- What counts as an AI agent
- Customer service: Klarna, Bank of America and the numbers that got published
- Sales and lead qualification: the client story I keep coming back to
- Coding and engineering agents that ship overnight
- Recruiting and HR agents screening before a human ever looks
- Finance and back-office agents doing the reconciliation nobody wants
- Marketing agents running the repetitive campaign work
- Real estate agents using agents to never miss a lead again
- Where these projects fail, and it's not the technology
- How to pick your first use case without wasting three months
- The unglamorous businesses making money from this right now
- Frequently asked questions
- Official documentation
Real AI agents are already answering customer service chats at Klarna, interviewing and matching job candidates at Mercor, turning coding tasks into pull requests on GitHub and matching invoices to purchase orders, with a human checking the exceptions. Klarna's own figures from February 2024 show its assistant cutting the time to resolve a query from 11 minutes to under 2. When projects fail, it's usually not the model: nobody planned an escalation path or a named owner for the conversations the agent gets wrong.
Most of what gets sold to small businesses as an "agent" is a chatbot with a memory, so the definition below matters before the examples.
If you already pay for Microsoft 365, the quickest agent to try may be one you can switch on, so check Microsoft 365 Copilot Business: Who Qualifies and How to Get It first.
What counts as an AI agent
Before the examples, a quick line in the sand, because the term gets stretched to cover everything. A chatbot answers a question. An AI agent takes an action, checks the result, and decides what to do next without a human clicking "approve" each time. If it can only reply, it's not an agent. If it can look something up, make a decision, and then do a second thing based on that decision, it is. I go into the mechanics of building one from scratch in this breakdown of what an AI agent is, but for this post I want the businesses that are already running them, warts and all.
Customer service: Klarna, Bank of America and the numbers that got published
Klarna is the case study everyone quotes, and it's worth quoting because the numbers are unusually specific for a company. In a press release dated 27 February 2024, Klarna said its AI assistant, powered by OpenAI, had handled 2.3 million conversations in its first month live, two-thirds of its customer service chats, doing the equivalent work of 700 full-time agents. Customers resolved their queries in under 2 minutes, against 11 minutes before. Klarna also said the assistant was on par with human agents on customer satisfaction, led to 25% fewer repeat inquiries because it resolved more accurately, and was available in 23 markets, around the clock, in more than 35 languages. Bear in mind these are the company's own launch figures, not an independent audit, and they describe the first month, not the years since. Bank of America's Erica is often named as the older, quieter version of the same story. Its usage totals are quoted widely and change often, so take any current figure from Bank of America's own newsroom rather than a total quoted second-hand. Klarna is the case to learn from, because it published dated before-and-after numbers you can hold a project against.
Sales and lead qualification: the client story I keep coming back to
A recruitment agency I worked with in Manchester was losing candidates simply because nobody could screen applications fast enough. Their consultants were spending around 40 minutes per candidate reading CVs, checking basic fit, and drafting a first response, and with 200 to 300 applications a week for busy roles, good candidates were sitting in an inbox for up to nine days before anyone replied. By the time a human got to them, the strong ones had already taken another job. We built a simple agent using a workflow tool connected to an LLM: it read each CV against the job spec, scored it against five fixed criteria, drafted a personalised reply within minutes, and only flagged the top-scoring candidates for a human to review and book in. Screening time dropped from 40 minutes to roughly 4 minutes of human oversight per candidate. Time-to-shortlist went from 9 days to 2. It didn't replace a single recruiter. It just meant the recruiters stopped doing the part of the job that was pure admin and started spending their time on the calls, which is where they were good. That's the pattern worth noticing: the agents making real money for real businesses right now are rarely doing the flashy, headline task. They're doing the boring middle step nobody wanted, which is exactly why the humans downstream get faster and better at the part that needs a human.
If you're in Queensland and want help finding that boring middle step in your own business, the Brisbane AI consultant page has the local detail.
