The measured size of it
- 80%
- Organisations meeting risky AI agent behaviour
- 10%
- Maximum CMA penalty of worldwide turnover
Lloyds Banking Group's agentic financial assistant began reaching customers in early 2026, built for a base that passes 21 million once the roll-out completes, and it is not arriving in one go: functionality expands in stages, EntrepreneurPlus reports. NatWest is running agent trials in complaint handling on a similar premise, inside a process where a regulator judges the result. The interesting question about both is not which model sits underneath. It is who signed off each widening of what the software may do on its own, and where that signature is filed.
Most firms are nowhere near that discipline. Manu Swami, chief technology officer at Sonata Software, writing for ET CIO, reports that McKinsey research found 80% of organisations have already met risky behaviour from an AI agent. His diagnosis is architectural. Enterprise systems were built on a division of labour in which applications execute, people decide and governance draws the line between them. An agent that reads a purchase order, chooses a next step and posts a transaction has been handed a slice of business judgement, and the control stack around it, identity, security and monitoring, was never asked to describe which slice.
So the question in the title is not about capability. It is whether your firm can produce, on request, three things: the decisions it has delegated to software, the limit on each one, and the person whose name sits beside the outcome. A firm that cannot produce those three things is not deploying agents. It is exposing itself to them.
The liability is already live, and it does not wait for agent-specific law
The GOV.UK guidance on complying with consumer law when using AI agents, updated on 30 October 2026, puts the position plainly: the same consumer rules apply whether a customer deals with a person or an agent, and you are responsible for what an agent does in the same way you are responsible for an employee. That holds even where someone else designed or supplies the agent on your behalf, which turns supplier due diligence into a consumer law question rather than a procurement nicety. The exposure complying with consumer law when using AI agents names runs to 10% of worldwide turnover, with compensation for affected consumers on top.
The regulators have since said the same thing in public. The four bodies that make up the Digital Regulation Cooperation Forum, the CMA, the FCA, the ICO and Ofcom, issued a joint statement on 30 October confirming that existing frameworks already cover personal AI agents, as DailyAI reported:
An AI agent's autonomy does not diminish the responsibility of organisations that deploy it to comply with applicable law and relevant regulation.
That statement adds no new rules, which is precisely why it should worry a board. Nothing needs to be legislated for a firm to be caught. The CMA has been able to decide for itself that consumer law was broken, without going to court, since April 2025, and the same reporting puts its ceiling at 10% of global turnover.
In the European Union the route is different but the documentary demand is the same. The AI Act Service Desk explains that "AI agent" is not a separate legal category under the AI Act, and that agents are covered because they are AI systems built on general-purpose models, so the rules for both apply to them. Transparency duties under Article 50 bite from 2 August 2026 where an agent interacts with people or generates content, and high-risk obligations follow from 2 December 2027 and 2 August 2028. UK and EU positions compared on the attributes that decide how you evidence a deployment:
The table below draws on Complying with consumer law when using AI agents - GOV.UK.
| Question | UK position | EU position |
|---|---|---|
| Is "AI agent" a defined legal category? | No. Consumer, data protection and sector rules apply to the conduct | No. Agents are covered as AI systems and as general-purpose models |
| What triggers the obligation? | The effect on a consumer, whoever or whatever produced it | The system's role: interaction with people, content generation, high-risk use |
| When do agent-facing duties start? | Already. Direct CMA enforcement of consumer law has applied since April 2025 | Transparency from 2 August 2026; high-risk requirements from 2 December 2027 and 2 August 2028 |
| What is the named exposure? | Up to 10% of worldwide turnover, plus consumer compensation | Set by the AI Act and the applicable market surveillance regime |
Read the two together and the practical consequence is identical. You cannot answer a regulator by describing your model. You answer by showing what the agent was authorised to do, what it actually did, and who held the authority in between. A defensible record for a single agent looks like this:
- the decision the agent may take, written in the language of the business process
- the systems it may read and, separately, the systems it may write to
- the limits, expressed in money, time, volume or category
- the escalation route and the named role that receives it
- the business owner who signs the mandate and answers for the outcome
- the review date, the change log and the conditions under which the agent is retired
Very few UK firms can produce that list today, which is a competitive opening. EntrepreneurPlus reports Gartner's forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear value and inadequate risk controls, and cites DSIT research from 2025 finding that only 7% of UK businesses using AI report using agentic AI specifically. A small estate is the cheapest possible moment to build the register, because the first version has few entries and the habit forms before the pressure does.
