AI customer service
AI customer service is two very different jobs wearing one name. Answering a question someone could have found on your site is low risk and pays for itself. Promising a refund, a delivery date or an exception is a commitment your business has to honour, and no amount of fluent writing makes that safe. The build is mostly about keeping those two apart.
A fluent wrong answer costs more than a slow right one — a promised refund or delivery date is a bill you now owe. The safe half and the dangerous half wear one name, and the whole build is keeping them apart.
It answers what is already true, and is stopped from promising what isn't
Tell me what your support gets asked
Two lines is plenty. What comes in most, and the things it must never promise. I read it myself and answer within one working day.
Sergey Avakyan. You are writing to me, not to an inbox — I am the whole company, and I am the one who builds it. More about me →
Questions with an answer that already exists get answered in your voice; anything that would commit you — a refund, a date, an exception — is stopped and handed to a person.
Projects go in the order they arrive — the earlier we start, the earlier yours is built.
- Answers only from what is already true — your policy, your order record, your stock — never invented.
- A written line between answering and committing, enforced in the code: a refund, a delivery date, an exception it may draft but never send alone.
- A second, independent model that signs off every reply before a customer sees it.
- The numbers that actually tell you it works — repeat contacts, escalation quality, blocked replies — not the deflection rate every vendor shows.
The exact thing this page describes is already running: routine buyer questions answered in the seller's own voice, every reply signed off by a second, independent model before it is sent — one of the thirteen systems of mine live daily, out of sixty-five built.
A quarter of whatever we agree is due at the start, and that quarter is the only money at risk under the two-week promise — nothing else is due until it is running. What puts you at the bottom of that band: answers that already exist somewhere true — a policy page, an order record — and a short, clear list of what it must never promise. What puts you at the top: several places the truth lives, a customer system it has to read from, and a support history messy enough that deciding what is authoritative is half the job. It sits in the middle band because writing a fluent reply is the easy half — the work is keeping answering and committing apart, and proving it stays apart.
- Two weeks, or the first stage is free. The first stage is one named piece, written down before we start. If it isn't running two weeks after the start, you pay nothing for it. The clock pauses on any day I'm waiting on you for an access or an answer.
- You write the pass mark before I write the code. Before anything is built we agree one sentence that says what working means — for example, "it takes the call, puts the booking in the calendar and never promises a price, on nine of every ten of fifty real calls". If it misses that, you choose: I finish it at my own cost, or you take that stage's money back and the work stays with me.
- Three months of fixes included. For three months after handover I fix whatever breaks in what I built, already paid for. Something you ask for later, or a service someone else changes underneath it, is new work — and I'll tell you which it is before I touch it.
“I” here is the whole company: AVAROPA SLU, one founder and one employee, both of them me. The guarantee is the company's, and I am the person who honours it — you deal with the same one throughout.
Every stage you have paid for is yours the moment it is paid — the code and the accounts, handed over as we go. Nothing here depends on you staying.
Prices exclude VAT — 21% in Spain, reverse charge for an EU business with a valid VAT number, and normally none outside the EU — confirmed on your invoice. The band is a starting point; the exact figure is agreed and fixed before any work begins, and it does not move after that. The two-week first piece is the part I guarantee; the finish windows are what a job this size usually takes.
How it works, what it refuses, and when you don't need it
What automated customer service is sold as, and what it is
- Sold as: it handles your support. It handles the repetitive part of your support, which is most of the volume and almost none of the difficulty. The hard tickets stay hard, and now they arrive with an annoyed customer attached who has already talked to a machine.
- Sold as: it learns from your history. It answers from whatever you point it at. If your help pages are three years out of date, you have automated the distribution of stale answers, and made them sound confident.
- Sold as: it never sleeps. True, and it matters. But a reply at 3am that invents a delivery date costs more than the ticket was worth. Speed without a limit is not a feature.
- Rarely mentioned: it will be confidently wrong. Not often. Often enough that the interesting question is what happens on that occasion, and whether you find out. That is the part I would build first.
