Laszlo Nagy, Founder and CEO of DuoClarity, on why an approved AI system drifts after launch and what it takes in people and budget to keep it working.
An AI project gets its hardest scrutiny in the weeks before it launches. A committee reviews the use case, security signs off, legal clears the data, and the system goes live. Six months later the model behind it has been updated twice, the data feeding it has moved, and the people using it have found applications that were never in the original scope. Almost nothing in the approval process covers any of that, and the questions it leaves are the ones that decide whether the system still works.
Laszlo Nagy is Founder and CEO of DuoClarity, an advisory practice that builds production AI systems for enterprise leadership teams and growth-stage companies. He came to the work through electrical engineering and industrial automation, then embedded and automotive systems, and then enterprise cybersecurity, where he held global product ownership for security infrastructure at a multinational manufacturer. He later served as SVP of AI Product Management at IgniteTech, an enterprise software company that runs operations on AI. His own firms run on AI in daily production, so the maintenance burden he describes is one he carries himself.
"Models are changing, capabilities are changing, data is changing, and usage is changing. AI is an IT system. An IT system needs to be maintained, and this is the most complex IT any human has ever seen," said Nagy. Traditional software only changes when someone changes it. An AI system can start producing different results while the code sits untouched, because the vendor updates the model underneath it and the data feeding it keeps moving. Keeping it working is a different job from getting it live, and it starts the day the launch team moves on.
Build and forget: A pilot proves less than most teams think it does. It runs for weeks against a fixed problem, watched closely by the people who built it, and none of those conditions survive a launch. Most of the organizations Nagy advises aren't close to that problem yet, since they are still working out how to start at all. "The biggest gap is everybody is thinking of AI like 'build and forget', which is totally wrong," Nagy explained. "You build it once and after that comes the fun part, which is much harder than the actual pilot."
A backlog of reviews: Speed is what breaks the review model. A system producing at machine pace generates more output in a day than a team can read in a week. Degradation surfaces late under those conditions, once someone notices the output has become unusable and the drift has been running for weeks. "Instead of having a huge backlog of work, you're going to have a huge backlog of reviews, which nobody is doing," Nagy said.
Nagy puts sampling in place of full review. Pulling a handful of outputs every day or two and checking them against the quality the system produced at launch catches drift early enough to act on. Nagy's own news editorial ran unattended for months, collecting from several hundred sources daily, removing duplicates, extracting against specific terms, and producing a daily brief without degrading. "It's a non-deterministic code-based operation," Nagy said. "You tell the AI, and that listens as best as it can or wants, and it's just like people."
Three named owners: Before a system goes live, Nagy wants three separate people assigned to it, each with a different job. Splitting that work across just one or two leaves gaps, which usually go unnoticed until something goes wrong. "For an AI system to run, you need a minimum of three responsible parties," Nagy noted. "You need financial, you need maintenance or operations, and you need governance and someone responsible for data."
Cost is metered: In Nagy's view, the financial owner has the hardest job of the three. A personal subscription is a fixed line on a budget, but a production system isn't, and the figure changes with every workflow and user added. Nagy put heavy enterprise usage in the many thousands of dollars, with caching available as one lever that cuts the bill on workloads that qualify. "When you put it in a system, a subscription-based approach doesn't work," Nagy said. "It's API and it's token-based, and it costs as much as you use, and for significant usage, it's significant cost."
Governance written for humans: Nagy warns against closing a legacy system before the ownership and the operating model are in place. A transformation that removes the old process without replacing the governance around it leaves an organization exposed from both directions. "We have our own and it's been working for 100 years, but it was written for humans, not for AI," Nagy said.
Many AI projects are sold on the headcount they will remove, and Nagy believes it's the wrong goal. A system still needs people to run it, so the realistic outcome is a smaller team doing different work. Nagy removes the manual and repetitive steps first, then rebuilds the workflow around what is left. "Unless you transform your operations, governance, and how you think, it's not going to work," Nagy concluded. "AI is a different speed at a different level on different steps, and you can't adapt it to a human workflow."
The best editorial systems don’t happen by accident. Outlever builds them.
