AI that survives the parking lot: deploying computer vision in the real world

We all love a great AI demo. But the places where AI earns its keep are not stages or slides—they are parking lots, hospital garages, and city streets. Out there, a model that is 99% accurate can still fail the moment it matters. If your occupancy count drifts by 10 spaces in a day, the digital sign misleads drivers, the owner stops trusting your system, and a data protection officer may call before lunch to ask how you store license plates. That is the part of AI that actually changes the built environment—and it is much less glamorous than a keynote.

Wemolo, a Munich-based startup, has spent the past six years putting computer vision to work in parking: scanning plates for ticketless entry, enforcing maximum stays, counting capacity, enabling bookings, and stitching private lots into city-scale supply. They now run thousands of cameras across six countries, process tens of thousands of park events per day, and answer several inquiries from data protection authorities each week. Their story is not about a flashy model. It is about building an operating system for AI in the wild—where accuracy compounds, privacy is productized, and ROI is negotiated contract by contract.

This article unpacks what it takes to deploy computer vision in public spaces. We will start with the insight that models are only a small part of the job. Then we will look at concrete examples from parking, explore the regulatory and commercial implications, and end with a pragmatic takeaway: if you want AI that works outside a demo, obsess over accuracy, privacy, and operations.

Start with the problem before you touch code. The team behind the parking platform did something unfashionable in a hype cycle: they walked from parking lot to parking lot before they wrote software. They pitched a camera-based solution as if it existed and watched who leaned forward. One store manager near Munich’s Olympiapark said he spent two hours a day policing Oldtimer meetups that filled his spaces and drove customers away. That was a clear signal—pain, urgency, a willing first customer—and only then did they build a prototype. This is not romance; it is risk reduction. Talk to customers first and you will ship fewer features that nobody needs, a discipline the founders sharpened in a US accelerator that literally forbade writing code in the opening weeks.

Accuracy compounds in the real world. Consider a busy facility with 10,000 park events per day and 500 spaces. If your counting pipeline is wrong by just 1%, you are off by roughly 100 events daily. That drift can translate to an occupancy display that is 10 bays wrong by the evening. After a week, you are not wrong—you are irrelevant. This is why accuracy is not a vanity metric in physical operations; it is a trust boundary. Owners will revert to manual checks, and drivers will stop believing your guidance. The startup wrestled with this early. Legacy solutions used inductive loops in the ground—expensive to install (on the order of tens of thousands of euros) and still not precise enough, often requiring daily manual resets. The team instead pushed their cameras to do double duty: automatic number plate recognition (ANPR) for entry/exit and computer vision for occupancy, wrapped with error modeling and self-correction so the count does not drift. The benefit is not just lower capex. It is a system that does not need humans to rebalance it every morning.

Make hardware decisions that align with your measurement problem. A single camera that can see a wide area is cheaper and easier to maintain than instrumenting every bay—but you inherit complexity: night lighting, weather, occlusions, glare, and the messy choreography of cars hunting for spaces. The trick is to see your model’s “error budget” as a product decision. Where do you place cameras to minimize compounding error? What confidence thresholds do you expose to the UI? When do you trigger a fallback (for example, temporarily showing ranges rather than exact counts)? This is product management for AI systems, not just modeling. It must be designed up front, then reinforced by operational playbooks for calibration, health monitoring, and interventions after storms or power work.

Build privacy into the product, not just the paperwork. Hundreds of cameras across multiple countries and thousands of daily users invite scrutiny—and rightly so. The company expects two to three inquiries per week from data protection authorities, most coordinated through Bavaria’s regional office. Their compliance stance is part of the product: prominent on-site signage, clear purposes, stringent retention policies, and a hard line against constructing individual movement profiles. They do not say “the Düsseldorf plate was at these seven locations”; they say “10% of yesterday’s visitors were out-of-town,” because aggregation is permissible and tracking individuals is not. This design choice does more than pass audits. It simplifies architecture, clarifies internal norms, and reduces the blast radius of a breach. In regulated environments, the path to scale runs through predictable, documented behavior that regulators can come to know and trust.

Solve one job well, then layer value. The initial use case was simple: scan plates on entry, enforce a maximum stay so customers can park, and remove the need for paper tickets or disc clocks. From there, they added booking for longer stays, flexible tariffs (for nights and weekends when lots are underused), and digital payment at kiosks or on the phone. On paid facilities—think hospitals or urban garages—ticketless entry eliminated queues and broken barriers, letting the computer vision system become the backbone for pricing and flow. The same cameras then powered live counts of free spaces and, in time, the ability to guide drivers directly. Each layer used the same operational footprint to unlock incremental ROI, which matters because hardware-heavy deployments only pay back over years. Contracts reflect that reality: five to ten years is typical, and churn is below one percent. In other words, when you make AI part of core operations, the stickiness comes from the utility, not the algorithm.

At city scale, AI enables new spatial choices. Many European cities are reclaiming street space for biking, trees, and terraces. The immediate obstacle is not will—it is storage. Where do the cars go? One answer is to unlock private capacity: supermarkets at night, office blocks on weekends, underused corners of multi-story garages. You cannot orchestrate that without reliable, priceable visibility into free spaces. This is where computer vision becomes an urban instrument: count accurately, surface availability, then redirect demand. It is not hard to imagine a near future where this same infrastructure supports dynamic exposure of capacity to navigation apps and, beyond that, feeds congestion pricing. The team already uses their stack to power toll roads in Austria; the step from facility management to city levers is shorter than it seems when the instrumentation is in place and trusted.

