Writing Agents
Give the agent a job description
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Klarna shows what happens when an agent is judged on cost alone.

In February 2024 Klarna made the announcement every chief executive wanted to hear. Its new assistant, built with OpenAI, had handled 2.3m conversations in its first month, two-thirds of all customer service chats, and was doing the work of 700 full-time agents. Customer satisfaction was on par with humans. Klarna expected it to add $40m to profit in 2024.
In May 2025 Sebastian Siemiatkowski, the chief executive, told Bloomberg that Klarna would hire people again, so that customers could always reach a human. Cost, he said, had been "a too predominant evaluation factor", and the result was lower quality.
Klarna did not give up on AI. It kept the assistant and, by its own account, gave it more work. What it had lost was a clear idea of which conversations belonged to the machine and which to people.
The hype has a forecast
Gartner, a research firm, expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls. It also reckons that only about 130 of the thousands of vendors selling agentic AI offer the real thing. The rest have renamed their chatbots. McKinsey found in 2025 that 23% of companies were scaling an agent somewhere in the business, but in no single function did more than 10% do so. Its 2026 survey counted 40% among companies with more than $1bn in revenue and 22% among smaller ones.
Write the job description first
An agent is software that takes a task from start to finish. It reads the request, looks things up, decides, acts and records what it did. That makes it closer to an employee than to a tool, and it should be hired like one. Before building anything, write down what a new hire would get on day one.
What is the job, in one sentence? Which systems may the agent read, and which may it change? Up to what amount may it decide alone: a refund of CHF 50, a discount of 10%, an order of 200 units? Whom does it ask when it is unsure, and how fast must they answer? How will you check its work in the first month, and what counts as good?
Klarna's first version answered the cost question very well and the quality question hardly at all. A job description puts both on the table before the first line of code.
Small jobs first
The agents that work in practice start narrow: one inbox, one kind of request, one system to write into, and a person who approves anything unusual. Every step is logged, so the agent can be audited like a colleague. Once it does that job well for a month, it gets a second one. That is how companies promote people. There is no reason to treat software more generously.
Gartner's 40% will mostly be projects that skipped this step, agents told to transform the customer experience with no limits, no owner and no measure. The ones that survive will look boring from the outside. They will still be running in 2028.
Sources
- Klarna press release via PR Newswire: Klarna AI assistant handles two-thirds of customer service chats in its first month, 27 February 2024
- Fortune, after Bloomberg: Klarna turns back to humans, 9 May 2025
- CMSWire: Klarna CEO, we're giving AI more customer service work
- Gartner: Over 40% of agentic AI projects will be canceled by end of 2027, 25 June 2025
- McKinsey: The state of AI in 2025, 5 November 2025
- McKinsey: The state of AI in 2026, on the road to ROI, 25 August 2026
More writing
ApplicationsWhy most AI pilots go nowhereMIT found that 95% of corporate AI pilots make no measurable difference to profit. The fault lies in the last mile, where software meets the company.Author: Rea Stark
IntegrationThe plumbing is the productRivals adopted Anthropic's MCP. OpenAI and 19 investors put over $4bn into a firm that places engineers with customers. AI's value is in the connections.Author: Rea Stark
LabThe smallest model that does the jobNVIDIA's researchers argue most work inside AI agents is better done by small models. For Swiss companies that is good news on cost, speed and privacy.Author: Rea Stark