Chatbot or agent assist: where should customer service AI start?

The same two proposals often reach a customer service leader together. One is a chatbot: an assistant on the website or in WhatsApp that answers customers directly. The other is agent assist: AI inside the agents’ own tools that suggests a reply with its sources, summarizes a long thread and proposes the right queue. Both run on similar models. They fail in very different places.

Start with agent assist. When AI helps agents, every mistake meets a trained person before it meets a customer, and each draft they accept, edit or reject shows which requests are predictable. Once that record points to a narrow request that agents almost never change, open self-service for that request alone, with a person one step away.

What is the difference between a chatbot and agent assist?

A chatbot talks to the customer. Agent assist talks to the agent. That one difference decides who sees a wrong answer first, how quickly you learn and what a failure costs.

Agent assist is not a timid chatbot. It is where you build what any chatbot needs anyway: current sources, a clear list of request types, and evidence about which answers are safe to send unread.

Why start with agents rather than customers?

Agents are expert reviewers already inside the workflow: they notice at once when a draft quotes an outdated return window or promises a delivery date the carrier never gave. Starting with them has four advantages:

• Mistakes stay internal. A wrong draft costs an agent seconds; sent by a chatbot, it costs a repeat contact or a complaint.
• Sources get fixed. Suggested replies expose outdated knowledge articles quickly, because agents reject the drafts built on them.
• You collect evidence. Every accept, edit and reject, with a short reason, is a labeled example. Grouped by request type, the record shows where the AI is reliable.
• Trust builds. Agents who shaped the drafts trust them, and customers later see answers agents already approved.

Every draft an agent accepts, edits or rejects is evidence about what is safe to automate.

Which requests are safe to automate for customers?

A request is a candidate for self-service when all five of these hold:

1. The answer comes from data, not judgment. It is read from a record, such as an order, a shipment or opening hours, not decided about the customer.
2. It is narrow and frequent. One clear intent, arriving in volume, so the effort pays back.
3. The customer can be identified simply. An order number plus the email or phone on the order shows the right record to the right person.
4. A wrong answer is cheap and easy to correct. No money moves, and nothing is promised that the company can’t keep.
5. Agents rarely change the draft. Measured over many cases, not remembered from a good week.

Order-status questions are a common first candidate. “Where is my order?” arrives constantly, the answer lives in the order system and the carrier’s tracking data, and the right reply rarely depends on judgment. Watch the edges, though. A parcel marked delivered that the customer says never arrived is no longer a status question. It is a claim, and it belongs with a person, as do refunds, complaints, damaged goods and exceptions to policy, whatever the acceptance numbers say.

What does the move from assist to self-service look like?

Picture an illustrative case: an online retailer whose support team answers many delivery-status questions by email and chat.

1. Suggested replies. For each delivery question, the agent sees a draft quoting the order record and the carrier’s latest tracking event, then sends, edits or rejects it, with a reason for any edit.
2. Agreement by sub-type. The team splits the results into “in transit, on time”, “delayed” and “delivered but not received”. Agents rarely touch the first kind of draft; they often rewrite the other two.
3. Shadow self-service. For on-time shipments, the system prepares a customer-facing answer next to the agent’s reply, without sending it, to show whether it holds up on its own.
4. Self-service order tracking. Once agreement for on-time shipments stays above a threshold the owner agreed in advance, customers can check that status themselves in chat. Delays and missing parcels still go to agents, with drafts.
5. Sampling continues. People review a sample of automated answers every week, and the step is reversed if quality slips, for instance when a carrier changes its status codes.

Each step is justified by the record of the one before, which is the principle behind our note on when AI agents should act alone.

How do you protect the customer experience?

Self-service is only good service if leaving it is easy.

• A person is always one step away. A visible way out at every point, not after three failed attempts, recognized when customers ask for it in their own words, Turkish or English.
• No loops. If the assistant can’t resolve a request after one clarifying question, it hands over. It never asks the same question twice.
• Context travels. The handover carries the order, what the customer asked and what was already answered, so nobody repeats themselves. Our note on handing a customer from bot to person covers the design.
• Answers stay inside the data. If the record doesn’t answer the question, the assistant says so and hands over. It never guesses a delivery date.
• Honesty. Customers know they are talking to an automated assistant, and when a person will reply if nobody is free.

What should you measure at each stage?

Measure what each stage must prove, against the baseline recorded before launch.

• Agent assist: acceptance and edit rates by request type, reasons for rejections, handle time, quality scores and knowledge gaps reported.
• Shadow self-service: agreement between the prepared answer and what the agent sent, and every case the system would have answered but shouldn’t have.
• Live self-service: requests resolved without a repeat contact through any channel, how often and why customers ask for a person, and satisfaction compared with agent-served requests.

Handle time alone will mislead you at every stage; the note on customer service AI metrics explains which numbers catch what.

Start with agents

Pick the request agents answer most often from data they look up, and give a small group of agents suggested replies with sources. The customer operations page describes how that first project runs, and commerce intelligence covers the same pattern for retail and marketplace questions. When the record supports self-service, that step is custom AI software built around your systems and limits.