--- title: "Why Computer Vision Projects Fail | Fortune AI" description: "Computer vision models improved, but most pilots still fail. Andrew Chen explains why deployment, retraining, integration, and economics now decide success" canonical: https://www.fortuneai.app/blog/the-models-caught-up-the-moat-moved/ language: en --- [Blog](https://www.fortuneai.app/blog/) AI landscape AI landscape September 13, 2026 7 min # The Models Caught Up. The Moat Moved. [By Andrew Chen](https://www.linkedin.com/in/andrewchennn/) Founder & CEO, Fortune AI Why computer vision projects still fail after the models got better ![AI video analytics deployment across a 3PL distribution center loading dock](https://www.fortuneai.app/assets/blog/models-caught-up-distribution-center-v2.webp) Two years ago, a useful camera measurement took a computer vision team, thousands of labeled frames, and months of tuning. Today, a frontier vision model can return a usable detection from one frame and one sentence in seconds The obvious question ## If the models caught up, why do most camera AI projects still fail? The old workflow worked on the footage the model had seen. Then the sun moved, a forklift parked in the wrong place, or a new rack blocked half the frame, and the count drifted Detection is now cheap. That should have changed the success rate for warehouses, 3PLs, distribution centers, and plants. Instead, most projects still struggle to get from a controlled pilot to daily operations The numbers have not moved ## Better models did not close the production gap Parsec found that only 10% of manufacturers use AI at scale. Lenovo's 2025 CIO Playbook shows 23 AI proofs of concept for every three production launches. MIT NANDA reported that 95% of enterprise generative AI pilots produced no measurable return on the P&L Read the figures together and something stands out. None of them is a model-accuracy problem. The models improved dramatically between 2023 and 2026, while the pilot-to-production gap remained Sources: [Parsec 2026 State of Manufacturing](https://www.parsec-corp.com/news-and-events/parsec-survey-72-of-manufacturers-have-adopted-ai-but-only-10-have-done-so-at-scale) · [IDC CIO Playbook 2025, commissioned by Lenovo](https://pages.lenovo.com/rs/183-WCT-620/images/CIO%20Playbook%202025%20-%20Its%20Time%20for%20AI-nomics_February%202025_AP242508IB.pdf?version=0) · [MIT reference to The GenAI Divide](https://mitsloan.mit.edu/shared/ods/documents?PublicationDocumentID=10819) ![Bar chart comparing AI adoption at scale, pilots that never reach production, skills gaps, and pilots with no profit and loss return](https://www.fortuneai.app/assets/blog/ai-pilot-deployment-gap.webp) Production gap The deployment problem persisted while model capability improved Where projects actually die ## Four problems appear after the demo works We have deployed vision AI in more than 50 paying sites across five countries, starting in a place that punishes mistakes harder than a warehouse: swimming pools, where a missed detection can have serious consequences. Here is what kills a camera project, in the order we see it - 01 The site changes Glare, a bumped camera, a new rack, and shifting traffic patterns turn a clean lab result into unreliable operating data. A one-time model build has no answer for a site that keeps changing - 02 Nobody owns retraining The integrator delivers and leaves. When false alerts start, the operations manager stops trusting the number, and the pilot dies quietly - 03 The output does not land anywhere A detection is not a result. A result is a number in the report the GM already reads, or an alert sent to the person who can act on it now - 04 The economics never close In our experience, a custom computer vision project can cost $50,000 to $150,000, before the annual cost of the engineer needed to maintain it. That often exceeds the value of the operating problem being solved Meanwhile, the floor got more expensive ## Warehousing productivity fell while unit labor cost rose Between 2019 and 2024, US warehousing output per hour fell 32% while unit labor cost rose 91%. Labor is now one of the largest cost lines for many 3PLs and distribution centers, yet many still cannot answer simple questions about station staffing or units per labor hour The cameras that could help answer those questions are already on the ceiling. They are recording, but nobody is turning the footage into an operating number Source: [US Bureau of Labor Statistics, Productivity and Costs by Industry, NAICS 493](https://www.bls.gov/news.release/archives/prin2_06262025.pdf). The chart compounds the reported annual changes from 2019 through 2024 ![US warehousing chart showing output per hour down 32 percent and unit labor cost up 91 percent from 2019 to 2024](https://www.fortuneai.app/assets/blog/us-warehousing-productivity-2019-2024.webp) US warehousing Output per hour fell as unit labor cost increased from 2019 to 2024 The moat moved ## The value moved from the model to the deployment If detection is commoditized, the value sits in everything around it: getting a measurement live on the site's existing cameras in days, keeping it accurate as the site changes, and turning the output into a number an operator will use That is why we built Fortune AI as a deployment engine rather than a custom model shop. You describe the operational event in plain language. Fortune AI builds the analysis, runs it on the cameras you already have, and delivers the number every shift Our deployment is roughly 10 times faster than a custom computer vision project, with about 90% fewer false alerts than the systems we replace, because the system retrains on the site's own footage instead of freezing on day one Case study · Bay Area ### Labor utilization, measured in two weeks A reusable-container operator could see pieces of the line, but could not quantify how paid time was used across stations ![Illustrated reusable-container processing floor with workers active, walking, and waiting across three stations](https://www.fortuneai.app/assets/blog/labor-utilization-case-study.webp) - Before The in-house line counter ran at roughly 70% accuracy , needed one person to maintain it, and still could not quantify how labor time was used - Deployment Fortune AI used three existing cameras to measure work, waiting, walking, and cleaning by station - Outcome Two weeks later , the team had a live labor-utilization metric to compare stations and review staffing against the actual workflow What this means if you run a facility ## Stop evaluating vision AI on lab accuracy alone Ask how many days it takes to get from footage to a live number on your existing cameras. Ask who retrains the analysis when the site changes, and what that costs. Then ask where the output lands: a dashboard nobody opens, or the report and workflow your team already uses If a vendor cannot answer all three in a sentence each, you may be looking at another pilot that will not go live The models did their part. The rest of the industry has to catch up on the part that was always harder: deployment Bring us the stuck project ## Show us one camera question Tell us what has to be measured and we will show you what the first week looks like [Request a demo](https://www.fortuneai.app/book/?source=blog-models-caught-up)[Explore warehouse AI](https://www.fortuneai.app/industries/warehouse-operations/)