AI that Actually worked
Proof of concept is not a use case. The Cases of PrezenceAI cover real implementations in production — with verifiable results, declared costs, documented challenges, and lessons that only appear when the system needs to function for 100,000 users at 2 a.m.
The Difference Between Demo and Production
The gap between an AI demonstration and a production implementation is immense — and systematically underestimated. Demos are built to work. Production systems need to function when data is messy, when users ask unexpected questions, when the model hallucinates, when the API fails, when costs scale 50x beyond projections. Each of these scenarios requires specific engineering that does not appear in a demo.
The PrezenceAI Cases start from verifiable evidence. Each case study includes: primary source (company report, research paper, public technical interview), concrete metrics (not "improved significantly", but "34% reduction in processing time"), declared technologies e challenges faced. Company cases that do not declare methodology are treated as marketing, not as evidence.
The most recurrent result pattern in verified Cases: the pilot phase exceeds expectations; the scaling phase reveals quality, cost, and maintenance issues that the pilot phase did not capture. This is not failure — it is engineering. But organizations that do not plan for the scaling phase from the beginning pay double to reach it.
Total Cost of Ownership — hardware + licenses + people + energy + maintenance over time
Project and AI Engineering Terminology for Pointy Cases. Used to describe implementations with technical and organizational precision.
| Term | Editorial Definition | Level |
|---|---|---|
| POC | Proof of Concept — prototype to validate technical feasibility; does not represent production performance | Gold |
| MVP | Minimum Viable Product — first functional version in production; success criterion defined in advance | Gold |
| ROI | Return on Investment — financial gain divided by cost; calculation frequently omits maintenance costs | Diamond |
| TCO | Total Cost of Ownership — hardware + licenses + people + energy + maintenance over time | Diamond |
| Drift | Gradual degradation of model performance when production data diverges from training data | Diamond |
| Human-in-the-Loop | Human supervision of AI decisions — mandatory in high-risk domains, impacts throughput | Gold |
| A/B Testing | Controlled comparison between version with AI vs. without AI — gold standard for measuring real impact | Diamond |
| P99 Latency | 99th percentile latency — measures the typical worst-case scenario; more relevant than average for production UX | Gold |
| Feedback Loop | User feedback capture mechanism to improve model — essential for systems that evolve | Gold |
| Shadow Mode | Run AI in parallel with current system without showing output to user — validates before replacement | Silver |
| Graceful Degradation | System behavior when AI fails — fallback to rules or human without interrupting operation | Gold |
| Observability | Ability to monitor AI system behavior in production — logging, tracing, alerts | Gold |
| Cold Start | Additional latency on the first call — relevant for serverless and models that need to load | Silver |
| Chunking Strategy | Decision on how to divide documents for RAG — direct impact on retrieval quality | Gold |
| Evaluation Framework | Set of metrics and criteria to evaluate output quality — necessary before, not after deployment | Diamond |
AI in Retail: How Computer Vision Reduced Losses in G21 Retailers
The application of computer vision to prevent losses in retail is one of the most mature and measurable AI use cases in operation. Unlike LLM applications, where outputs are difficult to measure, loss reduction is a number in accounting—verifiable, auditable, and directly comparable to the implementation cost.
The Case of the Supermarket Chain
One of Brazil's largest supermarket chains implemented computer vision in 47 pilot stores in 2-2024. The system uses shelf cameras to detect: items placed in bags without passing through the checkout, inconsistent self-checkout behaviors, and discrepancies between scanned and packaged items. Result after 12 months: 31% reduction in losses for monitored products, equivalent to R$2.3 million per store per year. The system costs approximately R$180,000 per store annually for hardware and software.
The Challenges Nobody Talks About
False positive rate: the initial system generated 2-3 false alerts per hour per camera—unmanageable volume that caused fatigue among security operators. Adjusting the threshold and retraining on local data reduced it to 0.2 alerts per hour. Privacy and LGPD: implementation required legal approval, mandatory signage in stores, and anonymization of customer data for training. Lighting variation: cameras that worked during business hours failed at night—adjusting the model for low-light conditions added three additional months.
"The system worked perfectly in the pilot of 3 stores. When we scaled to 47, we discovered that each store had different lighting, a different layout, and different customer patterns. Scale is always where reality appears." — Technology Director of the network, technical interview NRF 2024
Vector Balance: Efficiency × Privacy
Loss reduction is real and verifiable. The ethical question also: Continuous customer monitoring systems normalize surveillance in public spaces. The technological response — anonymization, local processing without storing faces, focus on behavior rather than identity — mitigates but does not eliminate the tension. Legitimate AI use cases often carry implications that warrant analysis beyond operational efficiency.