Agentic AI in the UK: The Question Is How Much Authority We Give It
Imagine an AI agent identifying a stock shortage, checking supplier prices and preparing an order. If the purchase fits an agreed budget, it can proceed. If something looks wrong, it stops and asks a member of staff to step in.
This is a practical example of what agentic AI could bring to businesses. Rather than simply answering questions or suggesting the next step, an agent can carry out tasks across different systems.
But getting an agent to complete a task is only part of the job. Businesses also need to decide what it can do independently, where it needs approval and what happens when it makes a mistake.
Start with the work, not the technology
Teams spend time on routine, multi-step tasks. IT staff investigate alerts, finance teams check invoices, and retail employees monitor stock and contact suppliers. An AI agent could take on parts of this work, as well as tasks such as processing insurance claims or coordinating healthcare referrals.
The risk depends on what the agent can do. Checking a report is different from changing a customer account or releasing a payment.
Automation also won’t fix a process that already has problems. Outdated stock records or unclear approval rules can lead to mistakes, whether the work is done by a person or an AI agent. Businesses should review how a task works before deciding to automate it.
Retail is one area to watch
Consider a customer who asks an AI agent to find a suitable product, compare prices and place an order. Instead of browsing several websites, the customer gives the agent a set of requirements and lets it do the comparison.
Technology companies are developing protocols to support this type of commerce. Examples include Google and Shopify’s Universal Commerce Protocol, OpenAI and Stripe’s Agentic Commerce Protocol, and Google’s Agent Payments Protocol.
Also Read: Agentic Commerce Update, Mid-2026: What's Actually Changed and What It Means for Retailers
For retailers, this raises a practical issue: can an agent find and assess their products using the information available? Specifications, prices, stock levels and delivery details all matter. If those details are missing or inconsistent, a product may be passed over even when it meets the customer’s needs.
A second issue arises at the point of purchase. Retailers need to know whether an agent is authorised to place an order and whether it has stayed within the customer’s instructions.
Businesses do not need to wait for agent-led shopping to become commonplace before addressing these questions. Improving product data and reviewing purchasing controls are useful steps in their own right.
When something goes wrong, who steps in?
An employee checking an AI recommendation can spot an error before acting on it. An autonomous agent may already have taken several steps by the time anyone notices a problem. An order might have been placed, a payment initiated or a customer record changed.
That changes the governance question. It is no longer enough to check whether an AI-generated answer looks right. Businesses must also control the actions an agent is permitted to take.
Before deployment, teams should be able to answer four questions:
What can the agent do? Access to a business system should not automatically grant permission to change records, approve payments, or contact customers.
Can its actions be reversed? Correcting an internal report is relatively straightforward. Reversing a payment or putting right a customer-facing error may be harder.
Can the business establish what happened? Records should show which agent acted, what information it used and which permissions applied.
When must a person take over? Agents need clear limits and a way to pause or escalate situations they cannot handle safely.
Not every task needs the same level of supervision. An agent preparing a report might operate with little intervention. One handling payments or changing customer information needs tighter controls. Some actions may require confirmation every time; others can be automated once the risks are understood. Businesses also need to establish responsibility before an agent goes live.
Organisations cannot treat the technology as a substitute for accountability. The legal position will depend on the circumstances and the rules that apply, but businesses should know who owns the process and who responds when something fails.
Also Read: When an AI Agent Acts on Your Behalf, Who Is Accountable?
Identity matters when agents work together An agent may need to access a supplier’s system, use an external service or ask another agent to complete part of a task. Each handoff raises a trust question: how does the receiving system know who or what is making the request, and whether it has the necessary authority? Giving an agent the same permissions as an employee is rarely a sensible default. Account limit permissions to what the task requires and removes them when it is no longer needed.
If an agent delegates work, the next agent should not automatically inherit broader permissions. The UK’s digital verification services trust framework provides relevant context for wider discussions about digital trust. However, verifying a person’s identity is not the same as establishing what an AI agent can do. Businesses will need workable approaches to agent credentials, delegated permissions and the ability to revoke access. A practical route forward for the UK Businesses can begin with tasks that have clear rules, measurable outcomes and manageable consequences.
They can test the agent, examine failures, and expand its responsibilities as evidence supports doing so. There is little benefit in every team developing separate controls from scratch. Common approaches to permissions, activity records and human intervention could make it easier to deploy agents across an organisation. Industry and policymakers will also need to address identity and accountability as agents increasingly interact across company boundaries.
Agentic AI will not be judged solely on how many tasks it can complete. Businesses will also need to show that they can limit its authority, trace its actions and intervene when necessary. That is the work required to turn promising demonstrations into systems organisations can rely on.
