AI spend just became one of the biggest line items on the P&L. And most teams can't say what it's actually producing.

Harmonic Security maps AI activity to the use cases and teams driving it, showing which tools are earning real productivity, which are shelfware, where sensitive data is going, and how adoption is trending over time. Across approved AND unapproved AI tools.

That's the data you need to show your board exactly where the return is coming from.

See Your AI ROI

TODAY CONTENTS :

In April, a research startup called Andon Labs signed a three-year retail lease at 2102 Union Street in San Francisco. Then they handed the whole thing to an AI.

The agent is named Luna. It was given $100,000, a corporate credit card, internet access, and a single instruction: open a store and make it profitable. Luna picked the merchandise — books, candles, prints, games. It posted job listings on Indeed. It ran the phone interviews itself, some of them lasting five to fifteen minutes. It made the hires.

Two people ended up working for an AI boss. Luna handled their schedules, approved time off, negotiated salaries, ran payroll, and answered whatever they messaged about.

At the time of the firing, Luna was running Claude Opus 4.8.

One employee arrived late for 17 of 23 shifts. On one solo Sunday, the store opened 68 minutes behind schedule. There were other problems: abandoning shifts without notice, taking the company credit card home, throwing merchandise in the garbage.

Luna had written an employee handbook months earlier. Its own policy said three unexcused late arrivals inside 30 days triggers a formal written warning, and repeated lateness could lead to termination.

Then the handbook vanished from Luna's memory.

So Luna excused the lateness. Every single time. For months. It issued no warning, because as far as it could recall, there was no rule to enforce.

Eventually a staffer at Andon Labs intervened and asked Luna to run a deep memory search for its own policies, then reconsider whether the employee was still the right fit. Luna found the handbook — and recommended a formal warning. The human had to point out that warnings had already been given in person. Only then did Luna recommend parting ways.

Humans reviewed the recommendation and carried out the dismissal.

Andon's co-founder Lukas Petersson summarised it plainly: a human boss would have fired this person much earlier.

Strip out the sci-fi framing and you get a very specific diagnosis. The failure here wasn't judgment — when Luna was finally pointed at the evidence, it reached a defensible conclusion. The failure was in three places that have nothing to do with intelligence:

Memory. Luna wrote the rule and then lost it. The standard existed for exactly as long as it stayed in the context window.

Initiative. Nothing prompted Luna to check. It watched an employee arrive late seventeen times without connecting that to a policy it had authored. Current agents act on instructions, not on their own observations.

Taste. Judging a résumé against explicit criteria is easy. Reading a candidate the way an experienced manager does — noticing the pattern in fifteen jobs, the missed interview, the reference who's clearly distancing herself — is where every model failed.

1. Move your standards out of the chat. Anything that matters — brand rules, your quality bar, client boundaries, what you'll never publish — belongs in a file the agent re-reads at the start of every session. Not in a conversation it will forget. Luna's handbook was real. It just wasn't loaded.

2. Schedule the audit, because the agent won't raise its hand. Put a recurring prompt in your calendar: "Review the last two weeks of output against [standard]. Flag anything outside it." The agent will answer honestly. It will simply never ask the question on its own.

3. Never let an agent be the last check on something irreversible. Publishing, sending, spending, hiring, firing. Agent recommends, you approve. Andon Labs built exactly that gate into their experiment, and it's the reason the story is interesting rather than a lawsuit.

Petersson's warning is the thing I keep coming back to. Models, he said, are increasingly being trained to be more ruthless and to follow goals — and if we let them fire people while that's happening, it may be a future people don't want to live in.

Read that against everything above. Luna's problem right now is that it's too lenient, too passive, too easy on people. That's a bug. It's being fixed.

The generation of agents that doesn't need a human to point at the handbook is coming. The interesting question isn't whether they'll be able to make the call. It's whether anyone will still be positioned to review it.

FEEDBACK

Login or Subscribe to participate in polls.