Increasingly, a situation arises when employees use neural networks more actively than the business owner. While he is eyeing the technology, the team is already preparing calculations and proposals using artificial intelligence (AI). It is difficult for the manager to assess the method of obtaining them. Responsibility for the business remains with the owner, and the rules for using AI are formed without his participation.
When a leader lags behind his own team
The use of AI outside of agreed rules and controls is called Shadow AI. An employee may genuinely want to do the job faster. However, if the owner himself almost does not work with neural networks, it is more difficult for him to notice the substitution of fact by assumption, check the calculation and assess the consequences of the recommendation.
An employee’s personal account reinforces this opacity. The manager does not know what instructions are given to the assistant, what files are downloaded, what is left in the history of the dialogue and what restrictions are taken into account. He agrees on the outcome, seeing no reason to refuse. A convincing form of response begins to work as an undeserved mark of quality.
The owner needs to understand what data the AI received, what he does not know and how to check the result. In the Work Trend Index 2026, Microsoft also notes the gap between employees’ capabilities and organizations’ willingness to use them.
Neural network sees the department and misses the company
An employee usually conveys to the AI that part of reality with which he works. He can describe his task well, but does not know the limitations of other departments and strategic priorities. Without shared sources, the personal assistant gets limited context. The phrase “think like a CEO” does not close this gap.
Imagine that a manager asks a neural network to prepare a profitable offer for a large client. The model recommends a discount and accelerated delivery based on sales history. However, the warehouse knows that the remainder is reserved for existing obligations. The finance department is aware of the customer’s late payment, and the procurement is aware of the delay in the next batch. This information was not included in the request. The conclusion, convincing for sales, can turn into a disruption of delivery and loss of money for the company.
In a conversation between colleagues, such a conflict can be found: one will clarify the availability of goods, the other will remind of the restriction. If the preparation of the decision was transferred to a personal chat, and the exchange between the departments was missed, the neural network will not receive these clarifications. It is able to prompt for missing data, but it cannot be expected to always recognize a gap.
The company risks automating departmental isolation. Recommending AI can improve a department’s score and make a business worse off. The overall context should include objectives, measure definitions, constraints, and approval procedures. At the same time, each assistant needs up-to-date information on the task and access by authority, and not all company data.
A beautiful explanation does not confirm the calculation
The manager receives a report signed by the analyst and assumes that the numbers are verified. The analyst could only edit the response of the neural network. In this case, an error occurs both in arithmetic and in the formulation of the problem itself: for example, in calculating profitability, they forgot returns and urgent delivery.
The request “show the course of reasoning” does not make the answer reliable. The explanation may also contain errors. For a significant calculation, initial values, formulas, assumptions and verification in the calculation tool based on the primary data are needed. A second neural network with the same incomplete sources can repeat the first conclusion. The company must determine who prepares the recommendation, who checks its grounds and who has the right to make a decision.
Data and ways of working remain out of control
To disclose sensitive information, it is not necessary to send the Client Base file. Correspondence, discount terms, cost and reasons for rejections in several requests together reveal the economics of relations with the client.
Transferring data to a service does not mean publishing it. The risk depends on the conditions of service, storage, use of data for training and access to them. When working through personal accounts, the company may not know these conditions. Removing the surname also does not guarantee depersonalization: a person or deal sometimes allows to find out the remaining details.
Instructions and patch history remain in your personal chat. Colleagues get documents, but the method of obtaining them may not be available. In my own system of working with AI, I separately set the task of collecting verified information, working proofs and analysis of failures. The company must maintain knowledge of how the result was obtained and why errors occurred.
Agent can be discreetly delegated
The AI agent is able to act in connected systems: send letters, change records, start operations. Microsoft’s agent security guidelines highlight agent accounting, minimum access rights, and activity monitoring.
Permission to prepare a letter is not equal to permission to send it to the client. Access to reading the transaction card should not automatically give the right to change its terms. Each assistant needs a responsible person, action limits and a way to stop work quickly.
What the owner should do first
It is worth starting with the training of the leader and the analysis of one process. For example, go with the team the path of a commercial offer: from a request to AI to promises to the client. Find out what services are involved, what information they receive and what they can do.
Next, establish general rules of work: permitted services and accounts, permissible data, verified instructions and mandatory approvals. Assign people responsible for the relevance of the knowledge base and restrictions on their direction. A local task must be associated with the interests of the entire company. Buying a corporate subscription alone doesn’t do that.
Employee training should include incomplete data. It is important to be able to set a task, check sources and calculations, notice the missing context and contact a colleague in time. The skill test is simple: give an example in which the answer for one department looks successful, but violates the obligations of another. The employee must detect the conflict before sending the result.
Security needs to be translated into actions: delimit access, define storage rules, record significant operations and revoke access when a role is changed or dismissed. For critical processes, you need an employee who can pick up work, and an understandable order of error reporting. Promotion of the initiative should be combined with responsibility for violating established boundaries.
I expect that in the next three to five years, with the proliferation of agents, the ability of the manager to coordinate the work of people and AI at the company level will become more and more important. Without rules, training and data protection, even conscientious use of neural networks creates significant risks. Employees can accelerate their tasks while the business loses understanding of its own solutions.
By Dmitry Miroshnikov, Vice President of Business Development Corporation for Corporate Strategy