Understand

AI agents for customer support

In customer service, what wears the team down is not how hard the answers are: it is their repetition and their timing. The same question comes back fifteen times a week, and the one that really matters arrives on a Friday at 7pm. An AI agent absorbs the repetition and watches the clock — it does not replace judgement, it gives back time to those who exercise it.

1. What an agent takes on

Four tasks take up most of a customer service's time, and delegate well because they follow a statable rule and can be checked at a glance.

Sorting what comes in — reading every incoming message, extracting the actual request, attaching it to the right customer and ranking it by urgency. That sorting is what is missing when everything lands in a shared inbox.
Answering routine requests — the ones whose answer already exists: a delivery time, a procedure, a password to reset. They make up the bulk of the volume and call for no arbitration.
Preparing the rest — gathering the customer's history, previous exchanges and what remains open, then presenting the case to whoever will answer, with enough to decide in two minutes.
Watching the deadlines — spotting what is approaching the deadline you set yourself, and flagging it before it is missed rather than after.

What these four tasks have in common: none requires judging a particular case, only consistency. That is exactly what a team does badly on a busy day.

2. What you must not let it handle

Three things stay on the human side, and not for technical reasons.

The commercial gesture. A refund, a credit note, an exception to the returns policy: these are commitments that cost money and set a precedent. They must go through a validation, not because an agent would be incapable, but because their cost is real.

The angry customer. An escalating complaint is defused by acknowledging the wrong, not by the accuracy of the answer. An agent that answers correctly but beside the point makes things worse — and that is what ends up in a public review.

Anything someone's safety depends on. A health question, a compliance matter, a risky use of a product: the criterion is not difficulty, it is what a mistake would cost. These requests can be recognised in advance and routed to a human by default.

3. First response time

It is the best first project, because it meets the four criteria of good delegation: it comes back constantly, its rule fits in one sentence, its result is visible within days, and a mistake is fixed with a word.

What customers most often resent is not the answer itself, it is the silence that precedes it. An acknowledgement stating what has been understood and when the answer will come beats a perfect answer three days later — and that is precisely what an agent produces without tiring, nights and weekends included.

In practice, the agent handles routine matters alone and raises the rest with its context. You then move the boundary once the quality suits you — that is the autonomy boundary, and it is set per request type, never once and for all.

4. Connecting your existing customer service

The question that stalls most projects is not "what can the agent do" but "do we have to change tools". The answer should be no.

A customer service rarely lives in one place: a contact address, a form on the site, a chat, sometimes a ticketing tool already in use. What matters is that requests arrive in the same place and leave through the channel they came from — the customer should never have to know whether a human or an agent is answering, nor switch channel midway.

In nullbot, every source connects through a single entry point, and replies are relayed back to the original tool. Common systems are recognised as they are; the others connect by pointing to where the sender, subject and message can be read. You keep your tool, you add agents inside it.

5. Measuring the result

Four figures are enough, and they must be read together.

First response time — the time between a request arriving and the first real answer. That is where the effect shows first, and it is what customers feel.
The share handled without human intervention — useful only when read alongside the next one: a high share obtained at the cost of unhappy customers is not a gain.
Satisfaction — asked after resolution, on requests handled by the agent as well as the others. If the gap widens, the autonomy boundary is badly placed.
Human rework time — how many minutes a day the team spends correcting the agent. If it does not fall, the rule is badly worded: that is what needs revisiting, not the agent.

A customer service that answers fast but demands systematic proofreading automates nothing: it shifts the load from whoever answers to whoever reviews.

6. Frequently asked questions

Will the customer know they are talking to an agent?

That depends on what you decide. In the most common setup, the agent prepares and a human sends: the customer receives a message from their usual contact. If you let the agent answer directly, honesty requires saying so — and several regulations now make it mandatory.

Can I take over a conversation in progress?

Yes, in one gesture: the conversation moves to the human side and the agent stops replying to it. That is essential — a system you cannot take back is a system you will not trust. The reverse move exists too, once the tension has passed.

Should I pay per resolution, as some vendors charge?

Not at nullbot: customer service is included in the subscription, whatever the number of requests handled. Per-resolution billing has a design flaw — it charges you more when your customers have more problems, which is exactly when your cash flow is already suffering.

Where should I start?

With sorting and acknowledgements, not with answers. Within a week you will see what the agent understands of your requests, without a single customer receiving an answer you have not reviewed. Routine answers open up afterwards, one category at a time.