> ## Documentation Index > Fetch the complete documentation index at: https://support.fieldko.com/llms.txt > Use this file to discover all available pages before exploring further. # AI and Photo Quality in FieldKo > What the artificial intelligence (AI) photo check in FieldKo does and doesn't do, and how to validate any image recognition tool before you rely on it. This page is for sales leaders, admins, and managers deciding how artificial intelligence (AI) fits into photo evidence. The outcome is a clear picture of what FieldKo's AI-Guided Capture does today, and a way to test any image recognition tool before decisions depend on it. > [!NOTE] > > **What FieldKo does today:** AI-Guided Capture checks a survey photo's **quality** on the device. It doesn't recognize products, read prices, count facings, or check planograms, and its result isn't saved. FieldKo Atlas makes no automatic image-compliance decisions: a person or your own process decides whether a photo shows compliance. ## What Good Looks Like * Photos are sharp, well lit, and free of glare, so a reviewer can judge them. * Everyone knows that a photo-quality check isn't a compliance check. * Any image recognition tool starts with a narrow task that can be checked against human answers. * A person stays accountable for decisions the tool informs. * Tool accuracy is measured by category and revalidated after packaging or model changes. ## When to Use It and Common Failure Patterns Use this page when you plan photo questions, or when someone proposes using AI to judge shelf or display compliance. Common failure patterns: * **Treating "Great Shot" as "compliant".** A sharp photo can still show an empty shelf. * **Automating a weak standard.** AI can't fix an unclear scoring rule or an out-of-date product list. * **One photo for everything.** Expecting one image to prove stock, price, placement, and campaign conditions at once. * **Trusting the vendor's accuracy figure.** A tool that works on one category or store format can fail on another, or when a promotional pack replaces the usual one. * **Hidden correction work.** Reviewers fix so many results that the time saved disappears. ## Operating Procedure ### Use AI-Guided Capture for Photo Quality 1. Ask for photos in survey questions, where AI-Guided Capture is available. 2. Coach reps to use **Smart Capture (Guided)** for evidence photos, and to retake when told. 3. Review a sample of photos each week for what they show. See [Capture Photo Evidence That Holds Up](https://support.fieldko.com/retail-execution/photo-evidence.md). ### Validate Any Image Recognition Tool Before You Rely on It If you're considering a third-party image recognition tool: 1. **Pick a narrow task** that a person can check, such as "Is the promotional sign visible?" or "Is this known product on the shelf?". 2. **Build a validation set.** Pick a representative sample of photos across store formats, lighting, devices, shelf density, packaging, and regions. Have a person record the correct answer for each. 3. **Measure the tool against it.** Track precision (how often a positive result is right), recall (how many true positives it finds), and the cost of each kind of error. 4. **Match the decision to the confidence.** Let the tool surface low-risk suggestions, send uncertain results to a reviewer, and require stronger proof before costly or customer-facing actions. 5. **Pilot against the human process.** Compare time saved, issues correctly found, overrides, rework, and whether more issues get fixed. Include setup and supervision costs. 6. **Keep a trail.** Keep the original photo, the tool's version, and any correction. 7. **Revalidate** after packaging, category, or model changes. ## How to Implement It in FieldKo ### AI-Guided Capture **Availability:** Built-in, FieldKo Atlas app version 1.0.70. * On a survey question that asks for a photo, the photo box says **AI-Guided Capture**. Reps choose **Smart Capture (Guided)**, **Quick Photo**, or **Choose from Library**. See [Take Photos with Smart Camera](https://support.fieldko.com/field-work/photos.md). * Smart Capture gives live prompts while the rep frames the shot, such as holding steady and leveling the phone. * After the photo is taken, the app checks it on the device for blur, brightness, and glare. The check doesn't need a connection and sends nothing to a server. * The rep sees **Great Shot**, **Acceptable**, or **Try Again**, with tips to improve. What it doesn't do: * It doesn't identify products, brands, prices, facings, signage, or planogram positions. * It