A firm that can handle agents writes the mandate before it picks the model
The document that separates a controlled deployment from an expensive experiment is a written mandate per agent, agreed by the business owner and enforced by systems the model cannot edit. Start from the decision. Take a refunds agent: it can read the order, the returns policy and the payment record, and computing the amount is trivial. The question that matters is whether it holds the authority to waive the extended returns window for a customer who is four days late, and no amount of model capability answers that. Someone in the business decides, and that decision becomes a rule the surrounding systems enforce.
The architectural move ET CIO argues for is the one that survives contact with an auditor: separate the intelligence that proposes an action from the authority that permits it. A mandate record can carry that separation explicitly.
agent: refund-assessment
business_owner: customer-operations
may_decide:
- calculate_refund_amount
- confirm_goods_received
may_recommend:
- waive_extended_returns_window
- refund_outside_policy
must_escalate_to: team-leader
authority_reviewed: quarterly
This is testable without opening the model. Either the waiver appears in the escalation log or it does not.
Microsoft's Cloud Adoption Framework sets out the same principle at estate level: a baseline of governance and security requirements every agent must meet before it operates, a central registry listing each agent's owner, purpose, platform and access scope, and a distinct identity per agent so that its actions are attributable and revocable. Its reasoning includes the observation that untracked deployments are a security and a cost problem at once, which is the argument that tends to move finance faster than any ethics paper. The framework also insists policy is applied the same way to first-party, custom and third-party agents, because the agent a department bought with a credit card is still your agent.
Monitoring, meanwhile, is not assurance. Availability and latency tell you an agent is running. Assurance has to test the conditions where behaviour changes:
- tools that return nothing, or return the wrong record
- instructions injected into the documents an agent reads
- edge cases sitting exactly on the boundary of its mandate
- a model version change, and a prompt change made on a Friday
- a supplier's silent API update, and a data source that quietly alters shape
- a policy change, such as a new returns window, that the agent has not been told about
- two agents that are each within their own limits but wrong together
Those checks need somewhere to go: the power to restrict permissions, to require human approval, to suspend execution and, where the underlying systems allow it, to reverse the action. Built after the agent is in production, none of that is available.
Cost, finally, is a governance signal. dp.vision's 2026 cost breakdown puts real numbers on the three tiers most firms move through:
| Tier | Published cost | Typical scope | What it demands of you |
|---|---|---|---|
| Off-the-shelf agent inside a tool you already pay for | $0 to $200 a month | Tier-1 support, meeting notes, first drafts | Named owner, clear disclosure to customers, a reviewed sample of outputs |
| Configured agent | $200 to $5,000 a month, with $5,000 to $30,000 of set-up | Workflows trained on your own documents and systems | Written mandate, escalation rules, logged actions |
| Custom-built agent | $30,000 to $150,000 to build, reaching $500,000 or more at enterprise scale, then $5,000 to $20,000 a month | End-to-end processes spanning several systems | Decision rights, continuous assurance, rollback, independent review |
The same analysis warns that the most common mistake is commissioning the custom build when a cheaper tier would do most of the job, and that automating a chaotic process simply produces chaos at higher speed. Both failures are governance failures wearing a budget's clothing.
Where the account sits is the last question and the one firms answer worst. Technology teams can build the controls, but they cannot decide what a customer dispute should settle at. That belongs to the business unit running the process, and it belongs to them at retirement as much as at launch: an agent should not stay live because the platform is convenient. Maturity, on this reading, is not the number of agents deployed. It is how much business-critical work you can hand over while still being able to name the person who carries it. Staged expansion, the pattern Lloyds and NatWest are following, only works if each stage is a decision someone signed before it happened.
Sources
- Complying with consumer law when using AI agents
- How are AI agents addressed within the AI Act?
- UK Regulators: Firms Can't Blame Their AI Agents
- Your AI agents are ready for business. Is your business ready to own them?
- What Are AI Agents for Business? UK Use Cases in 2026
- Govern and secure AI agents across the organization
- AI agents for business 2026: cost and where to start
Source: Your AI agents are ready for business. Is your business ready to own them?, ET CIO





