Most businesses got comfortable with AI as an assistant. You ask, it answers, someone checks the output, and a person decides what happens next.
AI agents work differently. You give them a goal and they plan the steps, open the systems, and carry the work through. They can update records, send messages, raise tickets, process refunds and move data between platforms, sometimes without anyone looking at each step.
That makes them useful, and it changes where the risk sits. When an assistant gets something wrong, a person usually catches it. When an agent gets something wrong, the action may already have happened.
For a scaling tech business, the question is no longer just which tool your team is allowed to use. It is what that tool is allowed to do on your behalf, and who carries the consequences when it does the wrong thing.
What changes legally when AI starts acting
An agent has no legal personality. When it acts, your business acts.
If an agent sends a customer a pricing commitment, your customer will treat it as your commitment. If it deletes or exposes personal data, that is your data protection incident. If it changes a contract record or approves a credit, your finance team and your auditors will want to know who authorised it.
Saying "the AI did it" is not a defence. It points to a gap in your controls.
So legal oversight has to move earlier. With an assistant, you can review the output. With an agent, the review has to happen up front, when you decide what it can reach and what it can do.
Set boundaries around the work, not just the user
Most AI policies focus on people: who has a licence, who can use which tool, what they can paste into it. That still matters, but it is not enough once agents can act.
Before an agent goes live, the business should be clear on:
- Which systems and data it can access, and which are off limits
- Which actions it can take on its own
- Which actions need a person to approve them first
- Who in the business is accountable for each agent and its settings
- Who can build, change or deploy agents, and whether those changes are reviewed
If nobody can answer these questions, the agent's real limits are being set by its default configuration, not by your business.
The actions that should always need a human
Not every action needs sign-off. Letting an agent summarise documents, triage tickets or draft a reply is low risk. The line sits where an action is hard to undo, leaves the business, or creates a commitment.
As a starting point, keep a person in the loop for:
- Sending anything external, including emails to customers, suppliers or regulators
- Agreeing prices, discounts, refunds, credits or contract terms
- Changing or deleting records in core systems such as your CRM, billing or HR platforms
- Making payments or approving spend
- Sharing personal or confidential data outside the business
- Making decisions that affect an individual's access to a service, a job or money
For longer workflows, build in checkpoints. Let the agent do the groundwork, then require approval before it moves to the step that matters. A review that sits at the end of a fifty-step process is easy to wave through. One placed before the risky step is not.
Automated decisions and UK data protection law
If an agent makes decisions about people, data protection law applies directly.
The UK General Data Protection Regulation (UK GDPR) rules on automated decision-making changed in February 2026. The old Article 22 was replaced with a new set of rules that give businesses more room to use automation, but with conditions attached.
Where a decision is made without meaningful human involvement and has a legal or similarly significant effect on someone, such as refusing a service, closing an account or rejecting a job application, you need safeguards in place. In practice, that means telling the individual, letting them make representations, giving them access to human review, and letting them challenge the outcome. Stricter rules apply where the decision uses special category data such as health information.
The key phrase is "meaningful human involvement". A person who clicks approve on everything the agent recommends is unlikely to count.
What your AI provider contract needs to say
Many agent tools are bought on standard online terms. Those terms were usually written to protect the provider, not to reflect what the agent will actually do inside your business.
Before you sign, check:
- Liability: Is the cap high enough to cover an agent taking the wrong action across your systems, or does it come to a few months of fees?
- Output exclusions: Does the provider disclaim all responsibility for what the agent does, leaving all the risk with you?
- Data use: Can the provider use your data or your customers' data to train its models?
- Data Processing Agreement (DPA): Does it reflect the processing the agent actually carries out, including the systems it connects to?
- Logs and audit access: Can you get a full record of what the agent did, and for how long is it kept?
- Model changes: Can the provider change the underlying model or its behaviour without notice?
An agent with access to your core systems is a significant supplier relationship. The contract should reflect that.
If your product puts agents in front of customers
Many tech businesses are now building agent features into their own platforms. That shifts the question to your customer terms.
Your contract should be clear on who sets the agent's permissions, who approves its actions, and who is responsible when it acts on a customer's instructions. It should also make clear how agent actions sit within your liability cap and what acceptable use looks like.
If your terms were written before agentic features shipped, they will not cover this. Enterprise customers will notice, usually during procurement.
Keep a record of what the agent did
When something goes wrong, the first question from a customer, an investor or the Information Commissioner's Office (ICO) will be: what happened, and who approved it?
You need to be able to show which agent acted, what it accessed, what it did, and where a person signed off. Without that record, you cannot investigate properly, defend a claim or show you had proper controls in place.
Audit trails can feel like admin. In a dispute, they are your evidence.
Summary
AI agents can take real work off your team. They can also take real actions your business is responsible for.
- Decide what each agent can access and do before it goes live
- Keep human approval for anything external, financial, irreversible or affecting individuals
- Check automated decisions against the current UK GDPR rules
- Make sure provider contracts reflect what the agent actually does
- Update customer terms if your product includes agent features
- Keep a clear record of agent actions and approvals
The businesses that get value from agents will be the ones that set the boundaries first.
If AI agents are moving into your operations or your product faster than your contracts and and policies, book a free 20-minute call with us today.