Coding and engineering agents that ship overnight
When Cognition Labs unveiled Devin in March 2024, it was billed as the "world's first AI software engineer". Its Wikipedia entry records both praise and scepticism, including doubts that it could cope with tasks with more intricate requirements, and YouTube channels such as Internet of Bugs criticising one of its promotional demos, an Upwork project. The working pattern that stuck is more modest: GitHub's documentation for its Copilot cloud agent says it works on one branch at a time and can open one pull request per task, so a person reviews the change as a pull request rather than watching the agent type. Teams using these tools report the agent handling small, well-defined fixes (a broken link, a failing test, a minor UI bug) end to end, while anything architecturally messy still needs a person from the start. That's the clearest statement of the rule for this whole post: agents are strong on narrow, repeatable, checkable work, and weak the moment judgement calls or missing context are involved.
Recruiting and HR agents screening before a human ever looks
Beyond the recruitment story above, venture-backed platforms have built the same idea at scale. Mercor, founded in 2023, started out connecting freelance programmers in India with companies in the United States and built an AI platform to interview them and match them with hiring companies, according to its Wikipedia entry. It has since shifted towards supplying expert contractors to AI companies and says it managed 30,000 contractors as of October 2025. The risk deserves saying plainly: these systems inherit whatever bias sits in the historical hiring data they're trained or tuned on, and a badly built one will quietly filter out good candidates for reasons nobody can fully explain after the fact. Any business running one of these needs a human spot-checking a sample of rejected applications every month, not just the shortlisted ones.
Agents work best when staff know how to supervise them, and AI training for the people who will run them covers that.
Finance and back-office agents doing the reconciliation nobody wants
Invoice processing and reconciliation is where agents are quietly saving the most money with the least attention, because nobody writes a case study about matching purchase orders. A typical finance agent is set up to read an incoming invoice, match it against a purchase order and delivery note, flag mismatches over a set threshold, and pass only the exceptions to a person. One mid-sized manufacturer I spoke with cut its accounts payable processing time from around three days to same-day for roughly 80% of invoices, with the remaining 20% (the mismatched or unusual ones) still going to a human. Treat that split as the shape to aim for across finance, HR and operations rather than a fixed law: agents clear the routine majority, humans handle the exceptions, and no one publishes a reliable ceiling for the share an agent can take. Raise it only after a month of checking that accuracy holds.
If you'd rather have a team build these for you, read about hiring an AI automation agency before you start.
Marketing agents running the repetitive campaign work
Marketing teams are using agents for the parts of the job that used to eat a Tuesday afternoon: pulling last week's campaign numbers into a report, drafting the first version of ad copy variants for testing, tagging and routing inbound leads by source and intent, and scheduling social posts based on when an audience is online rather than a guess. I've written in more detail about how this looks in practice, with named tools and real workflows, in this piece on AI automation for marketing teams, but the short version is the same as everywhere else in this post: the agent drafts, a human approves, and the time saved goes into the strategy work that moves numbers.
Real estate agents using agents to never miss a lead again
Property is one of the clearest wins because response speed is almost the entire game: a buyer who gets an answer while still looking at the listing is far more likely to book a viewing than one answered the next morning. Tools built for this space now qualify inbound enquiries by text or voice, answer the obvious questions about a listing, and only pass a warm, qualified lead to the actual agent. I've covered what this looks like role by role, including what a dedicated AI consultant for real estate agents does day to day, because the setup work here is fiddlier than most people expect once you factor in MLS data feeds and compliance around who's allowed to say what to a buyer.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Where these projects fail, and it's not the technology
The most quoted forecast comes from Gartner. Its press release of 25 June 2025 predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. The same release warns about "agent washing", vendors rebranding chatbots, AI assistants and robotic process automation as agents, and estimates that only about 130 of the thousands of agentic AI vendors are real. That's the definition test from the top of this page, applied to the people selling to you. I've seen this up close. A business I advised launched a customer support agent with real enthusiasm, then quietly turned it off eleven weeks later because nobody had built an escalation path for the roughly 15% of conversations the agent couldn't resolve, and those customers were leaving angrier than if a human had just answered slowly in the first place. The agent itself worked fine. The plan for its failures didn't exist. Agents also hallucinate, confidently, and the businesses getting burned are usually the ones that assumed an agent checking its own work is the same as a human checking it. It isn't. I've written a full explanation of why this happens and what reduces it in this piece on why AI agents hallucinate incorrect information, and it's worth reading before you let any agent send an email or make a booking without a review step.
