Enterprise AI Agents: What They Are and When They Pay

Seven hundred people, one month

In February 2024 a payments company called Klarna published a set of numbers about its new AI assistant.

In its first month, the assistant had 2.3 million conversations, two-thirds of the company's customer service chats. Klarna said that was the work of 700 full-time agents.

Customers got answers in under 2 minutes, where it had taken 11. Repeat enquiries fell by 25 percent. The company estimated 40 million dollars in added profit for the year.

It became the most quoted AI case study in business. Every board asked where its own 700 were.

Read it again and notice what kind of work it was. Millions of similar questions, with a clear right answer, where a wrong one could be corrected by the next message. Remember those three features.

The other number

Sixteen months later the research firm Gartner published a prediction.

Over 40 percent of agentic AI projects will be cancelled by the end of 2027, it said, because of rising costs, unclear business value, or inadequate risk controls.

Its analyst described most current projects as early experiments driven by hype and often misapplied.

Gartner also named a practice it called agent washing: vendors relabeling existing chatbots and automation tools as agents. It estimated that only about 130 of the thousands of vendors claiming agentic AI were the real thing.

So one company reports the work of 700 people, and an analyst expects four in ten projects to be scrapped. Both are describing the same technology. The difference is where it was pointed.

What an agent is

The word is used loosely, so here are three things that all get sold under it.

A chatbot answers questions. It may be very good. It takes no action in your systems.

A workflow is a fixed sequence of steps with a model doing some of them. Read the email, classify it, draft a reply, send it for approval. The path was decided by a person in advance.

An agent decides its own path. It is given a goal and a set of tools, and it chooses what to do next, looks at the result, and continues until it judges the job done.

That distinction comes from Anthropic's engineering guidance, which separates workflows, where models follow predefined paths, from agents, where models direct their own process. Its advice is to use the simplest thing that works and add autonomy only when it is needed.

That advice is worth more than most vendor pitches. A workflow is cheaper, faster, and easier to check. Choose an agent when the steps cannot be listed ahead of time.

What surrounds the model matters as much as the model, which we cover in what an agent harness is.

Finish-Line Firms: when an agent pays

The companies getting a return from agents did not start with the technology. They started with a piece of work that had a finish line. We call them Finish-Line Firms.

They pick volume. Thousands of similar cases a month. An agent that saves ten minutes on a task done twice a year saves twenty minutes.

They can define done. A refund processed. A ticket resolved.

An invoice matched. If two of your own staff would disagree on whether a case was handled well, an agent cannot be judged either.

They make mistakes cheap. Either the action can be undone, or a person approves it before it becomes final. Agents are wrong some of the time, and the design assumes it.

They fix the process first. If nobody can describe how the work is done today, automating it produces confusion at higher speed. See operations problems disguised as AI problems.

They measure cost per case. Not the license. What each resolved case costs in model usage against what it cost before.

The method is in controlling AI token spend.

They name an owner. One person answers for what the agent does. We explain why in who owns AI.

"We are not Klarna. Is this for us?"

You may run a firm of forty people and wonder whether any of this applies. The test is the same at any size, and a smaller firm passes it less often.

You probably do not have millions of similar cases. You may have a few hundred of something each month: enquiries, quotes, invoices, appointment changes.

For most of those, a workflow is the right tool. A fixed sequence with a model at two or three steps and a person approving the result. It is not an agent and does not need to be.

Be careful with the headline case too. It was one company's own account of one month, and customer service is close to the ideal use. Results elsewhere will be smaller.

Ask a vendor three questions. What does it do without being told each step? What can it change in our systems, and how is that limited?

How will we know when it is wrong?

For a structured way to think about that last one, the US standards body's AI Risk Management Framework is free and readable. Our own checklist for judging a product is in how to evaluate AI agents.

What to read next, and why

Klarna AI assistant press release (company announcement): The original numbers. Read it as a best case from the party with most reason to be pleased.

Gartner on agentic AI project cancellations (analyst press release): The prediction, the reasons, and the term agent washing.

Building effective agents (engineering guide): The clearest plain-language account of workflows versus agents, from a company that builds the models.

AI Risk Management Framework (NIST): A government framework for thinking about what can go wrong and who answers for it.

So the simple takeaway is this: before asking which agent to buy, find the piece of work with the most volume and the clearest finish line. If you cannot find one, you do not need an agent yet.