Back to blog

AI support tools: why Humane does not build ticket bots

Published on November 22, 2026

A Humane figurine with a headset raises a hand against a small ticket dispenser machine while another looks on.

You know the situation from the other side. You ordered from an online shop, the parcel does not arrive, you click “Help”. A window opens: “Hi, I'm your virtual assistant. How can I help?” You type “parcel hasn't arrived”. The bot replies with three buttons: “Track shipment”, “Delivery times”, “Other”. You click “Other”. The bot asks for the order number, which you are now looking for in the confirmation email. Then you get a link to the carrier’s page, which you already had open ten minutes ago. At the end it says: “Was I able to help?”

From its maker’s point of view, that bot worked. It “deflected” a ticket. From your point of view it cost you five minutes and showed you that the shop does not want to know you. This article explains why we deliberately did not build Humane that way, how a ticket bot differs from an agent that writes answers, and what that means for you as a merchant.

What a ticket bot optimises for

Ticket bots are products built around one metric: deflection. It measures how many customers gave up on their request or solved it themselves before a human was involved. That sounds sensible if you view support as a cost centre. Every deflected ticket is a ticket nobody had to answer.

The problem: the metric does not distinguish between “customer found the answer” and “customer gave up”. Both count as success. A bot that is so inconvenient that customers would rather not ask has an excellent deflection rate. And the customers who gave up do not come back. They leave reviews, they buy elsewhere, they tell people. According to Zendesk CX Trends 2026, 85 percent of CX leaders say a single unresolved issue is enough to lose a customer. A deflected ticket is often exactly that: an unresolved problem that just no longer sits in your inbox.

The second property of a ticket bot: it does not know the customer. It knows a knowledge base and a menu. That is why it asks for the order number, why it sends links instead of answers, why it can only respond to “parcel stuck for five days” with the general delivery time. That is not a weakness of the model behind it. It is the consequence of deciding to deflect the ticket rather than resolve it.

What an agent that writes answers optimises for

Humane is built the other way round. The agent does not sit in front of the inbox to keep customers out, it sits inside the inbox to answer requests. The ticket comes into being the way it always has: the customer writes an email. Then what happens is what a good support person would do too:

  1. Read and classify. What is it about, which order, what mood.
  2. Look things up. Order and customer data from Shopify, shipment status from the carrier, your stored rules.
  3. Reply. An email in your tone, with the concrete information the customer needs: date, amount, next step.
  4. Know when not to reply. Ambiguous order match, angry tone, legal questions, amounts above your limit: those go to you as a draft.

The metric we care about is not the deflection rate but whether the customer had to write again after the reply. A reply that closes the ticket because it was correct and complete is the success. A reply that makes the customer give up is not.

That is why Humane bills by answered customer requests and not per seat or feature. The incentive should sit on the answer, not on the deflection. More on that in Why we bill by requests.

The difference in three concrete cases

The customer without an order number

A ticket bot asks for the order number. If the customer does not have it, the conversation ends or turns into a form. The agent looks up the order in Shopify by the sender address. If it finds exactly one open order, it replies. If it finds two, it does not ask the customer; it puts a draft with both in front of you. The customer does not have to supply anything the shop already knows.

The parcel without a scan

A ticket bot says “Your shipment is on its way” because that is what its template says. The agent checks the status with the carrier. If there is no scan yet, it says so: registered, not yet picked up in the carrier’s system, usually changes by the next working day. That is less reassuring than “on its way”, but it is true, and the customer can see it is true on their next click. How this works in detail: Answering tracking requests automatically.

The threat of a lawyer

A ticket bot does not recognise the word, or replies with “I understand your frustration”. The agent recognises the tone and the trigger words, sends nothing automatically, and puts the ticket in front of you as an escalation. If you want, with a draft you can review. But the decision is yours, and it happens before anything reaches the customer.

Why this is more than a matter of style

You could object: that is just a question of implementation. A good chatbot can look up orders too. True, and there are tools that do. The difference goes deeper, into what the product counts as success and what it learns from that.

A bot that optimises for deflection learns to cut questions short. An agent that optimises for resolved tickets learns from your corrections what a correct reply looks like in your shop. At Humane, every change you make to a draft becomes a rule for similar cases. That only works because there is a draft to correct in the first place. A bot conversation the customer abandoned cannot be corrected. You do not even see it.

There is one more reason, and it is less comfortable. Customers do not particularly like chatbots. In a Bitkom survey from 2025, 62 percent of online shoppers want a quickly reachable human contact when there is a problem, only 36 percent want a chatbot, and only half of those who have used one were satisfied. We do not read that as “customers reject AI”. We read it as: customers reject being pushed into a menu when they have a question. A good reply by email that knows their order is not rejected, no matter who wrote it.

We described the difference between chatbot and agent at length in AI chatbot vs. AI agent in support. This article is about the question behind it: what is the tool actually supposed to do.

What an agent cannot do, and what we therefore do not promise

Honesty about limits is part of this positioning; otherwise it would just be marketing.

  • An agent does not solve problems that sit in the process. If your warehouse does not record tracking numbers in Shopify, nobody can answer delivery questions, no agent either. It makes the problem visible; it does not fix it.
  • An agent does not make goodwill decisions you have not stored. Refund above your limit, exception for a regular customer, replacement without a return: those are your decisions. The agent prepares them.
  • An agent does not replace a human in conflict cases. Second follow-up, threat, bereavement, complaint about a colleague: a human should reply here. The agent recognises these cases, and that is its contribution.
  • “Up to 90 percent automated” is our ambition, not a promise. How much it becomes in your shop depends on your categories, your rules and your training week. What happens beyond that line is described in The 90 percent limit.

In short

A ticket bot is a product that measures how many customers it keeps away from support. An agent that writes answers is a product that measures how many requests were answered well enough that nobody had to ask again. The technology behind them may look similar; the goal does not. We chose the second because the requests in your inbox are not an interruption but customers with a legitimate question. The bigger picture is in AI in customer service: what really works in 2026.

If you want to see what answers instead of deflection look like in your inbox: try 100 customer requests for free, no credit card; in the first week only drafts are created.

Sources