--- title: "Google Cloud Vision Integration" description: "Give AI agents production-grade image understanding through Google Cloud Vision. Agents run OCR on scans and multi-page PDFs, auto-tag images with labels, objects, and logos, moderate uploads with SafeSearch, and match photos against an indexed product catalog across 17 actions." url: https://flowrunner.ai/integrations/googlecloudvision date_modified: 2026-08-01T02:40:32-07:00 --- # Google Cloud Vision [AI](https://flowrunner.ai/integrations/category/ai-llms) Give AI agents production-grade image understanding through Google Cloud Vision. Agents run OCR on scans and multi-page PDFs, auto-tag images with labels, objects, and logos, moderate uploads with SafeSearch, and match photos against an indexed product catalog across 17 actions. [Verified](https://flowrunner.ai/integrations/verified "What does verified mean?") · 17 actions · API key · available [Google Cloud Vision website](https://cloud.google.com/) · [Platform Documentation](https://cloud.google.com/vision/docs) · Capability data verified 2026-07-27 1. A scanned document lands in the intake folder 2. Detect Document Text runs dense OCR with page, block, and word structure 3. Agent checks that required fields parsed with acceptable confidence 4. Clean documents post their extracted fields into the system of record 5. The queue owner gets a digest of what processed and what was flagged 6. A page whose OCR confidence falls short goes to a person to read before any record is created from it ## What This Integration Enables Google Cloud Vision gives FlowRunner agents eyes. They run OCR on photos, scans, and multi-page PDFs, returning not just text but its structure: pages, blocks, paragraphs, and per-word confidence. They tag images with labels and localized objects, spot brand logos and landmarks, extract dominant color palettes, suggest crop regions, reverse-search an image's presence on the web, moderate uploads with SafeSearch likelihood ratings, and match photos against an indexed product catalog. For volume, the async batch operations OCR whole document sets offline and write results to Cloud Storage. Every one of these operations is analysis; none of them changes anything. That is precisely why the human-in-the-loop question lives downstream. Vision returns likelihoods and confidence scores, not verdicts, and the costly act is what a flow does with them: rejecting a customer's upload, creating a financial record from a misread scan, publishing an image that should have been held. FlowRunner's job is to act on the confident cases and route the ambiguous ones to a person, with the model's evidence attached. ### Without FlowRunner **Data entry from scans**: Someone reads each page and types what it says into the system **All-or-nothing moderation**: Every uploaded image gets eyeballed, or none of them do **Unsearchable image libraries**: Thousands of photos with filenames as their only metadata ### With FlowRunner **Documents parsed with structure**: Pages, blocks, and words extracted with confidence scores and layout **Only the ambiguous reach the queue**: Clear images pass, obvious violations are held, and humans see just the middle **Images tagged automatically**: Labels, objects, logos, and dominant colors attached the moment a file arrives ## Use Case Scenarios ### Scan-to-record document intake A new file lands in [Google Drive](https://flowrunner.ai/integrations/google-drive-api). The agent runs Detect Document Text, walks the structured fullTextAnnotation to find the fields the document type requires, and posts clean extractions straight into the downstream system. Pages with low-confidence blocks are routed to the intake queue with the original image and the parsed attempt side by side, so the person resolves in seconds what the model could not resolve at all. ### Upload moderation that respects the gray zone A user image arrives from [Dropbox](https://flowrunner.ai/integrations/dropbox-service). Safe Search Detection returns likelihood ratings across adult, violence, racy, spoof, and medical categories. Images rated VERY\_UNLIKELY across the board publish without ceremony. Images rated VERY\_LIKELY on adult or violence are held and logged automatically. Everything in between, the POSSIBLE and LIKELY middle, goes to a moderator with the ratings attached, because likelihood is exactly the kind of judgment call platforms should not delegate to a threshold. ### Batch OCR at archive scale A digitization project stages thousands of scanned PDFs in [Google Cloud Storage](https://flowrunner.ai/integrations/google-cloud-storage). The agent kicks off Async Batch Annotate Files with a gs:// destination, polls Get Operation until the job completes, and the OCR results land as JSON back in the bucket, ready for indexing. The agent tracks progress from the operation metadata and reports completion with counts, turning a months-of-typing archive problem into a supervised