By Sonica Aron, Founder & Managing Partner, Marching Sheep
An agent can generate a sales forecast, identify accounts at risk of being lost and recommend where the sales team should focus. But what if the recommendation differs from the manager’s knowledge of the customer or the data on which it was based was flawed? Somebody must still be the judge, the arbitrator.
Which risks is the business prepared to take?
This is where agentic AI shifts the skills debate. As AI systems perform multi-step processes and even execute parts of workflows with minimal human involvement, employees will have to assume greater responsibility for goal-setting, setting boundaries and auditing decisions during the process.
The critical issue for organisations is how to get people ready for this. It’s one thing to train employees how to make use of the AI tools, but in working with agents, the judgement of knowing what to hand over, what to double check and when to step in is quite another.
Learning to delegate becomes a skill Learn to Delegate–The Ability to Let Go
Delegation, which has usually been teaching someone else to do the work, so they understand the result they should achieve, and then tracking the work. Now, you will need to do some of that thinking about systems that can do connected tasks, often very fast.
First, set the task correctly. “improve customer retention” is a nebulous task, which would generate an unreliable outcome. Workers need to determine the business problem, define the desired result, decide which information the agent has access to, and construe the context within which it can act.
And they need to determine what decisions they will empower them to make and which decisions they will need to refer to another party. The agent may be allowed to compose a reply to a customer but not be authorized to process a refund over a certain value.
These are operational decisions. Employees have to be familiar with the business process sufficiently, to know where it makes sense to give some autonomy and where that might be inappropriate.
Judgement matters when the answer looks convincing
Work produced by AI can be fluent, comprehensive and incorrect. A worker may perform a workflow properly at each step based on out of date data, ignore an exception or achieve a goal that doesn’t make sense any more.
Employees should also consider why a particular output has been generated and scrutinise the evidence behind it, rather than simply how the output looks. This involves verifying whether the output data is up to date, making assumptions and testing recommendations against the practical context of the business.
Think about workforce planning. An AI system could spot jobs where it seems like there is potential for improvement, because staff members are doing similar work. A manager still has to look at the nature of the work, the skills a company might sacrifice, and the impact on staff and customers. The analysis can guide but not make the call.
Such judgement is especially crucial in decision-making where it impacts individuals, money, customers or regulatory compliance. Employees need to be able to appreciate the potential outcomes of a mistake, and be willing to challenge an AI suggestion, even if it appears to have the upper hand.
Understanding the workflow is more important than knowledge of a single tool
With agents embedded within all aspects of business, work will flow through humans and software agents as well as automation. Disruption at one point impacts all later processes.
They will have to visualise how their role is linked to other processes, where data comes from, who relies on the output, and what the impact is if something goes wrong. As an example, a finance worker might need to be aware that a wrong entry in an automated approval process could impact payments, reporting, and audits later on.
This requires system thinking and a cross-disciplinary perspective. It will also alter the mode of cooperation. A worker will need to clarify a decision, work with other employees who control other aspects of a process and troubleshoot any exceptions to an agent’s commands.
Furthermore, managers have another role-ensuring the delineation between people and agents is defined and that exceptions go to the correct individual and accountability is not lost in between systems.
Learning must be tied to real work
The skills needed to work with agentic AI will differ by role and develop as organisations roll out new systems. No one size fits all AI training programme will be able to ready all employees for the decisions they may need to make.
More sensible is to focus on factors where agents will be imposed and understanding what employees need to learn to oversee them. Customer service teams might need to evaluate escalations as well as the quality of actions taken. HR teams might need to scrutinise suggestions on candidates as well as employees. Finance teams might need to understand approval limits, data integrity as well as audit trails.
Employees should be able to attempt the responsibilities in a set-up, such as with an agent producing a bad result or experiencing an exception. Learning how to work with a system when it is successful differs from understanding.