Integrations remain an opportunity and a constraint. From a driver’s perspective, the best experience is seeing availability directly in the navigation app they already use. Yet platform companies tend to prefer aggregators with comprehensive coverage over point solutions. That creates a strategic question for vertical operators: should you become an aggregator, partner with one, or expose APIs and let the market decide? Meanwhile, standardizing data models for occupancy, pricing, and enforcement would help everyone move faster. The point stands: AI only reaches end users at scale when it is integrated into their existing journeys—and that is as much business development as it is engineering.

Reliability is a growth strategy. A Net Promoter Score in the mid-80s is not common in B2B infrastructure. It is possible here because the customer sees daily outcomes: fewer complaints, better turnover, new revenue from off-hours bookings, faster exits. The team uses NPS as a steering tool: high-scoring accounts are candidates for early pilots; low-scoring ones trigger customer success investigations before a renewal is at risk. This sounds simple, but it is part of the discipline of running AI systems as services. The model is invisible to the buyer. What they feel is reliability and responsiveness. Manage those, and you can justify multi-year contracts and expansions without chasing growth with discounts.

Sales, installations, and supply chains are part of the AI. In the early days, the founders wrote handwritten letters and walked into stores. Today, growth is a machine: inbound marketing so lot owners can discover the solution, outbound SDRs who start conversations on LinkedIn and phone, and a strong presence at industry events where property owners congregate. Then comes the most underestimated piece: installation. A dedicated team designs placements, confirms power and connectivity, coordinates with local contractors, stages and ships hardware, and watches timelines and quality across countries. None of this is glamorous. All of it determines whether your computer vision system will still be working on a rainy Sunday when a hospital shift turns over. In applied AI, the last mile is literal.

ROI is plural. There is no single beer-mat formula because “parking” spans retail lots, hospitals, airports, municipal assets, and private developments—each with different economics and stakeholders. For a retailer, the win may be eliminating abuse by non-customers and increasing turnover. For an office complex, it may be monetizing overnight and weekend surplus without staffing gates. For a hospital, it may be throughput and reduced friction for anxious visitors. For toll roads, it is low-friction enforcement and dynamic rules. The platform’s revenue reflects this variety: revenue shares on fees where applicable, fixed monthly service fees elsewhere, and, in some cases, leases that shift demand risk onto the operator. The throughline is that AI is not sold as “AI”; it is sold as clear outcomes in a language each owner understands.

Use AI to see, not to pry. Once cameras are on site, the temptation to do more is real. The team is expanding carefully into event detection that property owners actually need: accidents, snow coverage and plowing status, and weekend gatherings like unauthorized tuning meets that create noise and litter. Notifications and a chat-like interface surface these events without sliding into surveillance theater. The rule of thumb is simple: if the signal increases safety, uptime, or compliance without profiling individuals, it is on the table. This keeps trust with both owners and the public—and aligns with the legal constraints already in place across Europe.

Focus beats possibility. Step into any meeting and you will hear dozens of adjacent ideas: cashierless stores, loss prevention analytics, curbside automation, EV charging optimization, fleet dispatch. Many are attractive; not all are wise. The founders are candid that the real challenge is not dreaming up new use cases but saying no to most of them, and doing a few to depth. In practice, this means a backlog full of intriguing but untackled work, and a roadmap anchored to what existing customers will pay for using the hardware already deployed. That is not a lack of ambition; it is how you earn the right to be ambitious later. In hard tech, compounding advantages come from reliability, not novelty.

What enterprises can learn from this playbook:

- Start with conversations, not code. Identify acute pain, a first customer, and a specific job to be done before you architect anything.

- Define accuracy as a product requirement. Write down your error budget, placement constraints, and drift mitigation before you choose a model.

- Treat privacy as a feature. Design for lawful purposes, minimal retention, and aggregation. Document your choices and make signage part of the UX.

- Build the go-to-market as seriously as the model. Plan for inbound, outbound, events—and the installation and support muscle to keep sites healthy.

- Use NPS operationally. Pilot with promoters. Intervene early with detractors. Reliability is what renews multi-year contracts.

- Layer adjacent value on the same footprint. Start with the MVP that pays the bills; add bookings, dynamic pricing, and event detection later.

- Expect regulators to be weekly stakeholders. Build relationships, not just filings. Predictability earns trust and speeds approvals over time.

Looking ahead, autonomy will test these systems in new ways. If self-driving timelines land closer to 15–20 years (with US cities moving faster than Europe), the economics of movement will shift. When vehicles can drive themselves, point-to-point travel becomes cheaper; most projections anticipate more traffic, not less. Parking demand will not disappear; it will reshuffle. Vehicles may stage in cheaper areas minutes away from destinations, then reposition on demand. Dynamic curb pricing and access rules will matter more, not less, and the winners will be the operators who already run trusted, accurate instrumentation across a city. The computer vision in a garage today could be the meter for a low-emission zone tomorrow. The throughline stays the same: accuracy, privacy, and clear value to each stakeholder.

The takeaway is simple, and it is easy to underestimate: the AI that changes cities will look boring. Cameras mounted properly. Contracts that run a decade. Checklists for signage, retention, and calibration. Customer success teams who call people back. But boring is what scales. If you are serious about putting computer vision into the world, trade a bit of demo magic for discipline. Go where the pain is loud, measure the thing that breaks trust if it drifts, and bake compliance into your design. Do that, and your AI will not just survive the parking lot; it will become infrastructure. Listen to the full episode here, out on all platforms: https://share.transistor.fm/s/9ddf69e0