Before any agent goes live, write down its escalation path. A workable starting checklist:
- Trigger words that send a conversation straight to a person: refund, complaint, cancel, legal, and any word your customers use when they're angry.
- A limit of two failed attempts: if the agent hasn't solved it by then, a human takes over and tells the customer so.
- For the first month, someone reads 10 random conversations every day, not just the flagged ones.
- One named owner who fixes the instructions when a mistake turns up, rather than a shared inbox nobody watches.
For a wider list of risks to plan for, the NIST AI Risk Management Framework, linked at the bottom of this page, sets out categories you can turn into checks.
How to pick your first use case without wasting three months
- Write down every repeatable task in the business that takes over 20 minutes and follows the same steps every time. That's your shortlist, not the exciting stuff.
- Pick the one where a wrong answer is cheap to fix, not the one touching money, health, or legal risk, for your first attempt.
- Set a number before you start: hours saved, response time, or cost per unit, and check it at 30 and 90 days.
- Build in a human checkpoint for anything the agent isn't fully confident about, and define "not confident" in writing before launch.
- Budget for the boring bit, the data cleanup, the escalation rules, the monitoring, because that's often most of the work and almost none of the demo.
Here's how that looks with example values for a small booking business. The task: replying to booking enquiries, which takes about 25 minutes a day, five days a week. The target: a first reply within 10 minutes of every enquiry, checked at 30 and 90 days. The escalation rule, written before launch: the agent hands over to a person when it isn't sure of the answer or the customer sounds annoyed. If the 30-day check shows replies inside 10 minutes and few handovers, widen it; if not, fix the instructions before adding anything new.
If this is your first serious attempt and you'd rather not learn the escalation-path lesson the hard way, working through the first pilot with someone who's built a few is usually cheaper than the failed one, and I break down what that support should cost in this guide to AI consultant costs.
The unglamorous businesses making money from this right now
Away from the big brand names, plenty of solopreneurs and small agencies are building narrow agents as a service, one that qualifies leads for local trades, one that handles booking confirmations for clinics, and charging a monthly fee for it. It's a repeatable small business model right now, and I've listed it alongside other realistic paths in this list of ways to make money online for anyone weighing it up as a service to sell rather than just a tool to buy.
Recruiters who want a screening agent like the Manchester one can read about working with an ai consultant recruitment agencies screening follow up, and if you've built an agent yourself and want to write about it, the llm (this site) guidelines explain how to pitch.
Frequently asked questions
How fast did Klarna's AI assistant resolve queries?
Klarna reported in February 2024 that customers resolved their queries in under 2 minutes, against 11 minutes before the assistant. That's the company's own figure from its first month live.
How do I stop an AI agent giving wrong answers?
You can't stop it completely, so plan for it. Add a human checkpoint for anything the agent isn't sure about, set the escalation path before launch, and read a sample of real conversations every day for the first month.
Why do so many AI agent projects get abandoned?
Gartner's June 2025 forecast points to escalating costs, unclear business value and inadequate risk controls, and in practice the most common cause I see is a missing plan for what happens when the agent gets something wrong, not a failure of the technology itself.
How much does it cost a small business to build its first AI agent?
There's no standard price. Simple agents built on existing tools like Make.com or Zapier connected to an LLM cost whatever those plans charge, which depends on how many tasks you run, plus your setup time. Custom-built agents cost far more once you add data integration and testing. Check the current pricing pages before you budget, and cost the time for monitoring too, not just the software.
Official documentation
Run a business, a brand or a blog? I feature tools, products and expert guest posts for an audience of 15,000. See how it works