Buy agentic AI customer service on how its mistakes get undone, not on how many conversations it closes without a person. Agentic AI customer service, meaning AI that takes actions inside your systems instead of only answering questions, changes what a failure is. Not a bad sentence. A completed transaction.
The category moved from answering to acting during 2026. Vendor documentation now describes agents that plan the steps and carry them out in the company's own tools: look up the order, change the address, issue the refund, write the result back to the CRM. CX Today calls it automation with authority. The liability question arrived before the capability did. In 2024 a Canadian tribunal held Air Canada responsible for what its chatbot told a customer about bereavement fares, and that chatbot could only talk.
We shipped AI with that kind of authority somewhere a wrong field delays a person's medication. For Samacare we built an LLM pipeline that reads scanned faxes, PDFs and handwritten notes, runs OCR where it has to, and extracts structured data for prior authorization and patient enrollment. The published results were "LLM-based automation of data extraction" and "Multi-provider patient enrollment automation". The extraction was not what took the time. The evaluation systems and the human review step were: the model finished the documents it was confident about, and people read the rest.
Confidence per case is the control, not accuracy in aggregate. A model that is right 97 times out of 100 tells you nothing about the document in front of you, because the three that are wrong do not arrive labeled. The decision that mattered on Samacare was which cases the system was allowed to complete on its own, drawn from what the system could say about its own certainty, one document at a time. An agentic support rollout faces that question with the nouns swapped: which actions may complete without a person.
Sort the actions by what it costs to reverse them. An answer can be corrected in the next message. An action persists, usually in a system of record and often in somebody else's: the refund is paid, the subscription is canceled, the address is changed. Reversing it is a second operation, performed by a person with more access than the agent had, and it starts after the customer has already noticed. Automating an action without staffing its reversal does not remove that work. It moves it to the worst moment to be doing it.
The part that ran longer than we expected on Samacare was keeping the review step real. Human-in-the-loop is one box on a diagram and a standing cost in an operation, and it is the first thing to look expendable once the automation number looks good. Every correction fed back to the model was a person's hour, not a model's. We could have shipped faster by accepting lower accuracy. Healthcare did not offer that option, which turned out to be a useful constraint.
If you are evaluating agentic AI customer service this quarter, write the list of actions before you watch a demo. For each one, record what it costs to undo, how long you have before it cannot be undone, and who performs the reversal in which working hours. Then set the threshold per action, not per deployment: a wrong order lookup and a wrong refund are not the same mistake. That is how we run call center delivery, with AI on triage, agent-assist and QA across every interaction and a person on the actions, through partner centers we select and manage, nearshore in Mexico and Colombia or offshore in the Philippines, India and Egypt. The middle of that cost order buys a reversal queue awake when your customers are, which is customer service nearshore. It is the same reason the handoff matters more than the deflection rate.
A wrong answer is a sentence. A wrong action is a record.