A service manager we spoke with had a blunt reaction to the phrase "AI agent": "The last thing I need is a robot deciding to close a customer's ticket without checking if they are actually happy." She was right to be cautious, and also right that AI agents have a role in her ticketing process — just not the role she was picturing. The useful question is never "should we use AI agents," it is "which specific steps in this specific workflow are suited to an agent, and which are not."
What makes a step suitable for an agent
Steps suited to AI agents share three characteristics: they happen often, they follow a definable pattern most of the time, and success can be checked. Logging a ticket, categorising it by type, checking whether it matches a known issue, drafting a first response, and routing it to the right team — all of this happens hundreds of times a week, follows a recognisable pattern, and can be verified against clear criteria. This is squarely the territory of [AI workflow agents](/solutions/ai-workflow-agents).
A simple filter for any given step
- 01Volume: does this happen often enough that automating it saves meaningful time?
- 02Pattern: does this step follow a recognisable structure in the majority of cases?
- 03Verifiability: can a human quickly confirm whether the agent got it right?
- 04Reversibility: if the agent gets it wrong, is the consequence easy to correct?
A step that scores well on all four is a strong candidate. A step that fails on reversibility — closing a customer account, approving a large payment, issuing a refund above a threshold — should keep a human in the loop regardless of how well it scores on the others.
Where agents do not belong
Judgment calls that involve relationships, exceptions, or reputational risk are poor fits for full automation, not because AI cannot process the information, but because the cost of a wrong call is asymmetric. A collections agent deciding whether to escalate a long-standing, generally reliable customer who has missed one payment is a judgment call that benefits from a human who knows the relationship — even if an AI system flags the pattern first. This is why [Finance & collections intelligence](/solutions/finance-collections-intelligence) is built to surface the pattern and draft the options, while leaving the actual escalation decision with a person when the account has any history worth weighing.
The same logic applies to service escalations involving an unhappy customer, procurement decisions involving a new, unvetted vendor, and any approval where the amount or relationship crosses a threshold the business has defined as requiring judgment. Automating these fully does not save meaningful time — it just moves the risk of a bad decision from a person who can be held accountable to a system that cannot.
The mixed model, in practice
Most real workflows are not purely automatable or purely manual — they have a routine path and an exception path, and the value of AI agents comes from handling the routine path completely while routing the exception path to a person with the right context already assembled. In procurement, this means the agent can process a standard, in-policy purchase order from request to approval without a human touching it, while an out-of-policy or unusually large request gets flagged with full context for a manager to decide — the model behind [Procurement & vendor management](/solutions/procurement-vendor-management).
In project delivery, an agent can track site progress updates against a schedule and flag drift automatically, while a decision to reallocate budget or extend a deadline stays with the project manager, supported by [Project & workflow management](/solutions/project-workflow-management) surfacing the relevant history rather than making the call itself.
Building trust gradually
Teams that succeed with AI agents rarely start by giving them full autonomy over a process. They start by having the agent draft and recommend, with a human approving each action, and expand the agent's authority only as its track record earns it — a pattern we describe in more detail in [a practical framework for an AI process audit](/insights/ai-process-audit-framework). This is not caution for its own sake; it is how any new team member earns responsibility, and an agent is no different in that respect.
The right question going forward
The debate about AI agents in business workflows often gets framed as a binary — full automation versus no automation. Neither extreme reflects how well-run operations actually work. The more useful framing is a map of the workflow, step by step, with each step honestly assessed against volume, pattern, verifiability, and reversibility. Some steps will clearly belong to an agent. Some will clearly belong to a person. The value is in being deliberate about the line, rather than letting a vendor's marketing draw it for you.