Air Canada learned an expensive lesson in February 2024. A grieving passenger had asked the airline’s website chatbot about bereavement fares, got the wrong answer, and booked on the strength of it. When the case reached a Canadian tribunal, Air Canada argued the chatbot was a separate legal entity, responsible for its own actions. The tribunal didn’t buy it.
Notice what that chatbot actually did, though. It talked. Nothing more.
Newer service AI goes further and changes things inside your systems. The NiCE agentic AI research hub describes systems that pull up an account, check billing policy, update the billing system and log everything in the CRM, with nobody on staff touching the case. So when one of these systems slips, the error turns up as a transaction someone has to unwind.
Below: what the term means, how it compares with chatbots and copilots, where it earns its keep, and what to build before you let it near a customer’s money.
So What Is Agentic AI, Exactly?
Until recently, most AI in customer service played a supporting role. It answered questions from a knowledge base. It summarized calls. It suggested replies that a human agent would read, tweak and send.
An agentic system starts from a goal instead. It figures out the steps, then carries them out using your company’s own software. Vendors tend to describe four abilities working together: perceiving what the customer wants, reasoning about it, acting, and learning from how things turned out.
Billing adjustments make a good test case. The AI looks up the account and checks the policy, then calculates the right amount. It pushes the change into the billing system, writes the case up in the CRM, and confirms with the customer. A person only gets pulled in when the request falls outside the rules.
Expect to see a lot more of this. Gartner predicts agentic AI will resolve 80% of common customer service issues on its own by 2029, and says operating costs could drop by roughly 30% as a result.
When the Agent Gets It Wrong
Picture a telecom (this one’s made up) whose AI agent handles plan downgrades and prorated refunds. One customer has two accounts tied to the same email, left over from a merger years ago. The agent grabs the dormant one. It applies the refund policy flawlessly, sends a friendly confirmation and marks the ticket resolved.
No rule was broken anywhere. Yet next month’s bill arrives with the full charge on it, and the customer calls back furious.
Meanwhile, the dashboard has already logged that case as a win for automation. A chatbot making the same slip would have typed one bad sentence. This agent moved money to the wrong account and wrote a false note into the CRM. It also nudged a performance metric in the wrong direction, all before anyone blinked.
Air Canada was held responsible for what its bot said. Would a tribunal be more forgiving about what a bot does? It’s hard to see why.
Chatbots, Copilots and Agents Side by Side
People mix these terms up constantly. Sorting them by who finishes the job, and by what a mistake looks like, cuts through most of the confusion.
| Tool | What it does | Who finishes the job | If it gets something wrong |
| Chatbot | Answers from a knowledge base | The customer or a human agent | The customer acts on bad information |
| Copilot | Drafts summaries and replies for staff | The human agent | Someone can catch it before it goes out |
| Generative AI | Produces text, images or code | Whoever is using it | A draft needs fixing |
| Agentic AI | Completes tasks across business systems | The AI, inside set rules | The error is already in billing, the CRM or the account |
They rarely work alone. A generative model might help an agent understand what the customer is asking, and when the agent hands a case off, a copilot can support the human who picks it up.
Buyers should keep the agent versus chatbot distinction front of mind. Gartner has called out “agent washing,” where vendors slap the agent label on old chatbots and assistants. By its estimate, only about 130 of the thousands of vendors claiming agentic AI actually offer it.
Start Small, Start Reversible
The NiCE hub frames adoption as four stages. First comes automation, which strips out repetitive manual tasks but sticks to rigid workflows. Augmentation adds copilots that help staff find answers, summarize and decide. With autonomy, AI runs multi-step workflows by itself. At orchestration, several AI agents coordinate across systems and departments.
Every step up widens the blast radius of a single bad lookup. By the orchestration stage, one wrong account match could spread into billing, logistics, and marketing before a human sees it. Most teams sensibly pick one workflow, prove it out and expand from there.
Which workflow? Look for routine, rule-bound requests that touch several systems. Account updates, subscription changes, refunds and billing adjustments, password resets and service activation all qualify. Then favor the ones you can undo. A botched password reset wastes a minute of the customer’s time. A botched refund wastes money and goodwill, plus hours of staff time fixing it.
Keep people on anything emotional, complex or sensitive. The Air Canada dispute itself began with someone arranging travel after a death in the family, about as clear a case for a human as you’ll find. Clearing routine tickets from the queue gives your agents more room for conversations like that one.
Patience has a payoff here. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, pointing to rising costs, fuzzy business value, and weak risk controls.
Four Things to Build Before Launch
Each of these blocks a particular way things go wrong.
Connected customer data. The AI needs the whole picture: calls, chats, emails, and past transactions. Patchy data produces confident mistakes, and the two-account refund above is exactly that kind of error.
Links into your business systems. Without access to the CRM, billing, order management, or identity tools, an agent can describe a problem and do nothing about it. Give it write access only where a given workflow truly needs it.
Rules, limits and handoffs. Settle ahead of time what the AI may do alone, such as refund caps and eligibility checks, and when it has to step aside. The NiCE hub names three common handoff triggers: confidence dropping below a set threshold, bumping into a policy boundary, and a sensitive customer situation.
An action log. Record every request, the data the AI used, what it did and how it turned out. You’ll need that trail to reverse a bad refund, answer a complaint and tune the system later.
Measuring Whether It Works
How many conversations the AI touched tells you very little. Outcomes tell you much more.
| Metric | What it reveals |
| Requests resolved with no human involved | How much routine work the AI really absorbs |
| Repeat contacts about the same issue | “Resolutions” that never fixed anything |
| First-time fix rate | Whether the right action happened on the first attempt |
| Cost per interaction | Real savings once rework is counted |
| Time to resolution | Speed as the customer experiences it |
Pay special attention to repeat contacts. NiCE runs a modeled example for a center handling 30 million contacts a year at $7 apiece. Push automated resolution from 20% to 35%, trim repeat contacts from 12% to 8%, and the model shows $56.7 million in yearly savings. NiCE frames that as an order-of-magnitude illustration rather than a forecast. Look closer, and $25.2 million of it, nearly half, rides on repeat contacts going down. A wrongly executed action pushes that number up.
Activity counts are also easy to game. Companies pushing staff to hit AI usage targets found that out the hard way during the AI sprawl crisis. Capture your baseline for the same request types before launch, so the before-and-after comparison is fair.
Your Team’s Job Description Is About to Shift
People stay in the picture. What they work on changes. Routine requests drain out of the queue, and human agents spend more of their day on tricky or delicate cases. New responsibilities show up as well, like supervising AI activity, reviewing exceptions, and refining the workflows the AI runs on. Good training and change management make that transition smoother.
One of those new duties sounds dull and may turn out to be crucial. Somebody has to read the action log and spot the refund that landed in the wrong account. Then they have to trace it back to the rule that caused it. Few service teams have anyone assigned to that today. Before you let an AI issue refunds by itself, decide who that person will be.
Related: Observe.AI Companion Agent Review 2026: Is It Worth the Cost?