The line: answering versus committing
The rule I use is blunt. If the answer already exists somewhere true — your policy, your order record, your stock — the AI may say it. If the answer creates an obligation, a person signs it. Nothing in between.
This is not a philosophical position, it is how the reply gets checked. In the customer-chat assistant on this site, one model writes the reply and a second, independent one reads it before a buyer sees it and blocks anything that promises what it should not. In its own evaluation four separate AI reviewers checked that safety layer and found zero failures.
- May answer: where is my order, what is your returns window, does this fit that, how do I change an address.
- May draft for a person: a goodwill gesture, a partial refund, an exception to policy.
- Never alone: anything that costs money, anything about a legal or medical consequence, anything where the customer is already angry. Anger is a routing rule, not a tone problem.
How you find out it is actually working
Deflection rate is the number every vendor shows and the number that hides the most. A ticket closed by a customer giving up is deflected too.
- Repeat contact within 48 hours. The honest version of deflection. If a third of the answered customers come back, the machine did not answer them, it delayed them.
- Escalation quality, not escalation rate. A high hand-off rate is fine. A hand-off with no context attached is not — the person now does the work twice.
- The blocked replies. Read what the safety layer stopped, weekly. It is the cheapest source of truth about where the thing is overconfident, and it is the file nobody opens.
- Cost per resolved conversation. Model calls plus the person hours it did not remove. Compare it with the hours it did. Some support teams find the answer is no, and that is a valid outcome of a free hour of my time.
Do not buy this if
- Your volume is a few conversations a day. A saved-replies file and a shared inbox will beat any AI customer service build on cost and on quality.
- Your policies genuinely vary by customer and nobody can write down the rule. The machine will pick one and apply it to everyone.
- Your product is the conversation — consulting, therapy, anything where the human is what was bought.
- Nobody will read the escalations the same day. Then all you have built is a politer waiting room.
Answering, already in production
- Customer-chat assistant — this exact argument, running: routine questions answered in the seller's voice, with a second independent model signing off every reply before it reaches a buyer.
- Returns & defects handler — the ticket type that costs money — every claim judged against your written rules the moment it lands, honest ones through fast, the rest queued for a signature.
- Team briefing bot — the unglamorous half nobody demos: the same summary of what came in, landing in the team chat every morning whether or not anyone is free to write it.
Honest answers
How much does AI customer service cost?
€4,800 to €8,300 as a one-off build, higher when the truth lives in several places and it has to read from your customer system. A quarter is due at the start and is the only money at risk under the two-week promise; nothing else is due until it runs. The first hour is free.
Is AI customer service worth it?
It pays at volume, on questions whose answer already exists somewhere true. If you handle a few conversations a day, it is not worth it — a saved-replies file and a shared inbox beats any build on cost and on quality.
What can AI customer service safely handle on its own?
Anything where the true answer already exists: order status, delivery windows, returns policy, sizing, how to change an address. The test is whether answering creates a new obligation. If it does — a refund, an exception, a promise about a date — a person signs it.
How do you stop it promising something we can't honour?
A second, independent model reads every reply before it goes out and blocks the ones that commit you to something. It is a separate check on purpose: asking the same model to grade its own answer mostly produces agreement. In our own evaluation of that layer, four separate AI reviewers found zero failures.
Will customers know they are talking to an AI?
Yes, and they should be told. It costs nothing in satisfaction when the answers are good, and hiding it costs everything the first time someone works it out. It also sets the right expectation about what the thing can decide.
What about angry customers?
Anger is a routing rule. A conversation with real frustration in it goes to a person with the history attached, and it does not go back to the machine afterwards. Trying to de-escalate automatically is the single most expensive mistake in this category.
Is this the same as an AI receptionist?
No. A receptionist handles the inbound contact — the call, the booking, the who-are-you. Customer service handles the case after it exists, with your order data and your policies behind it. They are often built together, but they fail in different ways and are worth scoping separately.
Send me last week's tickets
Fifty real conversations, including the ones that went badly. I will mark which the machine could answer, which it could only draft, and which must never reach it. You see that split in the first hour, which is free, and before you have spent anything at all.
Book that hour →