Which industries get real results from AI agents today? Banking, retail, healthcare, software and travel top the list. Each one has public numbers from live deployments.
An AI agent is software that finishes a task on its own. It can answer a refund question, change a booking, draft a clinical note or write a pull request. It only calls a person when it gets stuck.
That last part matters. An agent earns its cost where the same requests arrive thousands of times a day and the answers follow written rules. Some industries have far more of that work than others.
1. Banking and fintech
Money questions repeat. Where is my payment? Why was my card declined? When is the next instalment due? Banks and payment firms answer these all day, in many languages, and the answers live in account data an agent can read.
Klarna is the best-known example. Its AI assistant handled 2.3 million conversations in its first month of global use. That was two-thirds of all Klarna customer service chats, or the work of about 700 full-time agents. Average resolution time dropped to under 2 minutes, down from 11 minutes. Repeat inquiries fell by 25%. The assistant runs in 23 markets, 24 hours a day, in more than 35 languages, and Klarna estimated a $40 million profit improvement for 2024.
Bank of America has run its assistant, Erica, since 2018. By April 2024, Erica had passed 2 billion interactions with more than 42 million clients. More than 98% of clients get the answer they need from Erica within 44 seconds on average.
Banking still has hard limits. Fraud reports, disputes and hardship requests need a human, and regulators expect a record of every decision. So the agents that work best in finance do the lookup work and pass anything with judgement or risk to staff.
2. Retail and e-commerce
Online stores get the same five questions again and again. Where is my order? How do I return this? Does it come in a bigger size? Can I change the address? None of these need a person if the agent can see the order system.
New Look, the UK fashion retailer, shows what this looks like at scale. Its support team handles about 9,500 email and chat inquiries a week. With Zendesk AI agents, 42% of customer queries were resolved without a human. In total, 305,761 contacts were deflected across email and chat. First reply time fell by 99.5%, from 16.5 hours to about 5 minutes.
Retail has more choice here than any other sector. There are now dozens of AI agents that answer tickets, chats and calls, and many plug straight into Shopify or a helpdesk. Some are free for a small number of conversations each month, so a shop can test one on real traffic before paying.
The weak spot is product advice. An agent that guesses at sizing or stock can create returns. Feed it the real catalogue and return policy, and let it hand off when a shopper asks something the data does not cover.
3. Healthcare
Doctors spend hours on notes. Every visit needs a record, and much of that typing happens after clinic hours. Ambient AI scribes listen to the visit, with the patient’s consent, and draft the note for the doctor to review.
Kaiser Permanente has the largest public data set on this so far. Over 63 weeks, 7,260 Permanente physicians used AI scribes in more than 2.5 million patient encounters. The tool saved an estimated 1,794 working days in one year, which is close to five years of one person’s working hours. The Permanente AI scribe analysis also links the tool to better patient conversations and higher work satisfaction.
Healthcare agents go beyond notes. Front-desk agents answer calls, book visits and send reminders. Billing agents check insurance bills before they go out.
Still, healthcare carries the highest stakes on this list. Patient data falls under HIPAA in the US, so any vendor that touches it needs a business associate agreement. Clinical decisions stay with clinicians. The agent drafts, and the doctor signs.
4. Software development
Code is text with strict rules. It either runs or it does not, and tests show the result in seconds. That makes software one of the easiest places to check an agent’s work.
GitHub ran a controlled experiment with 95 professional developers. Each one had to write an HTTP server in JavaScript. The group using GitHub Copilot finished in 1 hour and 11 minutes on average. The group without it took 2 hours and 41 minutes. That is 55% faster. The Copilot group also had a higher completion rate, at 78% against 70%.
That study covered an AI pair programmer. Today’s coding agents go further. They read a whole repository, plan a change, edit several files, run the tests and open a pull request for review. AI Agents Choice has 67 agents in its coding category alone, next to 168 task and industry categories in total, with prices and user ratings for each.
Teams still need code review. An agent can pass the tests and still add a security hole or a slow query. Most teams give agents small, well-scoped tickets first and widen the scope as trust grows.
5. Travel and airlines
Travel has a support problem that spikes without warning. One storm can ground a fleet, and thousands of passengers want to rebook at the same moment. No call centre can hire fast enough for that.
Air India built its virtual assistant, AI.g, for this load. It handles about 40,000 queries a day and covers more than 1,300 types of questions, such as booking changes and refund requests. Since launch it has run more than 13 million conversations with a 97% success rate. Only about 3% of queries go to a human agent. Even so, about half of customers still choose a person first, so Air India kept both options open.
Hotels follow the same path. Voice agents now pick up front-desk and reservation calls, so staff at the desk can focus on guests in the lobby.
What these industries have in common
The five sectors look different on the surface. But the same conditions show up in each one:
- High volume of repeat requests. Klarna, New Look and Air India all handle thousands of near-identical questions a day. Volume is what turns a small time saving per request into hundreds of staff hours.
- Answers stored in a system. Order data, account data and booking data let the agent give a real answer. Without that access, the agent can only repeat the FAQ.
- A safe way to check the output. Code has tests. Clinical notes have a doctor’s review. Industries without a check step need one before an agent goes live.
- A clean handoff to people. Every example on this list keeps humans for the hard cases. The 3% that Air India escalates is still thousands of conversations a month.
How to test an AI agent in your industry
Pick one narrow job first. “Answer order status questions” is a better first project than “handle customer service.” Measure how long that job takes today, then run the agent on a slice of real traffic for a few weeks.
Track three numbers: share of requests resolved without a person, time to first reply, and how often customers come back with the same issue. Repeat contacts reveal if the agent fixed anything. A fast answer that does not fix the problem only moves the work to next week.
Write the handoff rules before launch, too. Decide which words or topics send a chat straight to a person, such as a complaint, a legal threat or a safety issue, and test those paths on the first day.
Finally, read the vendor’s data terms before you connect any customer records. In banking and healthcare, that step is not optional.
Common questions about AI agents by industry
Which industry uses AI agents the most?
Customer service across retail, banking and travel has the most live deployments, because the requests repeat and the results are easy to count. Software development is close behind.
Do AI agents replace staff in these industries?
In the examples here, agents took over routine volume and people kept the complex cases. Klarna’s assistant did the work of about 700 agents, but disputes, fraud and medical decisions still go to humans.
Are AI agents safe for regulated industries?
They can be, with limits. Banks need audit trails, and healthcare providers need HIPAA-compliant vendors and clinician review. Keep the agent away from final decisions until it has a long, checked track record.
