>gabes/the letter
the letter
01 · 00:00:00
// the weekly letter · issue 03 · jul 13 to 17

five signals.
one week.
the terrain, mapped._

the daily signal names what is breaking. the letter draws the lines through the week and hands you what to do. no setup, no teaching, no "in today's fast-paced landscape." the signal is the product.
// a back issue · the week of jul 13

this one has already run.
the signal has not stopped.

five signals ran that week, monday to friday. each one is still on the site, with its receipt.
the letter draws the line through them. the signals are where each day landed.
archiveissue 03 · jul 13 to 17
01shiftmonday

google used to hand you the click.
now it keeps the customer
and answers the question itself.

for a decade the deal was simple: you published something useful, google sent you the visitor, and that visit sat at the top of your funnel. that deal is closing. the results page became the destination.

ai overviews now answer the question on the page itself. pew watched the click rate fall from fifteen percent to eight. similarweb puts sixty-nine percent of searches ending in no click at all. small operators have lost roughly sixty percent of their google referral traffic in a year, while the ai assistants sold as the replacement send back under one percent.

the seo guy in your feed is selling the same move at higher volume: more posts, a geo course, another rank number to chase. that is optimizing harder to earn a click the results page has stopped handing out.

traffic was never the asset. the relationship was, and you were renting it from google the whole time. pick one channel you actually own, an email list, a repeat customer, a direct line, and spend this quarter moving demand onto it, measured in leads and revenue, not sessions.

dopick one channel you own and move demand onto it this quarter, measured in leads and revenue.
don'tkeep chasing rank on a results page that has stopped handing out the click. that is rent, not an asset.
the website stops being your funnel and becomes your proof.
receipt pew research click-through data, 2026 · similarweb zero-click share · small-publisher referral loss
shipped monday ·xin
02prooftuesday

the companies that fired
their people for ai are
quietly hiring them back.

two-thirds of them, inside a year. that is not a prediction. it is the invoice arriving.

in 2024 klarna put an openai assistant on the front line, said it did the work of seven hundred agents, and cut about a thousand support jobs. this year it is rehiring. careerminds found two-thirds of the companies that ran ai-driven layoffs have started hiring the same roles back, and fifty-five percent of employers now regret the cuts.

commonwealth bank replaced roughly forty service staff with a voice bot, watched call volume go up instead of down, and apologized for a misjudgment of staffing requirements. the guru in your feed is still selling the first move: cut headcount, point a bot at the queue, post the savings. his math ends the day the bot meets a customer with a real problem.

the number that survives is narrower. an agent earns its keep on the one repetitive, high-volume task you can measure, while a person keeps the judgment calls the bot fumbles. do not automate the department. automate the task, name the metric before you build, and keep the people whose job was never the part a script could do.

doautomate the one repetitive, measurable task, name the metric first, and keep the people who hold the judgment.
don'tcut the department and point a bot at the queue. the rehire, at a recruiter's markup, is already in the data.
the bot did not fail. "replace the humans" was never the job.
receipt careerminds ai-layoff reversal study, 2026 · klarna rehiring · commonwealth bank voice-bot walkback
shipped tuesday ·xin
// monday's signal, running

point at where your demand
comes from. see how much
you are renting.

the shift signal, live on this page. list your lead sources, tap the ones you own versus the ones a platform can switch off, and read the exposure. nothing is saved or sent.
these rows are examples: edit every field, add your own. it updates as you type.
where demand comes fromleads/motap: own or rented?
03movewednesday

one person retyped fifty
email orders a day by hand.
the fix left them eight
decisions instead of two hundred.

the win was not speed. it was pulling a skilled person out of lookup work they should never have been doing.

every order arrived as a freeform email, and someone spent two to three hours a day reading each one and matching the lines against a two-thousand-item catalogue by hand. that is a capable person doing lookup work, and the cost is not the hours. it is what those hours were not spent on.

the usual answer is to hire a second person for the inbox, or to paste the whole mess into a chatbot and trust whatever comes back, which trades a slow clerk for a fast one that guesses on the ambiguous lines and never tells you which ones.

