Automating customer service with AI (without dehumanizing it)

In short

• Yes, you can automate your customer service with AI: a well-built assistant handles on its own on the order of 30 to 60% of repetitive requests (tier 1), not 100%, and that's exactly how it should be.
• Escalating to a human is not a failure: it's a core feature. The assistant hands off with full context, and the customer repeats nothing.
• Reliability condition: the assistant is bounded to YOUR information (RAG), cites its sources, and says "I don't know" rather than making things up.
• Order of magnitude of the gain for an SME: 40 to 60 hours of tier 1 per month, that is 1,200 to 1,800 € at a loaded hourly rate of 30 €.
• You steer with two indicators: the share of sourced answers and customer satisfaction (CSAT), not just the deflection rate.

The problem, concretely

In most SME support desks, a large half of tickets look alike: "where is my order?", "how do I make a return?", "what are your lead times?". Answers that already exist, in the FAQ, the terms of sale, or the order management tool. Yet each one ties up 5 to 10 minutes of a human's time, and the customer sometimes waits hours for trivial information.

Facing this, the number one fear is legitimate: "AI is going to dehumanize our support". Everyone has already raged at a bot that loops endlessly and refuses to pass you to a human. But the real issue is not AI or human, it's the dosage: letting AI absorb the repetitive work so your teams spend their time where the relationship truly matters. Badly dosed, AI frustrates and drives people away. Well dosed, it improves both sides: the customer gets their answer in 30 seconds at 3 a.m., and your team only handles the cases that deserve a human brain.

Before / after: a ticket's journey

Before, the typical journey of a simple request looks like this:

• 1. The customer sends an email or a message: "where is my order no. 4812?".
• 2. The ticket enters the queue, behind 30 others. Wait: 2 h to 24 h.
• 3. An agent opens the ticket, looks up the order in the back office, writes the reply: 5 to 8 minutes.
• 4. One time in five, the customer asks a follow-up question, and it starts over.

After, with a properly bounded AI assistant:

• 1. The customer asks their question in the chat or by email, at any hour.
• 2. The assistant identifies the type of request. For an order status, it queries your back office; for a procedural question, it answers only from your documentation (FAQ, terms of sale, returns policy), citing its source.
• 3. If the information does not exist in its sources, or if the customer expresses annoyance, the assistant escalates to a human by passing along the entire history: the customer repeats nothing.
• 4. Your team now only sees the tickets that deserve it, with a summary already prepared.

The technical point that changes everything: the assistant is not a free-running ChatGPT, it's a RAG bounded to your content, with a "no source, no answer" clause (I detailed the mechanisms in how to stop AI from making things up). A chatbot that announces a false delivery time or an invented returns policy destroys more trust than it creates. This is exactly the setup running on BeForBuild.com, my B2B SaaS: the assistant answers from the business documentation, with sources to back it up, and hands off when it doesn't know.

Mini case: Casabello, a home decor e-commerce

A fictional but realistic case, assembled from situations I encounter. Casabello, an e-commerce store for decor and small furniture, 18 employees, around 1,100 tickets per month, two people on support. Analyzing a month of tickets shows that about half fall under three topics: order tracking, returns, and delivery times.

• Weeks 1 to 3: setting up the assistant, connected to the FAQ, the returns policy, and the order status API. The scope is deliberately limited to these three topics, everything else escalates by default.
• Month 1: the assistant runs in parallel, the team reviews its answers and fixes the documentation wherever it stumbles. This is often the real discovery: half of the AI's "errors" are gaps in the documentation.
• Months 2 and 3: about 40% of tickets are resolved without human intervention, that is 440 tickets per month. At 7 minutes on average per ticket avoided, that's about 51 hours per month handed back to the team. CSAT stays stable, and the time to first response drops from several hours to a few seconds within the covered scope.

A detail that matters: about 15% of the conversations handled by the assistant end in human escalation. That's not a flaw, it's the system working. The two support people were not replaced: they now handle carrier disputes, B2B orders, and unhappy customers, with more time for each one. For the same principle applied to the phone, see the AI voice agent for SMEs.

What you should not automate (the honesty of dosage)

A realistic deflection rate sits between 30 and 60% depending on the share of repetitive requests in your business. If a provider promises you 80 or 90%, be wary: either the figure counts customers who give up (that's not a resolution, that's an exit), or the assistant is answering things it should not touch. Some cases must stay human, by design:

• Sensitive complaints: a dispute, a defective product with harm, an angry customer. An automated reply, even a correct one, is perceived there as contempt.
• Vulnerable or struggling customers: a complicated personal situation, repeated misunderstanding. The assistant must detect these signals and hand off early.
• Anything that creates a legal or financial commitment beyond the written rules: an exceptional goodwill gesture, a waiver, a refund outside the policy.

On the compliance side, the CNIL also lays out the ground rules: clearly inform the user that they are talking to a bot, and limit the data collected to what is necessary. It's good practice anyway: an assistant that pretends to be human always ends up exposed, and trust never recovers.

How much it costs (and what it brings in)

On the market, a custom support assistant developed by an agency generally bills between 10,000 and 50,000 € depending on the channels and integrations. Turnkey SaaS solutions look cheaper to start with, but their per-resolution or per-conversation-volume billing climbs fast, and the assistant stays generic: it knows neither your back office nor your tone.

My approach is in between: the building block is already developed and proven in production, and I personalize it to your content, your channels, and your tone. The base feature starts from 2 640 € excl. tax, as a fixed package signed before we begin; personalization (CRM or back-office integration, additional channels) is priced on top, as a fixed package. The monthly running cost (AI API, hosting) generally stays at a few dozen euros.

The return calculation is simple. Take Casabello again: 51 hours of tier 1 saved per month, at a loaded hourly cost of about 30 €, that is on the order of 1,500 € per month. The base pays for itself in a few months, not counting what is measured less directly: the immediate 24/7 response, baskets saved by a quick pre-sale answer, and a support team that no longer wears itself down on the repetitive. Do your own: (repetitive tickets per month × minutes per ticket ÷ 60) × your loaded hourly rate.

Where to start

• 1. A free 30-minute audit: together we look at your ticket volumes, the share that is genuinely automatable, and your existing documentation. You leave with an honest opinion, including if the answer is "not yet" because the documentation isn't ready.
• 2. A deliberately reduced scope: 2 or 3 high-volume request types, the assistant bounded to your documentation, human escalation on everything else. Live in 2 to 4 weeks.
• 3. Measure, then extend: real resolution rate, share of sourced answers, CSAT. We widen the scope only when the figures justify it. And if the need goes beyond support (follow-ups, qualification, summaries), we move over to AI agents and automation.

Sources

• CNIL · Chatbots: advice for respecting people's rights
• European AI Regulation (AI Act, 2024/1689) · transparency of conversational systems
• Supabase · RAG in production (answer reliability)