I’d like to present an alternative perspective on AI for software development, using Klarna as an example.

Klarna

If we had to pick one company touted as the earliest adopter of AI, the first actually-using-it-in-the-enterprise company, It’s probably Klarna. Over the past year, The Swedish buy-now, pay-later provider claimed that it had stopped hiring, replacing turnover productivity with AI. Then it claimed to be shutting down it’s systems of record (including Salesforce and Workday) in favor of AI solutions. In addition, the company had replaced 700 customer service workers, about 2/3rd of their team, with AI-driven customer service. The CEO even gave his most recent quarterly update not as himself, but using an AI generated avatar.

That’s great, right?

Impressive.

Amazing.

But … what if it isn’t?

One thing you’ll notice about the examples above is they have almost no detail. What does the AI Chatbot do, exactly? Does it resolve customer tickets? HOW did they replace SalesForce with AI? Salesforce, for example, has a huge number of components, from CRM to tracking the sales process to generating quotes to lead generation and marketing. What did they do? The initiative answer is they used a tool like chatGPT to type in every step of the CRM process so they could ask for updates later, but … is that right?

A different interpretation

As it turns out, Klarna did not replace Workday and Salesforce with AI. Instead, they are using a collection of cheaper tools. When I heard about the chatbot, I expected I would be able to type in my order number and get specific information about my order. I expected the bot would take corrective action, say, initiating a fraud complaint for me. Instead, while the customer chatbot can serve up two-thirds of customer service answers, it seems to be just a way to enable self-service. Essentially, chatGPT reads all the existing documentation, then the conversational layer allows the user to “talk” to a bot that regurgitates the manual for that particular problem. This is a minor improvement over help guides and documents. In an era where interactive voice response has been resolving, 70-90 percent of calls can be resolved with Interactive Voice Response (pre-ai, phone call dial in), the number is less impressive. The avatar is interesting; the CEO seems to have taken his financials, asked an LLM to generate a script, tweaked it, then used tools to create a 2 minute presentation. Given how deeply he is involved in his finances, I’m not sure how much time this saved over just creating slides and making a presentation with Camtasia. Personally, I’m apt to sneeze, cough, or go off-script, so it might have saved a little time. Between tweaking the model and generating the slides, I expect not much time was saved at all.

So what is really going on here?

I suspect it is all just a gimmick.

Let’s take a look at Klarna profit, 2012-2023:

That giant blue line is the combined loss for 2022 – about $800 Million dollars. That’s a great deal of money to lose when you’ve raised a total of $4 billion in venture funding. While the company made a net profit in Q3 2024 (the last quarter we have records for), they lost money in the period Q1-Q3 2024 (source), as well as every year prior. Bear in mind that one US Dollar is about eleven SEK.

Now let’s talk about how much cash they had in the bank. First, the liabilities keep increasing, currently 132 billion SEK, or about 12 billion US Dollars. That’s a billion. With a B. Looking at asset to liability ratio, we see the ratio of assets to liabilities declining, from 1.22 in Q3 2023, to 1.19 in Q4 2023, to 1.18 in Q3 2024.

This is a company that was on an unsustainable path. To date, they had raised $4.1Billion in funding but only 2.4Billion left in assets, and that is assuming they have 100% of the seven point five billion US Dollars the customers owe them at their disposal, right now, in cash. In reality, customers will pay back slower.

In November and December, the company closed a $1 Billion dollar (US) funding round, filed with the SEC for a 2025 IPO to raise $20 billion more, and was fined by Swedish regulators for poor risk control. Over the years, the company took funding in larger and larger amounts, eventually getting to a $41 billion USD valuation, then entering downrounds as low as 6.1 billion US.

Okay, so the company is losing money and wants to raise a lot. So what?

It’s a publicity stunt.

The AI PR Move

Most of us are familiar with the idea of return to work mandates as a shadow layoff tool. Some number of employees will either find other jobs or simply be terminated for not returning to work. Because the job is available and it was the employees choice, the company may be able to skip unemployment, offer a lower severance, perhaps none at all .. and avoid a lot of bad press. California’s WARN law has a particularly perverse incentive in requiring companies to file if they plan to do layoffs in the next sixty days, and those layoff records are public records.

Layoffs look bad.

Admitting your old-school customer service process was expensive and silly and needed an overhaul looks bad.

Admitting you are losing money and can’t afford to hire looks bad.

The alternative is to dress it up.

You don’t have a hiring freeze because you are running out of money — no, it is because you are using AI!

You aren’t stitching together a home grown  CRM system to save money — no, you are building it with AI!

Create a couple of somewhat credible demos, create some buzz on the street, and when Venture Capital won’t give you any money, just get five times as much from those suckers at home! I mean um … retail investors.

My rhetoric here is a little over the top, but you get it. It might be a trick.

Or at least, sort of a trick. In this day and age where AI truly can summarize a letter to bullet points, the claims made above can be true, in a sense. It just isn’t quite in the sense that any reasonable person hears when they hear about the sound bite. To be fair, if you take the hour to listen to the CEO of Klarna talk about their use of AI, and read their published sales data, you can get a much more well-rounded picture. But how many people did that? Why is it Matt Heusser, the software tester, the one digging into this and figuring it out?

The Bottom Line

Claims about AI magical unicorns and rainbows that lack substance beg to be investigated. What is the person really claiming? Is it true? And, if it is true, so what? How does that change things today?

Without asking these questions we fall prey to both the charlatans as well as the honestly naive people who are repeating claims, misunderstanding things, and making hopeful guesses.

Does anyone else remember XML? It was supposed to revolutionize everything. There were XML Magazines (more than one!) XML conferences, big think books on XML at your local book store. Companies raised money to do vaguely defined things with XML. People were storing data in XML. One boss advised me to look for every and any opportunity to use XML as it was “the future.”

Twenty years later, XML is just another technology. It takes up a lot of disk space, isn’t particularly fast, but it is useful for interoperability, because you can add elements and maintain backwards compatibility. It’s also easy to build and manipulate data structures in memory that are XML with code libraries, either importing them from disk or exporting them to disk. XML has a place. Yet that niche is small, and within it, frankly, JSON and even YAML have advantages.

The world got a little better, the tool found its place. The thing to look out for was the people hyping it up as perfection.

I have real concerns about the malinvestment of capital that is happening due to unrealistic claims about AI for software.

What that means for testing, I’ll leave to the next blog post.