The appeal of agentic AI is easy to understand. If software can read information, reason about what it finds, open applications, and take action, businesses can hand over increasingly complicated workflows instead of merely asking a chatbot for answers. For Mikhail Voloshin, CEO and principal engineer at Mighty Data, that promise also exposes one of the most important problems in enterprise AI.
An AI system capable of taking useful action is also capable of taking the wrong action.
Voloshin has spent more than 25 years in software development, with experience spanning Microsoft, Google, finance, artificial intelligence, data science, and neuroscience. His work troubleshooting complex systems has led him toward a practical view of AI adoption. Companies need to think carefully about which decisions require machine flexibility and which boundaries should remain enforced by conventional software.
Instructions Are Not the Same as Constraints
Voloshin illustrates the distinction with an AI trading agent. A trader might want an agent to read financial news, examine charts, prepare a possible trade, and present the recommendation for approval. Because modern agents can operate browsers and software interfaces, much of that workflow can now be automated.
The danger appears when the agent can also execute the trade.
A user might instruct it never to submit an order without approval. Voloshin describes that as a “soft guard rail.” The restriction exists as an instruction the AI is expected to interpret and obey. Yet language models can misunderstand instructions, and an agent may find a different route to an outcome the user thought had been prohibited.
A hard guardrail works differently. Instead of telling the agent that it should not perform a sensitive action, the surrounding system prevents it from having the capability to perform that action in the first place.
That creates a more difficult engineering problem because useful agents need access. A financial agent may need to browse the web, retrieve information, interact with applications, and prepare transactions. The task is to preserve enough access for the system to be valuable while withholding the permissions that could create unacceptable consequences.
Let Software Do What Software Already Does Well
Voloshin sees another problem when businesses ask language models to handle tasks that conventional programs perform more reliably.
Imagine a spreadsheet containing hundreds of rows. A user tells an AI to perform the same analysis on every row. The model may process several entries and then produce language describing how it would continue through the remainder, rather than actually completing every iteration.
The mistake is treating an LLM as though it inherited all the reliable properties of conventional software simply because it runs on a computer.
“We already have machines that are capable of iteratively performing a repetitive task on a sequence of data,” Voloshin says. “They’re called computers.”
His preferred architecture combines the two. Traditional code can handle deterministic operations such as iterating through every row, maintaining state, or enforcing permissions. That software can call an AI model when a particular step requires interpretation or flexible reasoning. The LLM does the work for which probabilistic intelligence is useful without being asked to replace functions that ordinary software already handles predictably.
This distinction becomes more important as AI moves from generating text to operating business systems.
Better AI Starts With Knowing What Matters
Technical safeguards cannot solve a poorly defined business objective. Voloshin says one of the recurring problems he encounters is that companies optimize metrics before identifying what actually creates economic value.
He recalls working with a business-oriented social media company that viewed itself primarily as a provider of financial and corporate information. Yet its revenue came from advertising. From Voloshin’s perspective, the information attracted users, while the economically valuable product was their attention.
That distinction changes what data matters and therefore what an analytics or AI system should optimize.
The same principle applies to companies sitting on enormous quantities of unorganized data. Before selecting a model or deploying an agent, leadership needs to identify what the business actually provides, how that creates value for customers, and how that value becomes revenue. Otherwise, increasingly sophisticated technology can simply optimize the wrong objective more efficiently.
Voloshin expects AI to remain transformative even if today’s individual companies, products, and business models change dramatically. The opportunity for businesses is therefore larger than choosing the latest model. It lies in building systems that understand where AI is useful, where deterministic software is safer, and where human authority still needs to remain firmly in control.
Want more Plugged In? Read more or listen on Apple Podcasts.
