customer service ai agents escalation soporte

The solution is an AI customer service agent capable of answering common queries and passing difficult cases to a person: it responds using your company's information (documentation, policies, orders) citing the source, and hands over to a person when it doesn't have an answer, when the customer requests it, in the event of a complaint, or when the action exceeds its authority. Companies Automation kicks off its AI Factory with this task.

Most inquiries a company receives are the same old ones: where is my order, what are the hours, how do I return this, what does the plan include. Answering them doesn't require judgment, but it consumes hours of people who do have it. The rest (a complaint, an exception, a decision involving money) needs a human. A good system separates the two piles without the customer noticing the seam, and it does so with explicit escalation rules, not with the hope that the model will "realize" it.

What types of solutions exist?

There are three families, and they differ in how much they know about your company and what they can do. All three can escalate to a person; what changes is how much context they provide and how many queries they resolve before doing so.

1. Scripted Chatbot. A tree of options with fixed responses. It resolves what someone has foreseen and written; everything else ends in "let me transfer you to an agent." Cheap and predictable, but it doesn't understand a question phrased differently. We explain the difference from an agent in chatbot or AI agent: the differences.
2. AI Support SaaS. Platforms like Intercom or Zendesk incorporate an assistant that responds based on your help center and transfers the conversation to a human agent within the same tool. It works well if your support already lives there; its limit is that it knows what's in the help center and little about your orders or contracts.
3. Custom Agent connected to your systems. It responds with your documentation and also checks order status, invoices, or customer contracts, and can execute limited actions. How it is given that knowledge is covered in AI agents with knowledge of your company.

What does it solve alone, what does it escalate, and what should it never do?

The boundary is written before turning on the system, not after the first complaint. This table serves as a starting point; each business adjusts the rows.

Resolves alone	Escalates to a person	Must never do
Questions with documented answers: hours, deadlines, policies, features	Questions with no answer in the documentation or with contradictory sources	Invent an answer when it cannot find the source
Status of an order, shipment, or invoice by checking the system	Orders with issues: delay, breakage, shipping error	Promise deadlines or compensation not in the policy
Explain how to make a return or exchange	Authorize a refund or a policy exception	Move money without approval
Collect data to open an incident	Complaints, angry tone, threats of cancellation	Argue with an upset customer
Inform about contract conditions or plans	Contractual changes, cancellations, renegotiations	Modify contracts or prices
Respond in the customer's language	When the customer asks to speak with a person	Refuse to transfer to a person or pretend to be one

When should it pass the conversation to a person?

Well-executed escalation responds to explicit rules that the system checks at every turn. These should be in any deployment:

• It doesn't find the answer in the documentation or systems, or confidence is low. It says "I don't know, let me transfer you to someone" rather than improvising.
• The customer asks for it, using any phrasing. It is fulfilled on the first request, without pushing back.
• There is emotion or a complaint: anger, frustration, mention of a formal claim, threat of cancellation, or a negative review.
• The action exceeds its authority: refunds, discounts, contract changes, policy exceptions, third-party personal data.
• The conversation goes in circles: the same question twice without resolution.
• Sensitive customer or topic: key accounts, legal issues, health, minors.

The contextual handoff: what distinguishes good escalation

Escalating isn't hanging up and having the customer start over. The person receiving the conversation should see, before replying, a two or three-line summary of who the customer is, what they are asking for, what they have already been told, what sources the agent consulted, and which rule triggered the escalation. With that, they can answer in the first message instead of asking the customer to repeat everything.

The handoff also works in reverse: when the human resolves the case, their response is recorded and, if the question is likely to be repeated, it is added to the documentation so the agent can resolve it next time. This way, the percentage of cases resolved without a person grows with use, not just with more technology.

On which channels does it work?

The same agent serves several channels with the same rules; each channel changes the pace, not the criteria.

• Web. On-page chat: short conversations, many visitors who aren't customers yet.
• Email. Longer and less urgent messages; the agent can prepare a draft for a person to review instead of replying live.
• WhatsApp. Where the customer expects a quick response and where it hurts most not to have one outside of business hours. We cover this in AI agents on WhatsApp Business.

How is success measured?

Four figures, taken before starting and every month after. Without a baseline, you won't know if the agent is helping or just changing where inquiries pile up.

1. Percentage resolved without a person, counted only when the customer doesn't write back about the same issue. A closed case that reopens the next day doesn't count.
2. Response time, for both the agent and the person after escalation.
3. Satisfaction, with a question at the end of the conversation, separating those handled by the agent from those escalated.
4. Escalation false positives and false negatives: how many conversations it escalated unnecessarily (the human closes it with the answer already in the documentation) and how many it should have escalated but didn't (the customer complained later).

The cost depends on the solution family and how many systems need to be connected; the ranges are in how much a chatbot for a company costs, and the complete vision of the area in automating customer service with AI agents.

How Companies Automation does it

"Answering the same questions using your company's information and passing difficult cases to a person" is one of the tasks with which Companies Automation starts its AI Factory. It is the first level of its journey (one task, then an area like customer service, then the whole company) and follows its motto: AI produces, your team decides. The agent works with minimum permissions, every action is traced, and your company's knowledge is not used to train third-party models. Published timelines: free first process evaluation, 2-4 week operational diagnosis, and the first agent operating in 4-8 weeks, measured against a baseline from the first month.

Frequently Asked Questions

How does the agent know it doesn't know the answer?

Because it only responds with information it finds in your sources and cites it. If the search returns nothing relevant, if there are contradictory sources, or if confidence is low, the rule dictates escalation. An agent that cannot say "I don't know" is an agent that invents, and that is worse than having none at all.

What if the customer gets angry with the agent?

It escalates as soon as tone is detected: anger, frustration, mention of a claim or cancellation. The agent does not argue or try to calm them down; it says it is transferring them to a person and hands that person the summary of the conversation. Arguing with an upset customer is human work by definition.

Can the agent issue refunds or contract changes?

It shouldn't. It can explain the policy, collect the data, and prepare the request, but authorization is given by a person. It's a limit of permissions, not capability: even if the customer asks, the agent doesn't have the authority to move money or modify a contract, and that's how it's written in its configuration.

What do I need to have before starting?

Your documentation and policies in a searchable place, read access to the systems where orders or contracts are stored if you want it to consult them, and written escalation rules: what it resolves alone, what it passes to a person, and what it never does. With that, the first agent is set up on one channel and measured.

If you want to know how many hours are spent today answering the same things, the free diagnosis from Companies Automation consists of 9 questions, 2 minutes, and a figure in euros per year. And if you prefer to talk about it, the artificial intelligence consultancy begins with a free evaluation of the first process.