An IT auditor's view, drawing on research from PwC, McKinsey and Gartner.
From chatbots to agents
A chatbot answers questions. An AI agent takes action: it can plan steps, use systems such as email, ERP or CRM, and complete a task with little human input. For example, it can read an invoice, match it to a purchase order and route it for approval.
What the research says
- Adoption is real. In PwC's 2025 AI Agent Survey of more than 300 US senior executives, 79% said their organisations already use AI agents, and 66% of adopters reported measurable value, mainly in productivity. 88% plan to increase AI budgets because of agentic AI. (Digital Commerce 360 on the PwC survey)
- Scale is not. McKinsey's The State of AI in 2025 found 62% of organisations at least experimenting with AI agents, but only 39% reporting enterprise-level EBIT impact from AI. (CXOToday on the McKinsey report)
- Many projects will fail. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls, and warns of "agent washing", where ordinary software is rebranded as agentic. (IT Brief on the Gartner forecast)
Why AI agent projects fail
PwC observes that many companies still treat agents as "bolt-ons to existing workflows", and fewer than half (45%) are rethinking their operating models. McKinsey makes the same point: the value comes from redesigning the work, not from inserting an agent into an old process.
An IT auditor's checklist: 7 controls before you deploy an agent
- A clear owner and use case: one accountable business owner and a measurable goal.
- Least-privilege access: the agent gets only the system rights it needs, and nothing more.
- A human in the loop for high-impact actions: payments, customer communications, data deletion.
- A full audit trail: every action, decision and data source is logged.
- Testing and guardrails: including protection against prompt injection and manipulated inputs.
- Data governance and privacy: know what data the agent can see and where it goes.
- Cost monitoring and a kill switch: usage limits, and the ability to stop the agent immediately.
What it means for organisations in Pakistan
For banks, telecoms and other regulated sectors, an AI agent is a new type of privileged user. It should be governed like one, through access reviews, change management and logging, exactly as IT general controls require for any system that can move money or data.
The bottom line
AI agents are real and valuable. The winners will be the organisations that redesign the work and govern the agent, not those that simply buy the most tools.


