thinking
essays on ai strategy, governance, judgment, and craft.
latest
software that listens at work is switched on by one person, and it leaves a record that everyone else in the room has to live with. in a drug company, that record can carry a reporting duty. the rules for it have to be written down, because no settings screen will write them.
in a regulated company, more ai projects stall on where the data goes than on what the model can do. a model that runs on your own hardware removes that transfer for a class of work. when the job is still too big for the device, the useful question is what you can check about where it lands.
a performance review is one of the few places a person learns, in writing, where they stand. the judgment in it has to come from the manager who signs it, and when a tool shaped the wording, the employee should hear that from the manager too.
the on-label story is still locked to brand sites and review cycles. a chatbot often answers before anyone opens a brand url. the gap is getting verified words into that first reply, in language a person can use.
american businesses spent the summer moving everyday ai work onto cheaper models, and told the company that counts their spend that the work held up. the habit is one you can copy.
the head of alignment at anthropic put a number on the end of the world, in a reply, with no analysis attached. the number cannot be checked, and it has the same shape as the benchmark score in the vendor deck on your desk. 4 questions work on both.
a late software program was usually decided in its first weeks, before anyone agreed who it was for. here are the 4 things to ask for before you commit, and what skipping each one costs.
ai agents read your website so buyers do not have to, and what they can read decides what they say about you. that leaves one decision to make, page by page: free, metered, or blocked. for nearly everything, the answer is free.
gartner thinks 40% of these projects get canceled. the failures teach more than the wins, because they show you what the tools were amplifying.
the companies building the biggest ai models are in a race to grow, and this summer one of them lost control of its own agents for 2 months, then called the result intelligence. what happened, in plain words, and what it means for anyone being sold an ai score.
regulation and case law are closing the gap between running a model and owning what it produces. the enablement budget usually covered only the first half.
the last study in this series is also the oldest. 1989, two psychologists at michigan, and a finding about attention that has held up ever since.
read claire stapleton's memoir of google next to 4 years of ai coverage and the august 2026 deepmind reshuffle: the gemini meltdowns stop looking like a bug and start looking like a self-portrait.
risk management, record-keeping, logs, and oversight a person can actually exercise are properties of a running system. what a regulated build needs from day one, and why adding it at the end does not survive review.
the quiet weeks are the only forcing function the calendar hands you for free. most teams wait for september instead.
derek thompson coined vicemaxxing for american politics. the same gas-pedal pattern shows up in ai spend, inbox workslop, and the scroll, and the way out is deciding what to refuse.
why every retool app you've used was exhausting, and who's really paying for it.
notion gets adopted. the harder failure shows up after everyone logs in, when the workspace fills and the shared map never does.
iheartmedia and apple tv+ started labelling work as human-made. the label is a quality signal aimed at sameness, not a crusade against the tools.
ai shortened the gap between a hunch and something you can look at. who decides which hunch was worth having is still a human call.
linear answers every adoption complaint ever made about enterprise software, and teams still fail to adopt it. third in a series on rollouts, failing for the same reasons as the crms: missing transparency, missing plan.
design teams spend a week producing artifacts the user never sees, then ship one screen. process was meant to help teams avoid dumb mistakes, and somewhere it became something to stay faithful to.
an mit media lab study found the chatgpt group's brain activity was still sluggish weeks later, writing without the tool. we use ai every day, which is why the finding belongs in how we work.
attio is light and hubspot is dense, and the teams that land either one did the same 3 things before go-live. second in a series on what makes a rollout stick.
a summer series on the gap between how fast the ai tools move and how fast people actually work, and what the research says about who burns out first.
hubspot works for the revenue org it was built for. the failure is buying a platform of that weight for a team that is a different shape.
ai made the working prototype an afternoon's work, and the volume of mediocre software has never been higher. what is scarce now is someone with the taste and the standing to decide what deserves to exist.
the tool gets blamed in the board meeting. adoption usually failed earlier, when the org skipped whether the product fit how the team already works.
when the next module arrives, most saas companies stuff it into the product everyone already uses, because attach rate rewards it. the bill comes later, in onboarding time and a product that no longer fits in anyone's head.
koivisto and grassini's 2023 creativity study found the chatbots beat the typical person and never beat the best one. volume got cheap; the top of the range did not.
a year of product design has flattened into the same gradients and the same card layouts. when the floor rises for everyone, the ceiling is the only thing left to compete on.
the pilot works, the vendor gets picked, and then someone asks who owns the rollout. nobody answering that question is most of why enterprise ai stalls.