Klaudia Zaika is the CEO of Apriorit, a software development company that provides engineering services globally to tech companies.
Business leaders often see the release of an AI system as the finish line. I understand why. This is where the visible effort ends, and a system that has passed every pre-launch test looks like a success.
Yet, critical failures often start post-release, when your AI system faces real-world data and unexpected use cases.
I want to share where my team and I see originally well-built AI systems fail once they reach real users, and what we usually advise our clients to do about it.
Silent Degradation Of AI
Even if it performed well at release, an AI system may slowly degrade and drift over time, struggling to handle real-world data that diverges the training datasets.
Long-running sessions and AI agents can also degrade. A 2026 intelligence degradation study found that the performance of the open-source Qwen2.5-7B model was collapsing significantly once it reached about 40-50% of its maximum context window. And while AI labs work hard to combat context rot, other models are still prone to it.
AI degradation is usually silent and hard to detect. To prevent it, your team needs to use such techniques as population stability index and shadow traffic comparison. And to manage context rot, feeding your agents only the context required for their current step may not be enough. My team often relies on context compaction and sub-agent delegation to optimize the use of the context window of their AI tools.
Lack Of AI Output Quality Monitoring
Most production AI systems have some observability in place, partly because regulations push for it. They commonly focus on what the AI system does and why, but rarely pay attention to the actual quality of the outputs. So, while according to a LangChain survey, agent observability is implemented in 89% of organizations, what they observe are basic performance metrics and tool calls.
Instead, businesses need to pay attention to AI-specific metrics such as data drift, hallucinations, reasoning loops and token consumption. Having these metrics logged and assessed on a regular basis is necessary for monitoring and improving output quality.
You can start by defining the core use case your AI system serves and what an accurate output should look like for each. Then, use these references to evaluate the model’s live outputs and plan further improvements.
Liabilities Caused By Wrong Answers
One policy, price or refund rule invented by your AI system is enough to create a legal problem for your business. Courts are beginning to hold companies responsible for statements made by AI systems they deploy.
One of the first precedents came in 2024, when Air Canada was ordered to honor a bereavement discount hallucinated by their chatbot. And in 2026, a German court ruled that Google is directly responsible for false statements about identifiable third parties made by Google Search AI Overviews.
Regulators are formalizing these responsibilities as well. In particular, the EU AI Act requires providers of high-risk AI systems to continuously monitor them after release and report serious incidents.
Unfortunately, there’s no proven way to completely eliminate hallucinations from AI systems for now. But there are ways to minimize your legal risks. In particular, I always recommend that my clients identify which answers their business can’t afford to get wrong and assign a human owner to each. Another helpful measure is to keep well-documented monitoring logs in case you need to prove your diligence to a court or regulator.
Unlimited Access Granted To AI Systems
By default, an AI model can’t reliably distinguish legitimate instructions from malicious prompts hidden in the data it processes, so no prompt wording can fully protect a system exposed to untrusted content. The more access an AI system has, the more damage a single manipulated prompt can do. This is a risk that even a thorough pre-release security testing can’t close.
OWASP’s 2026 State of Agentic AI Security and Governance analysis identifies prompt injection as one of the leading causes of failures in production agentic systems. The worst part is that these malicious instructions, capable of compromising the entire system, may hide in any data your AI processes, from shared files to online pages.
In the projects my company works on, we always start by implementing least-privilege access and regular security audits of the live system. For workflows with critical impact, human approval should be mandatory before any actions are taken.
Release Is Where The Real Work Starts
What happens with your AI system after its release can affect your business operations, regulatory compliance and customers, so it’s important to not only recognize the need for continuous improvement of your AI system but to also actually plan key post-release measures well before you launch it.
With complex AI solutions, post-release work may be just as tedious and resource hungry as the initial development, and business leaders should account for this in advance.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?