Answering versus doing
A conversational system that retrieves information is useful but limited. It shortens the conversation and leaves the outcome unchanged. An agent that can act — check real availability, create the booking, write the record, send the confirmation — changes the outcome, because the customer's problem is resolved in the conversation.
Acting requires integration and it requires constraint. The agent needs permission to perform a specific set of operations against specific systems, and it needs hard limits on everything else.
Escalation is a design feature, not a failure
The most important design decision is where the agent stops. Good escalation is explicit: defined intents that always reach a person, confidence thresholds, a customer who simply asks for a human, and anything touching money, health or a complaint.
When a conversation escalates, the person receiving it should get the full context rather than a cold start. Making the customer repeat themselves is worse than not automating at all.
- Route clinical, financial and complaint intents to people by default
- Always honour an explicit request for a human
- Pass the full transcript and any captured data on escalation
- Log every action the agent took, so it can be audited
Where the gains actually come from
The realistic gains are coverage and consistency rather than headcount. Contacts that arrive outside office hours get handled. The tenth caller during a lunch rush gets the same answer as the first. Qualifying questions get asked every time, so records arrive complete.
Consistency has a second-order benefit that is easy to miss: when every enquiry is captured in the same structured shape, reporting becomes possible for the first time.
What to be careful about
Agents are confident by nature, which is a risk when the underlying information is wrong or stale. Ground the agent in a single source of truth your team controls, and treat that source as part of the system.
Be careful with anything regulated. Allergen information, clinical guidance, financial eligibility and legal terms all need a verified source and a named human owner — not a model's best guess.
Key takeaways
- Judge an agent on whether it completes tasks, not on how well it chats
- Design escalation first; it determines whether the system is safe to deploy
- Expect gains in coverage and consistency rather than headcount
- Ground it in one controlled source of truth, especially for regulated information