Automate customer replies without the robot feel

Every SME that opens a customer support project shows up with the same fear: “I don’t want my customers to feel like they’re talking to a bot”.

Fair fear. But the problem isn’t automation. The problem is automating the wrong things, or handing off to a human too late.

Here’s the method we use to automate customer replies in an SME without making the service feel like a robot.

Key takeaways:

• The annoyance doesn’t come from knowing you’re talking to an AI, it comes from dead-end loops with no visible human exit.
• Automate the low-risk repetitive volume (triage, FAQ, data collection), leave the emotional and ambiguous to humans.
• Define escalation rules BEFORE going live: which signal triggers the handoff to a person.
• On Navily we cut operational time on moderation and enrichment by 70% while keeping humans on the cases that matter.
• Without a baseline (what answering costs today) you can’t say the agent improved anything.

First: split tickets by risk, not by volume

The classic mistake is automating starting from the most frequent tickets. Wrong starting point. The right criterion is risk, not frequency.

Split tickets into three buckets:

• Low risk, high frequency: where is it, order status, opening hours, resets, document requests. The agent handles these end-to-end.
• Medium risk: requests the agent can prepare (collect data, draft a reply) but a human approves and sends.
• High risk: complaints, cancellations, emotional cases, legal or ambiguous requests. Immediate escalation to a person.

This map is 70% of the work. It’s the same triage and routing logic we apply on every customer support project.

Where the AI agent actually helps

The agent excels at the invisible work that drains the team without adding value:

• Triage: reads the incoming ticket, understands category and urgency, routes it.
• Instant first reply: for low risk it answers immediately, killing the wait.
• Context collection: before handing off to a human, it asks for the data needed (order number, screenshot, version), so the person starts already informed.
• Moderation and enrichment: on user-generated content, the agent filters and enriches at scale.

On Navily, a boating community, this approach cut operational time by 70% on moderation and enrichment. Not by replacing the team, but by taking the repetitive work off their plate.

Where it must hand off to a human

This is the difference between an automated service and a service that feels like a robot. You define the escalation rules before going live:

• At the first sign of frustration: if the customer raises their tone or repeats the same question, hand off to a person. Not on the fifth attempt.
• On explicit request: the word “agent” (or a button) must always reach a human. A hidden exit is the single most infuriating thing.
• On out-of-scope cases: if the agent isn’t sure, it doesn’t improvise, it escalates.
• On anything with financial or legal impact: cancellations, disputed refunds, formal complaints.

Golden rule: one escalation too many beats one too few. A human called unnecessarily costs a few minutes. A customer stuck in a loop costs the customer.

Tone: transparent, not fake-human

The AI Act classifies client-facing chatbots as limited risk: you must tell the user they’re talking to an AI. This isn’t a problem, it’s an advantage.

Customers don’t get angry because they know it’s an agent. They get angry when the agent pretends to be human and fails. An agent that clearly states what it is, answers fast, and hands off when needed beats any bot posing as a person.

No forced enthusiasm, no emoji spam. Dry, useful tone, with a human exit always within reach.

When NOT to automate (I’ll say it to your face)

• Few, highly variable tickets: below a certain repetitive-volume threshold, a custom agent is a waste. A template and a sharp person are enough.
• Hyper-sensitive brand voice on every reply: if every single reply is a marketing act, keep the human up front and use AI only for internal triage.
• Zero baseline: if you don’t know how much time or how many minutes of wait your support costs today, don’t automate yet. Measure first, then decide.

When the numbers are there, we build the agent on the real process, with an audit log on every decision and escalation rules written into the contract.

Next step

If your support is drowning in repetitive work but you don’t want to lose the human touch, let’s talk for 20 minutes. We’ll tell you honestly which tickets to automate and which not to, with the “you pay only if it works” guarantee.