What an AI agent does
An AI agent works towards an objective, plans intermediate steps, uses tools, observes results and adjusts its actions. Definitions vary, as the OECD’s review of agentic AI discusses.
A conversational assistant primarily responds to prompts. A rules-based automation follows predefined logic. An agent can choose some intermediate actions within its permissions. Products may combine these approaches, so ask a supplier exactly which decisions the system makes and which it cannot make.
Capability and reliability are different questions
Benchmark performance does not establish reliability in your workflow. Longer tasks create more opportunities for error, and errors can propagate between steps. As a simple illustration, ten independent steps that each succeed with 95% probability would succeed together about 60% of the time. Real workflows do not necessarily satisfy those independence assumptions; test them directly.
Forecasts of adoption or cancelled projects are forecasts, not measured outcomes. For an investment decision, examine a bounded task, its failure modes, full cost and business value.
Possible uses and prerequisites
Customer service. Search knowledge, investigate an issue, propose a resolution and escalate. Prerequisites include current knowledge, defined permissions and a clear route to a person.
Sales. Research accounts, help qualify enquiries and prepare CRM updates. Agree qualification criteria and approval rules for customer communications.
Operations. Reconcile information, check documents and assist with incident handling. Data access, permissions and exception handling need explicit design.
Finance. Assist with reconciliations, collect supporting information and prepare close activities. Preserve records of actions and controls over changes.
HR. Answer routine questions or assist with administrative preparation. Recruitment and employee assessment require particular care and may fall within high-risk AI rules, depending on the use.
When reviewing a case study, ask for the process before and after, volume, errors, human involvement and full costs. A supplier or customer announcement should be identified as such.
Start with the actual process
Procedures rarely contain every exception and informal judgement that people use. Map the work with the teams involved before giving an agent responsibility for it.
Four questions help identify a suitable candidate:
Is the actual workflow understood, including exceptions?
Is the task bounded, with a result that can be checked?
Are possible errors manageable, with a tested human recovery route?
Is a business owner responsible for the results?
Resolve unanswered questions before expanding autonomy.
Measure the first pilot
Record volume over a defined period, human intervention, detected errors and a sample check for silent errors. Measure supervision time and the full cost per successful task, including integration and review. Compare with the same process without the agent. Agree success and stopping criteria before starting.
Organisation and assessments
Make organisational decisions with a clear assessment
Our assessments identify practical changes to help your teams and processes work together.
An example from Spentia. Aria supports confidential individual interviews, with everyone in the agreed scope invited to participate. Collection is organised over three days. Senior human analysis and executive decisions remain distinct from the agent’s collection role. The boundaries of the work and the handling of individual contributions are defined in advance.
Supervision is an ongoing responsibility
Someone needs to monitor quality, adjust instructions, handle escalations, investigate failures and report results. Business ownership and technical support should work together. Confidence thresholds, anomaly checks, human sampling and a stop mechanism need to be designed and tested.
The EU AI Act classifies systems by use and risk rather than by the label ‘agent’. Applicable duties therefore depend on what the system does and the organisation’s role. The revised timetable places Annex III high-risk rules in December 2027 and high-risk AI embedded in Annex I regulated products in August 2028. Other duties, including applicable transparency rules, already operate. Check the relevant requirements for the particular use.
Frequently asked questions
What is an AI agent?
What is an AI agent?
A system that works towards an objective by choosing steps, using tools, observing outcomes and adapting within defined permissions. The degree of autonomy varies, so examine the specific product and workflow.
How does an AI agent differ from a chatbot?
How does an AI agent differ from a chatbot?
A chatbot primarily produces conversational responses. An agent may also take actions and choose intermediate steps across tools. Some products do both. Ask what it can change, what requires approval and how actions are recorded.
Will AI agents replace employees?
Will AI agents replace employees?
The effect depends on the tasks, performance and organisational decisions. Agents may take on activities while creating review, exception-handling and oversight work. Assess complete roles and workflows before drawing conclusions about staffing.
How many businesses use AI agents?
How many businesses use AI agents?
There is no single comparable measure across all definitions of an agent. General AI adoption statistics and supplier surveys do not necessarily measure production use of autonomous systems. Check the definition, sample and evidence behind each figure.
