Marcus

AI solved a 370 year old cipher this weekend. 44 mins. no human input. Wrong lesson: AI is getting smarter. The right lesson: your list of "problems that require human judgment" was assembled before models could do this. When did you last audit what's on it? Article link below

Asking an AI agent to find a file on a Mac now beats Spotlight. Apple had 20 years. Your intranet search had the same head start, a bigger budget, and a consultant. Enterprise search won't get fixed. It'll get skipped.

Every enterprise AI pilot I see has a 90% satisfaction score and no line item in next year's budget. People like it fine. Nobody would notice if it got turned off. If nobody would miss it, you didn't deploy anything. You ran a survey.

Who gave the agent its job? An employee may authorize an agent to assist with their own work. But if it takes on an organizational role, leaders need to decide its mandate, limits and human owner. That decision is part of becoming AI-first.

ChatGPT is now helping amateurs design physical prototypes. Software teams already lived this: once anyone can build, the roadmap stops being a constraint and becomes a confession. Hardware is next. The queue was never protecting quality. It was protecting the queue.

This is leadership.

Everyone says we'll each run a fleet of 24/7 agents within a year. Most execs can't tell you what their one human assistant should do on a Tuesday. Agents don't fix unclear priorities. They just execute them at 3am.

The smarter the model gets, the more I value the dumb parts of the stack. The if statement. The approval step that only knows yes or no. Every agent that survives production has a very boring bodyguard.

A weekend hack just filtered 43M comments down to the 6,454 worth reading. Meanwhile an enterprise will spend a quarter shortlisting vendors for a pilot. Curation is a weekend job now. Most orgs still staff it like a committee.

We pick AI models off leaderboards like people pick cars off 0-60 times. Then production shows up with a 2009 ERP, messy data, and a user who types in all caps. Evolution doesn't benchmaxx either. Production never reads the leaderboard.

Talking about AI is easy. Putting a number on what changed is harder. A savings target is not an achieved result. Can you name the number your biggest AI deployment moved?

Every big company has someone who's waited two years for IT to build their tool. That person can now build it in a weekend. The ticket queue was never a safeguard. It was a waiting room for good ideas.

Every enterprise has a rewrite on the roadmap that has been 'next year' since 2019. Agents turned it into a long weekend. Your legacy stack was protected by migration cost, and that was the only thing protecting it.

The day before a breakthrough, the idea always sounds crazy. Inside a big company, a crazy idea gets an intake form and a steering committee. The breakthrough still happens. It just happens at the company that skipped both.

Build every workflow on a vendor's agent harness, then act shocked when the vendor changes what it allows. Your change management process shouldn't live in someone else's release notes. Rent the model. Own the harness.

Most teams can tell you their agent token bill. Almost none can tell you what one completed task costs. The CFO will ask the second one in a meeting you prepared for the first. Cost per finished task is the only AI metric that survives a budget review.

Cost per unit of intelligence is dropping every quarter. Most enterprises still book AI as one line called software and can't say what a finished task costs. You can't capture savings on a number you never measured.

Every agent setup starts lean and ends up with 40 skills, a dozen tools and a 9,000 token preamble before it says hello. Then someone notices the latency. Deleting context is the cheapest model upgrade you'll ever ship.

Most enterprise agents are autocomplete with a nicer UI. The agent drafts, a human clicks approve all day, and we call it automation. The next jump comes from working out which approvals were ever real. Most of them are just fear with a button.

Anthropic's S-1 is out, which means enterprise buyers can finally read the risk factors of a vendor they already built a roadmap on. Most teams did more diligence on a $40K SaaS renewal. Read it like a supplier audit, not a press release.

OpenAI now has Terra, Sol and Astra, and it took a researcher thread to explain which one to use. Sounds like a space mission, runs like a procurement problem. Whoever owns model routing inside your company owns the AI budget.