Is your organisation ready for agentic AI?

Authored by David Sugden, Head of AI & Engineering at Axiologik.

Three shifts UK leaders should make in the next twelve months.

Agentic AI presents UK enterprises with a new challenge as it introduces actors inside the business that need managers, rather than engineers. The deciding factor for success over the next twelve months will therefore not be solely technical capability, but whether the organisation is ready to supervise an actor it did not hire.

The accountability and unpredictability problem

A 2022 Scientific Reports study by researchers at MIT, Harvard, Cambridge, and Monterrey observed temporal degradation in 91% of the machine learning models they tested (https://www.nature.com/articles/s41598-022-15245-z). That finding describes predictive AI, where the model returns a value, and a human or downstream system decides what to do with it.

Agentic AI, the category of system that plans and executes multi-step actions with limited human intervention, removes that decision point, resulting in the same drift playing out inside an autonomous workflow rather than at its boundary. Whereas a drifting prediction shows up as a wrong number on a report, a drifting agent shows up as a wrong action taken at speed, potentially repeated across thousands of interactions and all before anyone realises the workflow has changed.

Introducing the readiness model: evaluation, ownership, control

In May 2026, the NCSC, with its Five Eyes counterparts, listed five categories of risk that follow from this autonomy (https://www.ncsc.gov.uk/blogs/thinking-carefully-before-adopting-agentic-ai) and named strong governance, explicit accountability, and rigorous monitoring as essential safeguards and prerequisites for agentic systems. That translates to three connected capabilities that define whether an organisation is ready: 

• Evaluation is how it checks the actor's work. 
• Ownership is who manages the actor and who is accountable for what it does. 
• Control is what authority the actor has and how that authority is removed. 

Existing governance frameworks treat these as separate disciplines, but for agentic AI systems they collectively describe how an organisation governs a worker it did not hire.

Shift 1: evaluation moves from procurement to continuous, and from engineering to product

Most enterprises benchmark suppliers and components at procurement, then never again. Fortunately, when it comes to agents, continuous production evaluation patterns are already available, such as graders and LLM-as-a-judge, and deterministic guardrails can wrap non-deterministic decisions.

The harder question is who owns them. For a customer support agent, accuracy is important, but tone, compliance posture, and the escalation decision all matter equally, and none are technical judgements.

The board-level question is: who in your organisation signs off the escalation threshold for an agent that handles regulated customer interactions?

If the answer is the technology function, you are likely still treating the agent as a tool. Product should define what good looks like. Risk should define what acceptable failure looks like. And Engineering should instrument each workflow step.

Shift 2: ownership moves from functional silos to cross-functional accountability

Agents collapse the boundaries between functions, so ownership must collapse too. An agent touches data, application logic, downstream services and the business workflow it sits inside, and the team that owns it must oversee all four.

This is an operating model decision, not an engineering restructure. UK leaders should stand up cross-functional squads with end-to-end accountability for the agent.

The board-level question is: who in your organisation is the agent's manager?

If the answer is more than one name, or no name, ownership is incomplete.

Shift 3: control moves from approval gate to runtime authority

Traditional technology change management assumes the change is the code. For agents, change occurs every time the model is updated, the prompt is updated, a question is posed by the user, or an upstream data source drifts. Controls must therefore extend into operating at runtime:

• Detection - something is changing. 
• Containment - stop the change from spreading. 
• Reversal - return the system to a known good state. 
• Cost control - keep the runtime change inside the original business case. 

Together these are the equivalent of how an organisation supervises any other worker: it watches what they do, limits where they can act, can stand them down, and controls how much of the budget they spend.

The board-level question is: if a new hire started changing your customer-facing language without telling anyone, you would notice immediately. Would you notice if an agent did?

These three questions aggregate into a single readiness test. If organisations can answer them with a threshold, a name, and a runtime control, they are ready to put agents into production. If they cannot, the gap is organisational, rather than technical.

What the regulators are saying

Cross-sector UK regulators are converging on the same diagnosis. The Digital Regulation Cooperation Forum, comprising the CMA, FCA, ICO and Ofcom (https://www.drcf.org.uk/publications/papers/thefutureofagenticai), identified fragmented accountability as one of the top emerging concerns for agentic systems across financial services, consumer markets, communications and data protection.

Sheldon Mills, FCA Executive Director, put the question more directly at the launch of the FCA's long-term review into AI in retail financial services (https://www.fca.org.uk/news/speeches/fca-long-term-review-ai-retail-financial-services-designing-unknown) when he asked what 'reasonable steps' looks like when models update, incorporate components organisations don't directly control, or behave differently with new data. 

The regulators are asking the supervision question because the technology has produced a worker that existing supervision regimes were never designed to govern.

What readiness actually looks like

It’s clear the task is harder than buying and deploying technology. Evaluation owned by product and risk, ownership held by a single accountable squad, runtime controls authorised by the board – these are the structures of supervision, not of technology deployment.

If your organisation cannot define the threshold, name the manager, and enforce runtime controls, it is not ready for agentic AI. The UK organisations that lead the next twelve months will not be the ones with the best models or engineers; they will be the ones that learned how to manage a worker they did not hire.

This article first appeared as a guest blog on techUK.