doesn't check that the whole shelf or display is in the frame. * It doesn't save its quality result to Salesforce. Only the photo is attached. * It isn't used for visit, account, or pack-out photos, and there's no setting to turn it on for them. ### Third-Party Image Recognition **Availability:** External. FieldKo doesn't include image recognition. Survey photos are stored in Salesforce, so a third-party tool would read them from there and write any results back through your own integration. See [Export FieldKo Data and Photos](https://support.fieldko.com/admin/export-data.md#where-photos-are-stored) and [Integrate FieldKo with Other Systems](https://support.fieldko.com/platform/integrations/integrate.md). Validate the tool as described above before reports or decisions use its results. ## Fictional Example Fictional example: a made-up cereal brand, Maple Ridge Oats, tests a third-party tool on one task: "Is the in-store promotional sign visible?" Reviewers label 200 photos from 4 store formats. The sign is visible in 100 of them. | Result | Count | | - | - | | Tool says visible, and it is | 81 | | Tool says visible, but it isn't | 9 | | Tool says not visible, but it is | 19 | | Tool says not visible, and it isn't | 91 | Precision is 81 ÷ (81 + 9) = 90%. Recall is 81 ÷ 100 = 81%. Most misses come from one store format with strong glare. The team decides the tool can flag likely missing signs for review, but a person checks every "not visible" result before any store is marked noncompliant, and reps in that format are coached to use Smart Capture to reduce glare. ## What to Measure | Metric | Definition | Population | Denominator | Exclusions | Owner | | - | - | - | - | - | - | | Evidence acceptance | Sampled photos a reviewer accepts as usable evidence ÷ photos sampled | A weekly random sample of survey photos | Photos sampled | Photos for optional questions marked not required | Field manager | | Tool precision | Correct positive results ÷ all positive results | The validation set, by category | Positive results | Photos the reviewer can't judge | Retail execution lead | | Tool recall | Correct positive results ÷ true positives | The validation set, by category | True positives | Photos the reviewer can't judge | Retail execution lead | | Override rate | Tool results a reviewer changes ÷ tool results reviewed | Live results in the pilot | Results reviewed | None | Retail execution lead | Evidence acceptance needs a manual review process, because the quality result isn't saved. Tool metrics need your own validation records. ## What to Do When It Goes Wrong | Symptom | Likely Cause | What to Do | | - | - | - | | Reps get **Try Again** often | Glare from shelf lighting or the flash, or low light | Tilt the phone, turn off the flash, and coach on framing | | The app shows "Quality analysis unavailable" | The check couldn't run | Ask the rep to look at the photo before using it | | Sharp photos still can't be judged | The photo doesn't show the part of the shelf the question is about | Rewrite the question to say exactly what to photograph | | A tool's accuracy drops | New packaging, a new category, or a model change | Revalidate on a fresh sample before trusting results | | Someone reports AI-checked compliance | A quality result was mistaken for a compliance result | Correct the report: FieldKo makes no automatic compliance decision | ## Checklist * [ ] Ask for evidence photos in survey questions so AI-Guided Capture applies. * [ ] Tell managers that photo quality isn't compliance. * [ ] Sample photos each week and record acceptance. * [ ] For any third-party tool, define a narrow task and a validation set. * [ ] Measure precision, recall, and overrides by category. * [ ] Keep a person accountable for decisions the tool informs. * [ ] Revalidate after packaging or model changes. ## Industry Best-Practice Context FieldKo's guide [Using AI and Image Recognition Responsibly in Retail Execution](https://fieldko.com/best-practices/ai-image-recognition-retail-execution/) covers when shelf AI can help and how to validate it with a real field pilot. ## See Also * [Take Photos with Smart Camera](https://support.fieldko.com/field-work/photos.md) * [Capture Photo Evidence That Holds Up](https://support.fieldko.com/retail-execution/photo-evidence.md) * [Design Surveys and Tasks for Reporting](https://support.fieldko.com/platform/customize/design-surveys-for-reporting.md) * [Export FieldKo Data and Photos](https://support.fieldko.com/admin/export-data.md) This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.