The AI risks enterprises are creating faster than they can control

A square object with a purple light coming out of it

AI has moved from small trials into daily business operations. Employees use public tools to analyse documents, write code and prepare client work. Companies are also installing agents that can search internal systems, contact customers and act without waiting for instructions at every step.

The risk has changed with the technology. A wrong chatbot answer is annoying. An agent with access to customer records, payment systems or corporate communications can cause real damage before anyone notices.

Several business leaders are trying to address different parts of this problem. Cohere CEO Aidan Gomez has focused on enterprise data privacy and the danger of relying too heavily on a small number of external AI providers. While XFactorAi CEO John Margerison argues that companies should use controlled automation, with people checking and approving important actions. Accenture CEO Julie Sweet is working with large companies as they move AI from trials into core operations.

Together, their work points to five risks that enterprise leaders need to address now.

  1. Shadow AI is making company data impossible to track

Employees are not waiting for formal AI programmes. They are opening personal accounts, trying new tools and uploading work because it helps them finish tasks faster.

That creates an unofficial AI estate that IT teams cannot see. Confidential contracts, customer details and internal plans may pass through tools that have never been checked by security or legal teams. New AI functions can also appear inside existing software after its original procurement review, meaning an approved product can change without receiving fresh scrutiny.

Cohere has warned that tools approved for low-risk internal work can gradually move into higher-stakes customer tasks, while staff may begin using public services before clear rules exist.

  1. AI agents are receiving permissions built for people

An AI agent needs access to company systems to do useful work. The danger begins when it receives broad permissions, permanent credentials or access to information that has little connection to its task.

Traditional identity systems were built around human employees who sign in, complete an action and log out. Agents may operate continuously, move between several applications and carry out work on someone else’s behalf. That makes it harder to know which system acted, who authorised it and where responsibility sits.

Cloud Security Alliance research found that 53 per cent of surveyed organisations had seen an agent exceed its intended permissions. Almost half had experienced an agent-related security incident during the previous year.

Gartner expects the average global Fortune 500 company to have more than 150,000 agents by 2028, up from fewer than 15 in 2025. Yet only 13 per cent of organisations believe they have suitable governance in place.

That is not simply software growth. It is the rapid creation of a new workforce with unclear access rights.

  1. Small mistakes can now trigger chains of action

AI systems have always made errors. Agents raise the stakes because they can act on those errors.

A system might misunderstand an email, update a customer record and trigger a follow-up message. Another agent may then use that incorrect record to change an order, produce a forecast or flag an employee. What began as one bad interpretation becomes a chain of automated decisions.

Gartner predicts that 40 per cent of enterprises will reduce the powers of autonomous agents or remove them by 2027 after governance failures become visible through incidents. The firm argues that controls should match each agent’s degree of independence and access, rather than placing every system into one of two camps: fully blocked or fully trusted.

This is where Margerison’s position is most relevant. His strategy calls for people to remain active participants and approve the final step, with AI earning greater trust through controlled use rather than receiving full freedom at launch.

  1. Dependence on a few providers is becoming a board risk

Many enterprises are building important processes on models and infrastructure controlled by a handful of external companies. That creates exposure to service outages, price changes, policy shifts and restrictions on where data can be stored or processed.

Gomez has argued that organisations which surrender control of their AI infrastructure may also give up control over data privacy, security rules and continued access.

The issue is not that every company should build its own model. Most should not. But leaders need to know what happens if a supplier changes its terms, withdraws a model or becomes unavailable. Contracts, portable data and alternative providers now belong in business continuity planning.

  1. Pressure for quick returns is weakening oversight

Boards want evidence that heavy AI spending is producing results. That pressure can lead businesses to expand a promising trial before they have tested how it behaves with real customers, sensitive information or unusual requests.

Sweet and other advisers working on large AI programmes face a difficult balance. Companies need to move quickly enough to learn, but not so quickly that oversight is added only after an incident.

The answer is not another long policy document. Enterprises need a live record of their AI tools, strict access limits, named human owners and approval points for decisions involving money, staff, customers or sensitive data.

AI risk grows when nobody can see the system, explain its actions or stop it. The companies that act now will not eliminate every mistake. They will stop one mistake from becoming a business-wide event.