AI News That Actually Matters: Cheap Agents, Expensive Decisions

Klarna Replaced 800 People With Agents and Disclosed It. The Reckoning Has a Name Now.

I’m not going to dodge this one, because it challenges my thesis in the cleanest possible way. My view has been consistent: agents extend people more than they replace them. Klarna just replaced them. About 800 of them, concentrated in entry-level support and credit-risk workflows, and the company didn’t whisper it through a leak. It disclosed it. That matters. When a company says the quiet part out loud, you should assume it believes the market will reward the honesty.

My honest read is that Klarna wasn’t treating those roles as a talent strategy. It was treating them as a cost center. And the specific work described — routing, logging, first-line scripted support, repetitive assessments — was always sitting in the blast radius. That’s not agents replacing deep human judgment. That’s agents replacing human clipboard-holding. Harsh, but accurate.

But there’s a second page to this story, and too many people are pretending it isn’t there. Remaining staff reportedly saw higher workloads while learning to collaborate with the agents. That is the extension model, and it’s much messier than the replacement model. Once you automate the obvious front-end work, the humans left behind inherit the exceptions, edge cases, escalations, and oversight. The work gets more valuable and more cognitively demanding at the same time. So yes, Klarna displaced people. It also created a harder operating environment for the people who stayed.

That’s why I don’t think the takeaway is "AI replaces people, full stop." The real takeaway is uglier and more useful: without leadership intention, agents default to labor arbitrage. With leadership intention, they can become leverage. Most companies have not explicitly decided which game they’re playing. Klarna did. If you run a business, don’t fool yourself into thinking the technology will make that values decision for you.

The “AI Is Too Expensive” Objection Died This Week. Please Update Your Excuse Inventory.

I’ve been tracking this one for months because it maps directly to a question I care about: what constraints disappeared this week? This week, three major providers answered in unison.

OpenAI cut GPT-4o mini inference pricing by 40% and raised quota ceilings. Anthropic reduced Claude 3.5 Sonnet pricing versus March and expanded limits. Google launched Gemini 3.1 Flash-Lite at pricing that would have sounded unrealistic not long ago. Three different companies. Same direction. Same week. That is not noise. That is a market signal.

I documented back in October 2025 what it took to run a sales intelligence system across multiple models: over 100 nodes, 37-minute runs, rate-limit collisions all over the place, and API costs high enough to make a CFO start asking whether this was experimentation or self-harm. The difference between then and now is not cosmetic. It is structural. The set of workflows that are economically reasonable to build has expanded fast.

So let me say this plainly for the business owners in the back: the cost objection is over. Complexity still exists. Governance still exists. Bad implementation still exists. But “we’d love to explore agents, it’s just too expensive right now” is no longer a serious statement. It’s a stall.

And that has a second-order effect that I think matters even more. A whole class of enterprise vendors has been selling the idea that AI is too complex and too expensive for normal companies to touch without a paid escort. That moat is draining in real time. If their pitch starts with fear and ends with a markup, I’d reevaluate the pitch. The new question is not whether you can afford to test this. It’s whether you’re willing to let somebody faster turn your hesitation into their operating advantage.

Small Businesses Are Running Agents in Production and Have the Receipts. Here’s What the Numbers Actually Say.

This is my favorite kind of evidence because it doesn’t come wrapped in a giant brand logo and a choreographed keynote. A manufacturing shop improves lead-to-customer conversion from 8% to 12% and cuts the sales cycle by roughly 40% with a sales agent. A mid-sized SaaS company cuts inbound tickets by 35%, lifts NPS by 15 points, and reduces support headcount needs by one-third. A boutique e-commerce retailer cuts stock-outs in half, improves inventory turns from 4.2 to 5.6, and sees 12% monthly revenue lift.

That is the extension model in the wild.

This is the deployment pattern I’ve been predicting since I started writing this column: not spectacular humanoid theater, not agents replacing entire companies, but small and mid-sized businesses using software labor to operate above their weight class. A small team converting more leads with the same headcount is not a job-loss story. It’s a leverage story. A retailer preventing stock-outs without hiring a full-time analyst is not a gimmick. It’s an operating upgrade.

The inventory example is the one I keep coming back to because it captures what a lot of people still miss. AI is not just a feature you add to the website so you can say you did AI. In the right deployment, it becomes an operating system for decisions your business was already making badly, slowly, or inconsistently. If you’re a business owner, that should get your attention more than any frontier-model benchmark ever will.

I’ve said from the beginning that this column is for the businesses that don’t have a PR machine but do have payroll, margins, and very real constraints. These are the companies I care about because they’re the ones proving that agents can create practical leverage before the Fortune 500 finishes arguing about governance terminology.

Nobody Knows Who Owns Your AI Agents’ Access. That’s Not a Future Problem — It’s a Current One.

I called this in February, on record: Agent Identity Management was going to become a real topic in 2026 before most vendors had an adult solution for it. That call is aging well, unfortunately.

The number that should bother you is not just that only 57% of surveyed firms felt confident their agents were properly scoped. It’s the existence of the phrase “shadow AI agents.” Once the security industry names a category, the category is real enough to have become somebody’s problem. And this category means agents are already being created inside organizations without clear authorization, visibility, or auditability.

No, there hasn’t been a headline-grabbing public breach tied to unauthorized agent access in the last 30 days. That does not make this safe. It makes it early. “No breach yet” is the corporate equivalent of saying the stove hasn’t caused a fire yet while the burner is still on.

Here’s why this matters to my thesis. I believe in the extension model. But extensions require boundaries. If an agent can access systems, act on behalf of staff, move data, or trigger workflows, then “who gave it access to what, and how do we revoke it?” stops being an IT hygiene question and becomes a board-level operating risk question. Right now, a lot of companies do not have a crisp answer. Some don’t even have an inventory.

That’s the part I want business owners to hear: you do not need to become paranoid to become serious. But you do need to stop treating access governance as something you’ll clean up after the pilot. If agents are already touching production systems, then identity is not a future layer. It is part of the product.

The Multi-Agent Complexity Pitch Is Getting Louder — and That’s Exactly What Railroad Companies Do Before the Planes Come

I want to separate signal from vendor theater here. There’s a lot of increasingly confident talk about hierarchical orchestrators, event-driven reactors, state-graph workflows, and multi-agent systems outperforming single-agent approaches by 3–5x. Some of that may prove true in specific environments. Some of it is also being pushed by people who benefit when you conclude the architecture is too complicated to touch without them.

What do I trust more? The adoption data. Pilot-to-production conversion for enterprise agentic AI projects nearly doubled to 31% in Q2. MCP server registrations jumped 58% quarter over quarter. No-code and low-code keeps climbing toward a market large enough that nobody gets to dismiss it as toy infrastructure anymore.

That combination tells me something simple: the market is shipping before it is philosophizing. Companies are not waiting for perfect orchestration doctrine. They are putting simpler agents into production, learning where they break, and expanding from there. That is exactly how real technology adoption works.

So my current read is this: the integration layer is winning faster than the orchestration layer. MCP growth matters more to me right now than architecture arguments on Medium, especially when the strongest performance claims still aren’t paired with named enterprise case studies. I’m watching what survives in production six months from now, not what sounds smartest in a diagram today.

And this fits a prediction I made earlier this year: the camps of agentic development were going to become visible and debated. Here they are. DIY builders, platform buyers, orchestration maximalists, no-code pragmatists. Good. Let them argue. The market will decide based on what gets deployed, maintained, and paid for — not what wins the whiteboard battle.