AI agents for business: where they genuinely work and where they get in the way

AI for business is sold as a replacement for staff, but what it replaces is not people — it is specific operations. Here is what an AI agent handles reliably, where its mistakes are expensive, and what you need in place beforehand to avoid ending up with a polite chatterbox.

Conversations about AI agents for business usually open with "who will it replace". That is the wrong framing: an agent replaces not a person but individual operations inside that person's job. Where the operation is clearly defined and its result can be checked, AI works reliably. Where a decision carries accountability, it stays an assistant.

Below is that split: what you can hand over calmly, what needs supervision, and what it will not do however much you tune it.

Where AI agents for business work reliably

These tasks share one property: the result can be verified. You can look at the agent's answer and call it right or wrong without unpicking how it got there.

The first reply to a customer, around the clock: clarify the request, answer a standard question, capture the contact details and hand over to a person. Speed matters more here than perfect wording.
Processing incoming documents: invoices, delivery notes, free-form requests. The agent extracts the fields and the system checks them against your records.
Finding the right product from a description or a photo, when the catalogue is large enough that people get lost in it.
Drafts: a reply to a review, a letter to a customer, a product description. Someone edits and sends — faster than writing from scratch.
Listening to sales calls and scoring them against criteria set in advance. The agent does not get tired and covers all of the recordings rather than a sample.

Notice that in every case there is either a checking system or a person at the end. That is not distrust of the technology but ordinary engineering insurance.

Where its mistakes get expensive

The danger is not that AI makes mistakes — people do too. The danger is that it makes them confidently, in exactly the same tone it uses when it is right. The person reading the answer cannot tell the difference.

Firm commitments: deadlines, prices, availability, warranty terms. The agent must take these from the system rather than phrase them itself.
Disputes: returns, complaints, conflicts. These need a person entitled to make the decision and answer for it.
Anything with legal weight. A sentence that reads smoothly can be costly.
Any answer the customer will read as a promise from the company.

A simple dividing line: if the agent's answer implies a commitment by the company, it must rest on data from the system or pass through a person. Everything else can be automated.

What you need in place first

An AI agent is not the first step of automation but a layer on top of data that already works. If the data is missing or contradictory, the agent will start inventing things politely.

A current source of truth: catalogue, prices, stock, order statuses. The agent answers from it, not from memory.
A destination for the result: a CRM, a table, a task. Otherwise conversations stay conversations.
Written rules: what to answer to common questions, when to hand over to a person, what not to promise. These are precisely the instructions you would give a new hire.
A person who reads what the agent has been answering all week — at least for the first few months.

If the first two are missing, it makes more sense to sort out the data first. That usually turns out to be the real project, with AI added afterwards, when it can actually deliver something.

How to tell whether the agent is doing well

Judge it by numbers that existed before the rollout, not by the impression its dialogues make.

Time to first reply — the most honest metric, and it moves immediately.
Share of enquiries closed without a person, and separately the share handed over. The second figure should not be zero: an agent that never calls for help is probably answering at random.
How many answers had to be corrected. Rising numbers mean the rules are vague or the data has gone stale.
Whether it reached sales. If enquiries went up and revenue did not, the issue is not the agent but what happens after it.

How to check your own setup

Three questions worth asking before paying for a rollout.

Do you have one place where product, price and order data is known to be correct? If there are two and they disagree, the agent will confidently tell customers the wrong thing.
What do customers ask most often, and are there agreed answers? If every rep answers differently, the agent has nothing to stand on.
Who will read the agent's conversations and adjust the rules? Without that person the system degrades within a couple of months.

An AI agent takes load off well where the work is repetitive and checkable. It does not solve the absence of a process — it makes an existing process faster. If there is no process yet, there is nothing to automate.