Is that account's 50% what I should expect in my business?

Not necessarily. It is a building management company, with fairly repetitive requests and a lot of written information. A business with more variable questions will land lower, and one with a heavily documented catalogue can land higher. The only number that matters is the one you get from counting your own conversations.

What happens to a request it cannot resolve?

It is handed off by rule, not by chance. You define which topics always stop. When one stops, an email arrives with the reason and a direct link to the conversation, and whoever picks it up sees the whole thread.

How do I check it is answering well and not making things up?

You open the conversation and read it. The actions that matter are logged right there as private notes your team sees and the customer does not: an order created, a past order looked up, a voice note transcribed, and a handoff with its stated reason.

Can it be tested before it talks to real customers?

Yes. It is tested in a safe environment, picking real contacts from your address book to see how it would respond in concrete situations, and adding notes to a contact to simulate unusual cases.

How many support requests can AI actually resolve?

The short answer

How many support requests AI can resolve depends almost entirely on what you asked it to resolve. There is one reference number worth keeping close: according to a global market-research survey of 5,728 customers, only 14% of customer service issues are fully resolved in self-service.

That 14% is the ceiling of classic self-service: the FAQ page, the menu tree, the form that hands you back a help article.

In one account we run, a building management company with around 312 conversations a month, AI resolves roughly 50% and hands off about 40% to a person. That is one account, not a market average. But the distance from 14% is not an accident, and it is worth understanding where it comes from.

Why self-service stalls at 14%

Traditional self-service does not converse. It matches your question against content that already existed and hands it back.

That works when your question is written down exactly as you asked it. It breaks the moment the question carries context: "the lift in block B has been making a strange noise since Thursday and I already reported it last week." There is no help article for that.

The outcome is familiar. The person tries, does not find it, and writes in anyway. The request was not deflected. It was delayed.

What an agent with context does differently

Three things move the number, and none of them is magic.

It reads your documentation, not a keyword list. If you uploaded the building rules, the price list or the returns policy, it answers from those. The answer to a question you never anticipated may well already be in your own documents.

It remembers the customer. Someone who wrote in last month does not start from zero. That alone removes a large share of the back-and-forth that makes people give up.

It knows when the request is not its to take. This is the one that weighs most. An agent that hands off well resolves more, because it does not burn the customer's patience trying to solve what it cannot.

The 40% that gets handed off is not a failure

In the account above, close to 40% of requests escalate to a person. It is tempting to read that as the share the AI "could not do". It is the other way around.

Those are requests that need coordinating with a contractor, sending someone to the building, or making a decision that costs money. No automation should be closing those on its own. What was gained is that they arrive sorted, classified and with the full conversation attached, instead of showing up as a loose message among forty others.

A separate survey, across 3,566 customers, found that 87% believe a company using generative AI in support has to give them access to a human. The route to a person is not the fallback plan: it is part of what people expect before they accept any of the rest.

How to estimate your own number

You do not need to start from a hypothesis. One week of your own conversations is enough:

Count how many requests came in.
Mark how many are answered with information already written down somewhere in your business.
Mark how many need a decision, money, or someone physically going somewhere.
The first group is your automation floor. The second is what you will always hand off.

The first group is almost always bigger than the owner estimates before counting. And the second is almost always bigger too, relative to what any demo promises.

One forecast and one warning

The firm behind that 14% figure projects that by 2029 AI will autonomously resolve 80% of common customer service issues. That is a five-year forecast, about common issues, across companies of every size.

The warning is that the number describes a destination, not a starting line. If you begin expecting 80% next week, you will read a real 50% as a failure. And a real 50%, with the other 40% handed off in good order, is already the difference between a team underwater and a team that keeps up.

Conclusions

A global survey measured that only 14% of issues are fully resolved in self-service, across 5,728 customers surveyed.
In one real account with around 312 conversations a month, AI resolves close to 50% and hands off close to 40%. One account, not a market average.
The difference comes from three things: answering from your documents, remembering the customer, and handing off well.
The handoff share measures what should not be resolved automatically. Reading it as failure is the most common mistake.
87% of customers expect to be able to reach a person, per that same research.
Estimate your own number by counting one week of conversations before you buy anything.

How many support requests can AI actually resolve? What self-service promises and what one real account delivered