pipeline. ## Human-in-Loop Highlight The gate on this page is not a Vision operation, because Vision only reads. The gate is the takedown that follows it. Rejecting a paying customer's upload on a wrong guess costs trust; publishing an image that should have been held costs more. So when Safe Search Detection returns a middle rating, the FlowRunner agent quarantines the image rather than deleting it and posts to the moderation channel: "Upload from account 8841 rated racy: LIKELY, adult: POSSIBLE, violence: VERY\_UNLIKELY. Publish, reject, or escalate to policy review?" The moderator sees the image and the ratings together and answers in one tap. The agent executes the decision and logs the rating, the decision, and the decider next to the file. The model provided likelihoods; a person owned the verdict. That is [human-in-the-loop](https://flowrunner.ai/concepts/human-in-the-loop) for read-only AI: gate the consequence, not the analysis. Agent processes routinely Detects exception requiring judgment Clear match Continues automatically Ambiguous Routes to human via preferred channel Human decides Agent resumes with decision ## Agent Capabilities 17 actions ### Text and Documents 3 - **Detect Text (OCR)** Extract text from images with per-word bounding boxes and locale. Best for sparse text: signs, labels, product packaging. - **Detect Document Text** Dense OCR optimized for scanned pages and screenshots, returning a structured annotation of pages, blocks, paragraphs, words, and symbols with confidence and layout. The workhorse of document intake. - **Annotate File** Run Vision features against a PDF, TIFF, or GIF file per page or frame, up to 5 pages per synchronous request. The right tool for a short multi-page scan. ### Classification and Objects 4 - **Detect Labels** Tag an image with broad categories and concepts, each with a confidence score, for general classification and search. - **Detect Objects** Localize multiple objects with names, scores, and bounding polygons, including several instances of the same object type, when the question is where things are, not just what. - **Detect Logos** Recognize product and company brand logos with bounding boxes, for brand monitoring and ad verification. - **Detect Landmarks** Identify well-known landmarks with geographic coordinates, for geo-tagging and location recognition. ### Moderation and Faces 2 - **Safe Search Detection** Rate an image across adult, spoof, medical, violence, and racy categories on a five-step likelihood scale. The evidence a moderation decision runs on. - **Detect Faces** Detect faces with landmarks, head pose, and emotion likelihoods. Does not perform recognition or identification, a boundary worth knowing before designing a flow around it. ### Image Analysis 3 - **Image Properties** Extract dominant colors with pixel fractions and scores, for palette extraction and color-based matching. - **Crop Hints** Suggest crop regions that keep the important content in frame, optionally tailored to target aspect ratios for thumbnails and banners. - **Web Detection** Find visually similar images, pages that include the image, and best-guess labels from web context. Reverse image search as an agent capability. ### Product Search 1 - **Product Search** Match an image against a pre-indexed Vision Product Set, returning scored product matches with grouped bounding-box results. Requires a product set built and indexed in advance. ### Batch and Advanced 4 - **Annotate Image** Run multiple detection features against one image in a single request with full control over the feature list. - **Async Batch Annotate Images** Start offline batch annotation of image sets in Cloud Storage, with JSON results written to your own bucket. - **Async Batch Annotate Files** Start offline annotation of multi-page PDF, TIFF, and GIF files in Cloud Storage, the volume path for document OCR. - **Get Operation** Poll a long-running batch operation's status, progress, and outcome until it completes. The heartbeat of every large OCR job. ## Frequently Asked Questions ### What can FlowRunner do with Google Cloud Vision? FlowRunner agents can run Detect Labels, Detect Text (OCR), and Detect Document Text in Google Cloud Vision, plus 14 more actions. ### Does connecting Google Cloud Vision to FlowRunner require OAuth? No. Google Cloud Vision connects to FlowRunner with an API key, no OAuth flow required. ### Can Google Cloud Vision trigger a FlowRunner workflow automatically? Google Cloud Vision doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger. **Work at Google Cloud Vision?** This integration exposes Google Cloud Vision to AI agents on every FlowRunner plan, including through MCP, at no cost to you. [See what FlowRunner offers integration partners](https://flowrunner.ai/integrations/partners), including how to keep this page current. --- Markdown version of https://flowrunner.ai/integrations/googlecloudvision. Site index: https://flowrunner.ai/llms.txt