Back to blog

The 90 percent limit: how much support automation is too much

Published on October 15, 2026

A Humane figurine with a headset stands at a dial turned almost to full, one last drawer deliberately left open.

Our homepage says “up to 90% automated”. Now and then someone asks why not 100. Technically it could be done, at least on paper: every email gets read, every one gets a reply, nobody has to look at the inbox any more.

We decided against that deliberately, and this article explains why. Not because we couldn't manage the last ten percent. But because the last ten percent are the cases where a customer decides whether to come back, and where an automatic reply achieves the opposite of what it's meant to.

You'll learn which cases those are, how a system recognises them, what the limit looks like in a concrete shop, and why the question “how many percent” is actually the wrong one.

Why 100 percent is a promise nobody should make

A support inbox isn't a uniform mass. It consists of a large block of recurring requests whose answer sits in a system, and a small remainder made up of exceptions: the parcel that was scanned as delivered and isn't there. The customer writing for the third time. The case where two orders run under the same address and nobody knows which one is meant.

The large block can be automated, and well. Gartner predicts that agentic AI will resolve 80 percent of common customer service issues without human intervention by 2029. The word that matters is “common”. The forecast refers to the standard cases, not to everything.

The small remainder has three properties that make it unsuitable for automation: the data is incomplete or contradictory, it's about discretion rather than rules, or the customer is in an emotional state where they don't want an answer but someone responsible. At 100 percent automation, exactly these cases land with a machine that can't solve them but replies anyway. That's the worst combination.

The cases that always need a human

The following list is the basis for what never goes out automatically at Humane. It isn't complete, but it covers most of it.

  • Contradictory data. Shipment “delivered” according to the carrier, customer says nothing has arrived. Two orders with the same email address, customer doesn't mention a number. A reply that simply skips one of these contradictions is very likely wrong.
  • Anger. The third email on the same topic, capital letters, words like “lawyer”, “consumer protection”, “review”. The customer doesn't want the best wording here; they want the feeling that someone is taking over the case. An automatic reply, however good, sends the opposite message.
  • Money leaving the shop. Refunds, goodwill, replacement shipments. The AI can make the suggestion and prepare everything. Approval should sit with a human, because mistakes here get expensive directly, and because patterns that hint at fraud only show up across several cases.
  • Discretion. A return after the window has closed, damage where it's unclear who caused it, a regular customer with an unusual request. There's no rule that always fits, and the decision shapes the relationship.
  • Anything the AI isn't sure about. Questions about products that aren't in the knowledge base. Statements on warranty or withdrawal that go beyond your stored rules. The right answer here is “I don't know”, and only a human can phrase that so it doesn't feel like being brushed off.

How to handle the second point when it lands on your desk is in Complaint management: de-escalation in 5 steps.

What customers say about it

The numbers are clear, and they contradict the image of the customer who just wants a quick answer. According to a Gartner survey from 2026 of 3,566 customers, 87 percent say a company using generative AI in customer service must offer a way to reach a human. At the same time, half of respondents find that interactions have become easier through generative AI.

Together, that gives the actual rule: customers accept automation as long as it doesn't become a wall. The path to a human doesn't have to be used often. It has to be visible.

On the company side, another Gartner survey from 2026 shows that 85 percent of service leaders are expanding the responsibilities of their human staff rather than cutting them. For a small shop that doesn't mean you have to hire someone. It means the time automation frees up has a purpose: the cases where a human makes the difference.

What the limit looks like in a concrete shop

A worked example with assumptions. Suppose a shop gets 500 requests a month. Suppose the distribution looks roughly like this: 180 delivery questions, 100 returns and refund questions, 70 order changes, 80 product questions, 40 complaints, 30 other.

Of the 180 delivery questions, most have a clear status and can be answered automatically. Some show a contradiction or no scan for days. Those go to you as drafts. Of the 100 returns questions, the ones about process and deadlines can be answered automatically. The refunds themselves go through an approval. The 70 order changes depend on the order status; the ones where the order hasn't shipped are easy. Product questions go automatically if the knowledge base has the answer, otherwise not. The 40 complaints all go to you, with a prepared draft you adapt.

Add it up and in this example you land at a high share answered without you, and a remainder of perhaps 50 to 80 cases that need your time. That's the range we mean by “up to 90 percent”. The exact number depends on your range, your rules and the quality of your knowledge base. It isn't the goal. The goal is that the 50 to 80 cases are the right ones.

Here's how it works at Humane: the agent decides per request whether it's confident enough. Only what falls into an approved category and where the data is unambiguous goes out automatically. Everything else is prepared as a draft. Angry customers are recognised and handled differently, and escalations can go straight to a human. Which tickets go out without approval in practice and which never do is in Auto-send: which tickets Humane answers without approval.

How to tell you've gone too far

The limit moves, and you should watch it. Four signals show that too much is running automatically:

  1. Follow-ups rise. Customers write again after the automatic reply. That means the reply didn't close the question.
  2. Reviews mention support. Sentences like “you only get standard replies” or “nobody reachable” are a direct signal.
  3. Drafts get approved unread. If you're only waving drafts through, the category is either ready for automation or you've already given up control without deciding to.
  4. A case surprises you. A customer calls or writes on another channel, and you hear for the first time about a problem that's been running for three emails. Then automation kept a case it should have handed over.

In all four cases the reaction is the same: switch the affected category back to draft, watch for a week, sharpen the rules, then enable again.

Why “how many percent” is the wrong question

The number on the homepage is an ambition, not a promise. It says the standard cases should run without you. It doesn't say your support runs without you.

The better question is: which cases should land with me? Once you have that list, the percentage follows on its own, and it's different for every shop. A shop with a simple range and clear return rules sits higher. A shop with products that need consultation or lots of special requests sits lower, and that's as it should be.

A vendor promising you 100 percent has either not yet seen the difficult cases or plans to handle them with a standard reply. Both are reasons to look more closely. The difference between a system that answers and one that looks things up is described in AI chatbot vs. AI agent.

In short

Support automation has a natural limit, and it isn't set by the technology but by the cases: contradictory data, anger, money, discretion, not knowing. These cases need a human, not because the AI couldn't answer them, but because an automatic reply here costs the trust the other 90 percent have built. Automate the standard cases fully, keep the path to a human open, and measure follow-ups and reviews to notice when the limit slips. The full overview of AI in support is in AI in customer service: what actually works in 2026.

If you want to see which of your requests the agent would hand to you and which it wouldn't: try 100 customer requests for free, no credit card, with a training week in which everything stays a draft.

Sources