A customer changes an order. An AI agent checks stock, revises the delivery schedule, updates the finance system and sends confirmation. Nobody moves information between applications or chases a colleague for approval. The work progresses because software can interpret the request and act across the business.
For enterprises, that possibility changes the automation conversation. Traditional systems follow predetermined steps. Agents can select tools and decide how to pursue an objective. The opportunity extends beyond completing individual tasks faster. Yet giving software the ability to act creates a management challenge. Which decisions should it make? How much authority should follow it between systems? And how can leaders establish whether greater autonomy improves the business?
Answering these questions requires enterprises to examine how decisions and measures of success are organised.
Start With Decisions the Business Understands
The strongest candidates for autonomy have recognisable boundaries. Businesses need a clear objective, an acceptable range of outcomes and a practical way to detect failure.
“The best candidates share three traits: they are reversible, bounded and observable,” says Claire Lebarz, CTO of . “If an agent gets it wrong, you can undo the action cheaply. The action stays within a clearly defined scope. And you see the outcome quickly enough to correct course.”
That supports delegating activities such as routing requests, reconciling records, preparing shortlists and monitoring anomalies. High volumes and established procedures also provide a baseline against which an agent’s performance can be assessed.
The enterprise opportunity becomes particularly interesting where work repeatedly stalls between departments. A process may take days even though its individual tasks require only minutes.
Simon James, Managing Director of AI & Data Strategy at , points to commercial lending, where several experts may need to value a property before a contract can be prepared. Much of the elapsed time disappears into busy people’s inboxes. Agents could coordinate these dependencies and move work forward.
However, flexibility does not remove the need for clarity.
“What businesses aren’t ready to delegate is anything where the definition of success is still being argued about internally,” Lebarz told Silicon UK. “An agent can’t resolve an ambiguity the organisation itself hasn’t resolved.”
Before deployment, leaders should, therefore, map the decisions inside a workflow. Which rely on established policy? Which require interpretation? Which create commitments? This reveals where autonomy can safely begin.
It also discourages handing an entire process to an agent simply because its opening steps appear straightforward. Longer sequences introduce more opportunities for mistakes and misinterpretation. Checkpoints should sit where uncertainty or consequences increase.
Keep Human Judgement Where Consequences Matter
An agent can assemble evidence without having authority to decide what happens next. Enterprises need to preserve that distinction, particularly where decisions affect employment or significant financial commitments.
“Generally, final decisions should default to a human when there are material consequences for customers, colleagues or the organisation,” James commented. “Completing the work is one thing, signing off on the work being satisfactorily done is another.”
Procurement illustrates how substantial automation can coexist with human decision-making. An agent might engage suppliers and develop negotiating options while leaving the award decision with a sourcing professional.
“Any commitment that binds the business should stay with a person,” says Keith McFarlane, CTO of . “Our agent running in autonomous mode negotiates with suppliers and produces award scenarios, but it never makes an award on its own. A member of the sourcing team makes that decision.”
This division lets specialists focus on commercial judgement and relationships. It also makes responsibility explicit when price, resilience and long-term supplier value point towards different choices. Human oversight must nevertheless be designed around circumstances. Requiring approval before every action can undermine automation’s purpose, especially when delay itself creates harm.
Chris Talevi, Chief Information Security Officer at , identifies isolating a compromised host or revoking a leaked credential as situations where an agent may need to act first and face review afterwards. The conditions supporting that authority must be engineered in advance. “At a minimum, there should always be humans in the loop reviewing exceptions,” Talevi told Silicon UK.
Managers must, therefore, define escalation conditions and recognise when circumstances have moved beyond the agent’s remit. They also need enough capacity to intervene effectively. An approval queue that nobody can properly review offers little protection.
Domain expertise remains essential. Someone must understand when an apparently reasonable result misses a commercial relationship or an exception that historical data does not capture.
Build Authority Into the Operating Model
Agents working across systems create a problem that conventional application permissions may fail to address. Access to customer records, billing information and communication tools can combine into considerably more power than each permission suggests individually.
“Bounded autonomy has to be designed into the architecture, not described in a prompt,” says Cobus Greyling, Chief Evangelist at kore.ai.
Each agent needs its own identity, a named owner and permissions tied to its task. Financial thresholds, permitted recipients and approval requirements should constrain execution directly. These controls must remain effective when an agent encounters unexpected information.
External content introduces another complication. Supplier messages and customer submissions can contain instructions designed to manipulate an agent. Businesses should treat this material as untrusted input and carefully control how reading it can lead to external action.
“A prompt is a suggestion; a permission is a control,” Lebarz told Silicon UK.
Multiple agents also create coordination risks. A purchasing agent pursuing lower prices could weaken supply resilience. A service agent offering compensation might conflict with a finance agent restricting expenditure. Both could meet their targets while damaging the wider business.“ The difficult part of a multi-agent system is that every agent can be doing its job correctly and the business can still get the wrong outcome,” Greyling commented.
Leaders need clear priorities and a mechanism to resolve trade-offs. Shared measures should reveal whether improvements in one function create costs elsewhere. Where rules cannot resolve a conflict, a named person must own the decision.
Accountability also requires a reliable record. Enterprises should capture the information available to an agent, the tools it used, the actions it took, and the permissions in effect at the time. Changes to its authority need their own approval history.
Without that evidence, investigators may see where a mistake became visible while missing the earlier decision that caused it. Responsibility remains with the people who authorised and operated the system.
Measure Outcomes that Survive Scrutiny
The easiest autonomy metrics are often the least revealing. Tasks completed and hours apparently saved indicate activity. They do not establish whether customers received better service or the business became more profitable.
A useful measurement framework begins with the process outcome. Customer service teams might assess successful resolution and repeat contact. Procurement teams could examine realised savings alongside supplier participation and resilience. Finance teams should consider accuracy and how quickly usable information reaches decision-makers.
Coverage matters too. Agents may enable enterprises to address work previously left untouched because employees lacked capacity. That creates value even when the existing process becomes only marginally faster.
“In accounting or bookkeeping, and the world of finance more generally, an agent that is fast but wrong doesn’t save anyone time,” Stephen Edginton, Chief Product and Technology Officer at , told Silicon UK. “It simply moves the work to later, when someone has to find the mistake and fix it or pay a fine.”
This makes correction rates and downstream consequences essential measures. Your business should track how much work finishes correctly without intervention, how often employees revisit it and whether those patterns improve.
The financial calculation must include operating costs that demonstrations can obscure: model usage, integration, retries, monitoring, review and recovery. Lebarz recommends measuring “the true cost per successful outcome, including inference, retries, errors and the human time spent reviewing”.
Your business also needs to establish what happens to released capacity. Time savings become valuable when employees use that time to improve service or undertake work that previously went undone. Assuming every saved minute translates directly into cash will ratify the business case.
Deployment should progress through observable stages: shadowing existing work, making recommendations, acting with approval and then receiving limited independent authority. Expansion should depend on evidence from the relevant decision type, with reassessment when systems, policies or business conditions change.
The autonomous enterprise ultimately demands more deliberate management. Leaders must decide what good performance means, who owns each trade-off and how failure will be contained.
As Lebarz says: “Agents are very good at execution. Deciding what’s worth executing, and for whom, is still a deeply human job.”