What actually separates an AI agent from normal software?
Traditional business software is deterministic. Someone wrote the rules in advance, and the same input produces the same output every time. If a lead scores above a threshold, it moves to a queue. If an invoice matches a purchase order, it posts. The behaviour is entirely contained in the rules, which is why it is predictable, auditable and brittle at the edges.
An AI agent is probabilistic and goal-directed. You give it an objective and a set of tools, and it decides the sequence of steps. Given "follow up with this lead appropriately", it reads the history, picks a tone, drafts the message and may choose to wait instead of sending. Nobody enumerated those branches. That is the capability, and it is also the reason the agent needs supervision that a rules engine does not.
Why does this distinction matter to a buyer?
Because it changes what "working correctly" means. Traditional software either matches its specification or has a bug, and you can test it exhaustively. An agent produces a distribution of outputs, most of which are good and some of which are not. You cannot test it exhaustively; you can only sample it and measure the rate.
That reframes procurement. The right question stops being "does it have the feature" and becomes "what is the quality rate on my work, and what happens on the bad cases". Buyers who carry a traditional software checklist into an agent evaluation usually end up with a product that demos beautifully and disappoints in month two.
What has not changed?
More than vendors like to admit. Agents still need the same unglamorous foundations as any other business system.
- Integrations. An agent with no access to your CRM, help desk or document store is a very expensive text box. If it can't reach your systems, it can't do the job.
- Identity and permissions. An agent acts as somebody. Deciding who, and what that identity may touch, is a normal access-control problem.
- Data quality. Wrong records still produce wrong answers, and an agent will state them more confidently than a report would.
- Change management. People still have to trust it, adopt it, and know what to do when it is wrong.
- Support and ownership. Something still breaks at an inconvenient time, and someone still has to own it.
The plumbing is unchanged. What sits on top of the plumbing has become far more capable and far less predictable at the same time.
Where do agents clearly beat rules-based automation?
Agents earn their place where the input is unstructured and the branching is impractical to enumerate. Reading a free-text email and deciding what it is about. Summarising a long thread for the person picking it up. Extracting terms from documents that no two counterparties format the same way. Drafting a reply that has to reflect context rather than a template.
The tell is simple: if you've ever tried to write the rules for a task and given up because there were too many exceptions, an agent is a reasonable candidate. If you wrote the rules and they work, an agent will mostly add cost and uncertainty.
Where is traditional software still the better answer?
Anywhere the requirement is exactness, repeatability or provability. Financial postings, tax calculations, compliance gates, anything where a regulator or an auditor expects the same input to yield the same output forever. Deterministic code is also cheaper to run and easier to reason about, which matters at high volume for a simple task.
The strongest systems combine the two rather than choosing. A rules engine enforces the constraints that must never be violated, and an agent handles the fuzzy interpretation in front of it. Let the agent decide what the email means; let the rules decide what may actually be committed.
How does the cost model differ?
Traditional software is largely a fixed cost. Once it is built or licensed, running it one more time costs almost nothing, so the incentive is to use it as much as possible. Agents carry a real marginal cost per unit of work, because every run consumes model inference. That inverts a habit many teams have: with agents, running the process on everything by default is a decision with a bill attached.
There is also a hidden cost that rarely appears in a business case — review time. Output that is nearly right still needs a human to read it. When you compare an agent against the status quo, count the review and correction minutes, not just the generation. Agents that save time on paper and lose it in review are the most common disappointment in this category.
What changes about security?
Two things, and both are genuinely new. First, an agent can be steered by the content it reads. Text arriving from outside — an email, a web page, an uploaded document — can contain instructions, and an agent that treats data as instruction can be manipulated into acting against you. Traditional software does not have this failure mode.
Second, agents concentrate access. To be useful, an agent often needs read access across several systems at once, which makes its credentials more valuable than any individual user account. The mitigations are conventional but non-negotiable: least-privilege scopes, approval steps on anything irreversible, complete action logs, and a tested way to revoke access quickly.
How should you decide between the two?
- Describe the task as a rule. If you can write it in a page and it holds, buy or build deterministic software.
- Count the exceptions. Many exceptions, or exceptions that keep appearing, point towards an agent.
- Check the cost of a wrong answer. High-consequence and irreversible steps stay deterministic, or stay behind human approval.
- Check the volume. Very high volume with a simple rule favours code; moderate volume with messy input favours an agent.
- Assume a hybrid. In practice the answer is usually an agent for interpretation with rules for enforcement.
Frequently asked questions
No. A chatbot converses; an agent acts. Agents are given tools and permissions to change things in your systems, which is what makes them valuable and what makes their governance a different problem.
Usually not. Agents tend to sit on top of existing systems and act through them. Your CRM, help desk and finance tools remain the record of truth; the agent is a new way of working with them.
They are less predictable, which is not the same thing. Deterministic software fails the same way every time; an agent produces a range of outcomes. That is why agents need sampling, measurement and human approval on high-consequence actions.
Most business agents do not retrain a model. They are configured with your context, connected to your systems and given examples or instructions. Ask specifically whether your data is used to train shared models and how to opt out.
Per unit of work, usually yes, because every run consumes inference. Whether it is more expensive overall depends on the labour it displaces and how much review time the output requires.
Yes, and that is normally the best design. Let the agent interpret unstructured input and propose an action, then let deterministic rules validate and commit it.
Compare AI agents by category to see which tasks are well served today.