Welcome to The Hidden Layer.

I'm Ankit Gordhandas, and at Eigenomic we build AI systems inside real companies: law firms, freight brokerages, healthcare back offices, and other similar ops-heavy businesses. Every week, this newsletter takes apart what actually happens when AI meets a P&L.

What is an AI agent?

AI agents are all the hype these days, and in talking to people, we realized that most have different understandings of what an AI agent is or does. So we published a longform essay this week that answers these questions.

The short version: Think of a music band. The model is the lead singer: brilliant, but on its own it just sits in the room and talks. The tools are its instruments. The context is the song. Permissions are the venue rules, evals are the audition, and human review is the producer. The line I'd underline: "The model is maybe ten percent of it. The other ninety percent is the tools, the context, the permissions, the evals, and the review gates."

If you read one thing before your next AI vendor meeting, make it this.

→ Read the full guide

The Teardown: Klarna hired the humans back

In 2024, Klarna's CEO told the world AI was replacing his customer service team: two-thirds of conversations handled by a chatbot, resolution times down from eleven minutes to two, hiring frozen. Fifteen months later he was on Bloomberg conceding the quality had dropped, and Klarna was hiring humans back.

Every headline called it a failure of AI. It wasn't. The technology worked, and the deployment worked on the metrics they were watching. What broke was three specific architectural layers that got cut before launch. If you're deploying AI right now, they're probably missing from your plan too.

We take the whole story apart in this week's episode, with primary sources and the customers that the dashboard never saw.

Workflow watch: client intake at a law firm

Intake is the least glamorous workflow in a law firm and the one we'd automate first.

Today it works like this: a new matter arrives by email or phone. Someone re-keys the details into the practice management system, runs a conflicts check, and waits for a partner to decide who takes it. At most mid-size firms this takes two to four days, and nobody owns it end to end.

With AI in the loop: the system reads the inbound email, extracts the parties, matter type, and urgency, drafts the conflicts-check query, pre-fills the intake record, and routes it to the right partner with a three-line summary. A human still verifies the conflicts result and still decides whether to take the engagement. Those two gates don't move.

What actually changes is response time, not headcount. Intake goes from days to hours, and the firm that responds the same afternoon wins engagements the slower firm never knew it was competing for. Speed becomes a business development weapon, which is a strange thing to say about an admin workflow.

The failure mode to design for: models will confidently mis-extract a party name. A missed conflict is malpractice, not a bug. That's why the conflicts gate stays human, and why "full automation" is the wrong goal here.

Three levels of AI adoption

Where does your company actually sit?

Level 1 is a chat assistant. Individual seats, individual productivity. Useful, and invisible on the P&L.

Level 2 is a workflow copilot. AI wired into one real workflow with a human review gate. Measured in cycle time.

Level 3 is an AI operating system. Your firm's playbook encoded into software that runs the process end to end. Measured in cost-to-serve.

One more thing

Reply with one workflow you want torn down. I pick one every issue and take it apart the way we took intake apart above.

Next issue: the freight broker that cut quoting from 20 minutes to 32 seconds and started answering the third of its inbox it never used to reach. What it means for every business that bills for time.