Vooban Quit Selling Advice - Then Made AI Earn Its Keep on the Factory Floor

A potato slides under a camera. A de-icing truck waits beside an aircraft. A recycling shredder chews through contaminated plastic. These are not the usual props in an artificial-intelligence pitch, which is precisely why Vooban is interesting. The Quebec City company has spent years putting algorithms in places where a wrong answer has a cost you can count: wasted cement, idle machinery, late pickups, extra fuel, overtime and aircraft sitting on the ground.

Vooban is a custom software and applied-AI firm, not a laboratory selling access to a single model. Its teams combine data engineering, operations research, computer vision, web and mobile development, cloud architecture and, now, cybersecurity. The customer is usually a mid-sized or large organization with a stubborn operational process. The deliverable is usually software woven into that process. That distinction keeps the story grounded. The company does not need to claim intelligence in the abstract when it can point to tons per hour.

First, stop selling the advice

Kevin Moore founded Vooban in 2011 after a decade as a developer. The first client was Canada's Department of National Defence, and the name nodded toward Sébastien Le Prestre de Vauban, the French military engineer associated with fortifications that influenced Quebec City's Citadelle. The original spelling was unavailable, so two O's did the administrative work of a moat.

The consequential decision came in 2017. Vooban stopped offering consulting services and moved fully to turnkey, project-based delivery. It formed its AI department in the same year. A consultant can hand over a recommendation; a builder must live with the integration, the interface, the dirty data and the operator who has to trust the result on a Tuesday morning. Vooban's differentiation grew from accepting that larger mess.

The market around it is crowded. A buyer can call Accenture, Deloitte or CGI, hire a specialist shop such as Moov AI or Osedea, or build an internal team. Vooban's answer is breadth with a deliberately industrial accent: the data pipeline, optimization model, custom application and deployment can sit with one project team. It is less a model vendor than a general contractor for complicated digital machinery.

The failure arrives before the model

Vooban's strongest case studies begin with the thing that failed first. At a plastics recycler, production varied by as much as 80 percent depending on the operator. Maintenance for shredder knives followed fixed intervals because the plant could not see how each action affected wear. Downtime consumed 27 percent of production time, and more than half of it was unplanned. The initial failure was not a missing neural network. It was an operation leaning on cameras, intuition and a calendar.

Vooban built a predictive model to regulate material flow and an operations-research system to select knives according to wear, production goals and incoming material. The reported result was a 40 percent lift in hourly output, from 2.67 to 3.74 tons, plus a reduction of more than 50 percent in unplanned stoppages. The useful lesson is almost boring: stabilize the process, then optimize it.

The same pattern appeared at OSCAL.AI, a waste-collection platform. Its old routing method simulated combinations without a true mathematical optimization engine. Humans still prioritized customers, multi-day planning remained awkward, and 58 percent of collections happened too early. Margin pressure and environmental requirements changed the calculus. Vooban built a constraint-aware engine that considered inventory, recurring stops, vehicle compartments and service windows. OSCAL.AI reported optimal-window collections rising from 34.5 to 93.1 percent and as many as 30 percent fewer routes for the same territory.

“Our teams in the field love it; they wouldn't want to go back.”Gabriel Lépine, co-CEO of Aeromag

What it costs, and what it has to return

Custom AI is not cheap, and Vooban rarely publishes contract prices. One public example is unusually specific: the company describes an unnamed maker of industrial sealing solutions that spent $1.2 million on an AI project. Vooban says the system now saves $3 million each month and avoids 300 metric tons of carbon emissions. Those are company-reported figures, not audited accounts, but the disclosure makes the right question unavoidable. A model is not valuable because it predicts; it is valuable because the prediction changes an expensive decision.

That structure explains Vooban's sales funnel. An innovation sprint identifies and ranks opportunities. A project blueprint examines feasibility, budget and estimated return. A proof of concept tests whether the idea survives contact with the client's data. Only then does the expensive scale-up begin. In its Google-backed AI Horizon program, the sprint ran four to six weeks, followed by a two-to-three-week project plan for candidates with proven integration potential.

The product is the changed decision

Consider Mirage, the hardwood-floor maker. Workers used Excel and hand sketches to plan how boards of different lengths should be packed, while robots waited for instructions. Vooban used combinatorial optimization to produce balanced, varied boxes with little leftover stock. Imagine Tetris, except the pieces are inventory and a bad score becomes waste. Deployed in roughly six weeks, the system doubled the factory's production rate, according to the case study.

At CSL Group, Vooban built models to predict ship arrival times and fuel consumption. The company reported 30 percent more reliable arrival estimates and 96 percent accuracy in fuel predictions, which can inform speed, route and hull-maintenance decisions. At Aeromag, models for de-icing mix and scheduling reduced average ground time by a minute at two airports in under three months. Vooban says the system creates more than $1 million in annual savings for airline partners. These examples make the company's customer profile clear: operators with physical constraints, existing systems and enough repetition for small improvements to compound.

The business model follows. Vooban earns project and services revenue from discovery workshops, blueprints, proofs of concept, custom development, cloud and data engineering, integration, security and support. It is not primarily a self-serve SaaS company. That makes revenue less automatic, but it also lets the firm chase problems that cannot be solved by installing another dashboard.

Buy the data layer, then sell the future

For 12 years, Vooban grew without outside investment. In September 2023, CDPQ made an undisclosed strategic investment, after Vooban said revenue had been growing about 50 percent annually. The institutional investor brought more than cash: access to a portfolio of companies that might need automation, and support for expansion into Ontario and the United States.

Six months later, Vooban acquired Stratéjia, a business-intelligence specialist. The deal expanded staff from roughly 130 to 190 and filled a practical hole in the AI pitch. Companies cannot deploy credible machine learning while their data lives in feuding spreadsheets. Vooban could now sell the sequence from digital strategy through data architecture to production AI. It opened a Toronto office in 2025, its third after Quebec City and Montreal.

The latest move adds product ambition. Vooban launched Vooban Labs to develop agentic and multi-agent systems, committing several million dollars over three years, and Vooban Cyber to secure models, agents and their daily operation. Labs says an internal multi-agent system changed how Vooban builds software and may be commercialized. Morphe, another offering, preserves expert knowledge in a searchable base for employees and agents. This is the interesting tension ahead: can a company trained to deliver bespoke projects turn its own methods into repeatable intellectual property without losing the operational intimacy that made the projects work?

The part you can steal

A manager does not need 225 technologists to copy the core move. Start with a repeated decision that already hurts: a late route, a rejected weld, an idle machine, excess overtime. Write down the current performance before anyone says “AI.” Find the person who makes the decision and the data they actually see. Estimate the annual value of a modest improvement. Then test one slice with a fixed deadline and a kill criterion. If a proof of concept cannot move the baseline, stop. A failed small test is cheaper than a successful demo nobody uses.

There is a limit to the playbook. A low-frequency process with poor data may never earn back custom development. A safety-critical decision may require human control, extensive validation and regulation that lengthen payback. A company unwilling to redesign work will turn even a good model into an ignored tab. And an agent with broad system access creates a larger security surface, which is why the Cyber division is more necessity than accessory.

Vooban has plenty left to prove. The funding amount and valuation remain private, the agent products are young, and case-study numbers are presented by the vendor. But its best idea is already portable: stop treating AI as the product. The product is a changed decision in a real workflow, measured against what happened yesterday. On a factory floor, there is nowhere for a vague promise to hide.