Generative AI has already changed the way people search for information, create content and interact with software. The next development receiving significant attention is the AI agent — a system designed not simply to answer a question, but to take actions towards completing a broader objective.
For businesses, that creates some genuinely interesting possibilities. An AI agent might gather information from several systems, interpret it, decide what needs to happen next and initiate the appropriate action.
That sounds powerful. It also makes it important to separate the processes where agents can genuinely add value from those where conventional software remains the better choice.
Not every workflow needs an AI agent. Sometimes a good old-fashioned
if statement is still doing perfectly respectable work.
What Is an AI Agent?
Traditional software generally follows predefined logic. When a particular event occurs, the application performs a known sequence of actions.
An AI agent introduces a degree of reasoning into that process. Rather than being told every individual step, the agent can be given an objective and use available information, tools and instructions to determine how to proceed.
For example, an agent handling an incoming customer request might:
- understand what the customer is asking;
- retrieve relevant information from internal systems;
- determine which actions are required;
- use connected applications or APIs;
- prepare or execute an appropriate response; and
- escalate the request when human judgement is needed.
The important difference is that the workflow does not necessarily have to be completely predetermined.
Where AI Agents Can Add Value
AI agents are particularly interesting where work involves a combination of information, judgement and multiple systems.
Consider an operational team that receives requests by email. Staff may need to interpret each request, retrieve information from several applications, check internal rules, update a system and prepare a response.
Traditional automation can handle parts of that process, but variations in language and circumstances can make complete automation difficult.
An appropriately designed AI agent could potentially coordinate more of the workflow while still involving people when necessary.
Other potential applications include:
- research and information gathering;
- document classification and processing;
- internal employee support;
- customer service workflows;
- data investigation;
- monitoring operational exceptions; and
- coordinating tasks across multiple applications.
AI Agents Are Not a Replacement for Conventional Automation
This distinction matters.
If a process is deterministic and can be expressed clearly as business rules, conventional software is often preferable.
A calculation such as:
Quantity × Price = Value
does not need artificial intelligence.
Neither does a workflow where every possible state and transition is already known.
Traditional software provides predictable behaviour, straightforward testing and clear auditability. Introducing an AI model into such a process can add cost and uncertainty without providing meaningful benefit.
AI handles ambiguity. Software handles certainty.
In practice, some of the strongest solutions will combine both approaches.
An AI agent might interpret an unstructured document, for example, while conventional application code validates the extracted values and performs the financial calculations.
In other words, there is no prize for making every part of a system artificially intelligent.
The Importance of Tools and Integration
An AI model by itself cannot transform a business process.
To perform useful work, an agent normally needs controlled access to tools such as APIs, databases, document repositories, business applications or purpose-built functions.
This makes integration architecture particularly important.
An organisation needs to decide:
- what information the agent can access;
- which systems it can interact with;
- which actions it can perform;
- what limits apply to those actions; and
- when human approval is required.
Giving an AI agent unrestricted access to operational systems simply because it is technically possible is rarely a good idea.
A useful rule of thumb is that an agent should receive the minimum access necessary to perform its role — much like a human user, except the human user usually does not attempt twenty actions in three seconds.
Human Oversight Still Matters
The more autonomy an agent receives, the more important appropriate controls become.
Different actions carry very different levels of risk.
Drafting an internal summary is not equivalent to approving a financial transaction. Looking up an order status is not equivalent to changing that order.
Organisations should therefore think in terms of levels of autonomy rather than treating automation as an all-or-nothing decision.
An agent might be allowed to complete low-risk actions automatically while preparing higher-risk actions for human review.
This can provide meaningful efficiency without removing necessary accountability.
Security, Auditability and Governance
Agentic systems introduce many of the same concerns organisations already manage with traditional applications, but some become more significant because the system may be making decisions and invoking actions.
Important considerations include:
- what data the agent can access;
- how authentication and permissions are managed;
- which tools it is authorised to use;
- whether actions and decisions are logged;
- how sensitive information is protected;
- how unexpected behaviour is handled;
- how the system is monitored; and
- where human approval is mandatory.
For business-critical applications, these controls should be part of the architecture from the beginning rather than something added after the AI functionality has already been developed.
Start with the Business Process, Not the AI
Perhaps the most useful question for an organisation considering AI agents is not:
“Where can we use AI agents?”
A better question is:
“Which of our processes are difficult to automate today, and why?”
If the difficulty comes from interpreting documents, understanding natural-language requests, gathering information from different sources or dealing with reasonable variations in a workflow, AI may provide a useful new capability.
If the problem is simply an inefficient but completely predictable process, conventional software automation may still be the better answer.
Technology should follow the requirement rather than define it.
Where RAG Can Fit into an Agentic Workflow
Retrieval-Augmented Generation, or RAG, can also play an important role in agent-based systems.
An agent may need access to internal policies, procedures, product documentation, contracts, technical information or other organisational knowledge before it can make a useful decision.
Instead of relying entirely on information stored within the underlying language model, a RAG-based approach can retrieve relevant internal information and provide it to the model at the point it is needed.
This can make the agent more useful in organisation-specific workflows, particularly where decisions depend on proprietary or frequently changing information.
However, RAG does not automatically make an agent correct. Retrieved information still needs to be relevant, appropriately permissioned and presented within a carefully designed workflow.
Agents Work Best with Clear Boundaries
The most useful business agents are unlikely to be completely autonomous digital employees wandering freely across every system in the organisation.
A more practical model is an agent with a well-defined role.
For example:
- analyse a particular type of incoming request;
- retrieve information from approved systems;
- apply a defined set of policies;
- recommend or execute permitted actions; and
- escalate anything outside its operating boundaries.
The narrower and clearer the responsibility, the easier the system is to test, monitor and govern.
A Practical Path Forward
AI agents are likely to become an increasingly important part of business software as language models become more capable and integration frameworks continue to mature.
But successful adoption will depend less on creating agents that can theoretically do everything and more on identifying specific, controlled workflows where reasoning adds measurable value.
A sensible approach is often to begin with one well-defined process, establish the level of autonomy required and measure whether the agent genuinely improves speed, quality or operational effort.
If it does, the approach can then be expanded carefully.
The Future Is Probably Hybrid
For many organisations, the most effective architecture will not be AI instead of software.
It will be traditional applications, deterministic automation and AI working together — each handling the parts of the process they are best suited to handle.
Conventional code can provide reliability and control. Automation can execute predictable workflows. AI can help interpret ambiguity, work with unstructured information and support more flexible decision-making.
That combination is where agentic AI becomes considerably more interesting: not as a technology demonstration, but as another practical tool for building better business systems.
And that is probably a healthier objective than trying to give every spreadsheet its own AI agent.
