorealistic views are rendered from design data, optionally with simulated wear, dirt and poor lighting. Research from 2024 to 2026 shows that models trained mainly on such synthetic images, supplemented with a few real photos, can reliably recognize industrial parts. Vendors such as nyris and Synthavo offer this as a service.

For many machine builders, this is the decisive advantage: the CAD data for their own parts already exists. Where parts can’t be photographed, because they’re too large or already installed, it’s even the only way. At Bosch, around 10,000 of the 30,000 parts couldn’t be scanned.

Step 4: Follow-up, confirmation and handoff to humans

A good agent never orders on a hunch. This flow has proven itself:

• Confident match: The agent shows the identified part with catalog image, part number and description and asks: “Is it this part?”
• Several candidates: The agent shows the three to five most likely parts side by side and lets the customer choose, or asks a targeted question: “Does the valve have a threaded or a flanged connection?”, “Can you measure the outer diameter?”
• Uncertain: If no candidate exceeds a defined threshold, the agent hands over to the back office, with all photos, the identified machine and the suggestions. The employee doesn’t have to start from scratch.

An example of what the chat dialog could look like:

Customer: (photo of a broken valve) Need this one, line is down. Agent: Thanks. To find the right part: could you send a photo of the machine’s nameplate? Customer: (photo of nameplate) Agent: I found your machine: filling line series X, serial number 4711, built 2009. Two matching valves are used in this machine. (two images) Which one is it? Customer: The left one. Agent: Solenoid valve 24 V, part number 12-345. Your price: €186.40, in stock, ships today. I’ve added it to your cart. (link) Shall I also add the matching seal?

For safety-relevant parts, you should also have a qualified person check before the order is placed.

Step 5: ERP check

Before the part lands in the cart, the agent checks the ERP for:

• Customer-specific price based on price list, framework agreement or discount tiers
• Stock and delivery time, possibly across several warehouses
• Successor part if the identified part has been discontinued
• Accessories and wear kits that are typically replaced at the same time, such as seals

The agent only reads; it doesn’t change any ERP data. How to connect even older ERP systems without a modern API is covered in our guide to adding AI to your existing ERP with n8n.

Step 6: The cart

The last step turns a good answer into revenue: the part lands directly in the cart. How that works depends on the shop system:

Important: the agent fills the cart; the customer places the order. That keeps responsibility for the order with the customer, and wrong matches are caught before purchase. New standards for AI-agent purchasing, such as the Agentic Commerce Protocol from OpenAI and Stripe or Google’s Universal Commerce Protocol, are currently of secondary importance for B2B spare parts in the mid-market.

Vendor or build your own?

In practice, a combination is common: a specialized vendor provides image recognition, your own agent connects it with channel, ERP and shop. Vendors cite high accuracy figures, for example nyris 99.3% on a DMG MORI dataset or Synthavo over 98%. These are vendor claims; there are no independent comparison tests yet. So always test with your own parts and real customer photos.

What does it cost?

Costs are made up of several components:

• Image recognition: Specialized vendors rarely publish prices. Quanos, a provider of spare parts catalog software, quotes €0.36 per part per year for visual search.
• Training data: photographing existing parts or generating images from CAD data, one-time per part
• AI model for dialog and nameplate recognition: usage-based, usually a few cents per inquiry
• WhatsApp: per-message fees via a Business Platform provider
• Integration: connecting ERP and shop, building the agent, pilot phase
• Operations: adding new parts, analyzing mismatches, updates

On the other side are the savings: less back-office time per inquiry, fewer wrong orders and return visits, and more orders through the shop instead of by phone and email. At what inquiry volume an AI agent pays off in service in general is covered in our article At What Ticket Volume Does an AI Chatbot Pay Off?.

Legal: what to consider

• AI disclosure: Since August 2, 2026, AI systems that interact directly with people must be recognizable as such (Art. 50(1) EU AI Act). A short notice at the start of the chat is enough. More in our EU AI Act compliance checklist.
• Data protection: data processing agreements with WhatsApp or your messaging provider and every AI service, a privacy notice in the chat, metadata stripped from photos, deletion periods defined.
• Trade secrets: CAD data and BOMs are valuable. Clarify contractually that vendors may not use them for their own purposes or for training general models.
• Liability: For a mismatched part, contract law usually applies, i.e. wrong delivery or faulty advice. The new EU Product Liability Directive, to be transposed into national law by December 9, 2026, will also cover software and AI systems, but only for personal injury and property damage. Having the customer confirm before ordering and a qualified check for safety-relevant parts therefore also make legal sense.

Pilot in 30 to 60 days

1. Choose a series: a machine series with many spare parts inquiries and a good data foundation.
2. Prepare data: BOMs per serial number, part master data, CAD data or photos of the 500 to 1,000 most important parts.
3. Build a test set: 50 to 100 real photos from past customer inquiries with known correct part numbers.
4. Test recognition: measure first-place and top-three accuracy, with and without narrowing via the BOM.
5. Connect the chain: channel, agent, ERP lookup, cart, initially for a pilot group of customers only.
6. Measure: time to cart, rate of handoffs to the back office, mismatches, order value per inquiry.

Checklist: your path to a spare parts agent

Frequently asked questions

Conclusion: the agent connects what used to be separate

The building blocks for spare part recognition by photo have existed for years: visual search, text recognition, spare parts catalogs, B2B shops. What’s new is that an AI agent connects them into one continuous workflow, from the photo in the chat to a prefilled cart, with follow-up questions, ERP checks and handoff to humans when needed.

The key to success lies less in the AI model than in the data: if you know which BOM is in which machine, you narrow the search so much that recognition becomes reliable. If you have CAD data, you can train image recognition without elaborate photo shoots. And if you let customers confirm the order themselves, you gain speed without risking wrong orders.