Yearly running cost Assuming 240,000 parts a year
By hand $4.8K / year 200 hours of work
Hosted service $7.7K+ / year
This workflow $379 / year Machine usage only; setup, hosting and review are extra. Images and camera Where will the part images come from?Inspect saved images, or connect an existing camera to an operator review queue.
Inspect saved images A trained model and a page to review your part images. $1,440 – 4,080 setup Connect my inspection camera Add image capture and a live operator review queue. $2,160 – 6,240 setup
What is included Adapt a vision model to your agreed visible defects. Return pass or hold recommendations with part IDs, images and model versions, and evaluate it on held-out production examples.
Start with 1 part family, 1 image per part and up to 5 known visible defect classes. Supply checked good and defective examples under your agreed camera view and lighting.
Camera hardware, lighting, fixtures, installation and labelling are separate. This workflow recommends pass or hold; it does not control machines.
Defects and false rejects How many defects must it catch?Set a defect target without rejecting too many good parts. Both numbers must pass together.
Balanced inspection Catch known defects without holding too many good parts. Defective parts held ≥95% · Good parts passed ≥98% Included Fewer missed defects More error analysis and tuning for difficult defect examples. Defective parts held ≥99% · Good parts passed ≥98% Setup +$432 – 1,440 My own requirements Set your own defect, false-reject or capture requirements. Quote separately
How inspection quality is measured Proposed targets, not tested results. Use parts from held-out production lots. Report every defect class separately; an overall score can hide rare defects.
Targets for your selected inspection What is checked Target Defective parts heldHeld defective parts divided by all labelled defective parts, reported for every agreed defect class. Defective pass decisions are misses. Pair with good-part yield so holding everything cannot pass acceptance. ≥95% Good parts passedPassed good parts divided by all labelled good parts. Unnecessary holds, abstentions and missing results count against yield. Report false holds alongside missed defects. ≥98% Unusable images heldHeld unusable inspections divided by all labelled unusable inspections, including blur, occlusion, wrong view and capture failures. Do not silently remove them from the test. ≥99% Traceable inspection resultsSubmitted parts with a valid decision and correct part, image and model IDs divided by all submitted parts. Missing results, duplicate decisions and stale image matches fail. ≥100%
Holding everything fails the good-part target. Missing, blurred or stale images cannot pass. Keep the existing inspection process during the pilot and let your quality owner approve operational use.
Speed How quickly should each part be checked?Choose the time to return a complete result, not just the model's processing time.
Within 5 seconds Included Within 2 seconds Setup +$216 – 720 Within 1 second Setup +$432 – 1,440
Timing and line speed At least 95% of test parts should meet the target at 1 concurrent inspection. Start at image submission, or at the camera trigger for camera scope. Include capture, transfer, queueing, checks, retries and result storage as applicable.
Mechanical handling and staff review are separate. Agree arrival rate and queue limits; a fast model alone does not prove the workflow can keep up with your line.
Running-cost target How much should each inspection cost?Set your machine-cost budget per part. This may change the model, hardware and scheduling.
Up to $0.01 Included Up to $0.005 Setup +$144 – 480 Up to $0.003 Setup +$288 – 960
Cost details Reference core machine cost: $0.002 – 0.004 per part at 20,000 parts a month. Includes allocated inference capacity, queue, logs and image evidence. Reinspection costs another run.
Hardware, training, labelling, operator review, agent calls and maintenance are separate. The budget is a target, not a measured bill. Local cost includes hardware use and electricity, not just tokens.
Hardware and environment Where should inspection run?Use your cloud or inspection computer. Keep the camera view, lighting and part positioning consistent.
Run it onMy cloud account My inspection computer or server
Hardware and image requirements Reference sizing: 4 vCPU, 8 GiB memory, images up to 4 megapixels. These are untested sizing assumptions; confirm the selected model and speed on actual equipment.
Agree the smallest visible defect, focus, lighting and acceptable variation. More pixels alone do not guarantee detection. Training compute and any GPU or camera purchase are separate.
Run the model in your cloud account. Camera scope uploads authorised images; include network delay in the test.
Use approved model weights inside your environment. Production images are not sent to an external model API. Agree setup downloads, licences and any agent-interface access.
Interface How should your team use it?Use an operator page or connect your agent. Camera capture is chosen above; no interface controls the line.
My existing AI agent Let your agent submit images and retrieve inspection records through MCP. Setup +$72 – 360 A quality review agent Use an agent to review results and explain recorded holds, without controlling the line. Setup +$144 – 480 An operator review page Inspect images, review holds and export evidence records. Included
Estimated setup price $1,440 – 4,080
This guide price is a starting point for your conversation with a provider. Agree the final scope and price directly with your provider.
Your requirements
Capture Inspect saved images
Inspection targets Balanced inspection
Speed target 5 seconds / part
Cost target $0.01 / part
Environment My cloud account
Interface An operator review page Hardware and training costs are separate. No machine actuation.
Prepare my brief → Generate a draft prompt to share with your connected RenX agent.
Enable JavaScript to configure a brief, or prepare your task in RenX .