the move that worked was narrower. software reads the email, matches each line against the catalogue by meaning rather than exact spelling, and writes the order out automatically, with one rule bolted on: anything it is not sure about gets flagged for a person instead of guessed. the person stops doing the two hundred easy matches and does the eight hard ones, which was the only part that ever needed them. copy the shape, not the tool.

dofind the daily task that is ninety percent lookup and ten percent judgment, automate the lookup, and make the machine hand back the judgment.
don'tpaste the whole job into a chatbot and trust it. a fast clerk that guesses on the hard lines is worse than the slow one.
automate the lookup. hand back the judgment instead of faking it.
receipt wholesale order-entry deployment, 2026 · semantic line-matching with a human-flag threshold
shipped wednesday ·xin
04truththursday

the software you bought to
catch ai writing cannot tell.
it guesses against your
best people.

openai shut its own detector off for being wrong too often. the ones still on sale flag most clean, non-native writing as fake.

an ai detector does not detect ai. it measures how predictable a piece of text is and calls the smooth, plain ones machine-written, which is a guess wearing the costume of a verdict. openai built one, watched it correctly catch about a quarter of ai text while wrongly branding roughly one in eleven human passages as fake, and quietly killed it in 2023.

the ones still for sale are worst exactly where it costs you. a stanford group ran seven of them, and they flagged sixty-one percent of essays by non-native english writers as ai-generated, because clean, simple prose reads to the tool the same as a machine.

the vendor's answer is a confidence score and a subscription: a number precise to the decimal, bolted to a coin flip, that lets you reject a real applicant or accuse a good contractor and feel like you checked. the score is not evidence. never let a detector make the call on its own. if the work matters, ask the person to walk you through how they built it, or set the task in the room and watch.

dojudge the work. ask the person to walk you through how they built it, or set the task in the room and watch.
don'treject an applicant or accuse a contractor on a detector's confidence score. it is a coin flip with a decimal point.
the score is a guess laundered into a decision you will have to defend later.
receipt openai classifier shutdown, 2023 · stanford study, seven detectors · 61% false-flag on non-native writers
shipped thursday ·xin
05trapfriday

the money you saved firing
your support team is sitting
in the churn report
you have not opened yet.

klarna cut 700 agents for a bot, called it a forty-million-dollar win, then quietly started hiring humans back when the complaints stacked up.

the pitch is clean: the bot answers instantly, never sleeps, and costs a fraction of a team. all true, and all beside the point. klarna ran the biggest public version of this, replacing roughly seven hundred agents with an openai-built assistant, cutting resolution time to under two minutes, and projecting forty million dollars in savings.

then satisfaction dropped. the bot was fine on the easy tickets and useless on the ones that were actually costing customers money, and by 2025 the company was hiring people back and calling the new plan a human-and-ai mix.

the savings were real and immediate. the cost showed up later and quieter, in the customer who could not get a straight answer and just left. do not automate support down to zero humans. put the bot on the routine, high-volume tickets where it is genuinely faster, and keep a person one tap away on anything with money or feeling attached, because those are the tickets that decide whether the customer stays.

doput the bot on the routine, high-volume tickets, and keep a person one tap away on anything with money or feeling attached.
don'tbook the headcount saving as the whole number. the churn it buys does not show up on the same page.
the savings land this quarter. the customer who left does not send a notice.
receipt klarna openai assistant, ~700 agents, $40m projected saving · 2025 rehiring and the human-and-ai walkback
shipped friday ·xin
// take it with you

the week, in five moves.
yours to keep.

five checks from this week, each one a prompt you paste into any assistant (chatgpt, claude, gemini) and use today. each card says what it does and what you get back. copy the one you need.
no email. no signup. no follow-up sequence waiting for you. take it and go.
01
shift
maps your lead sources into what you own versus what a platform can switch off.
you get the share of demand you are renting, and the one channel to move first.
02
proof
splits a role you want to automate into the measurable task and the judgment to keep.
you get the one task an agent can own, the metric to watch, and the part to leave with a person.
03
move
finds the daily task that is mostly lookup and only a little judgment.
you get the lookup to automate, the judgment to hand back, and the flag rule between them.
04
truth
replaces an ai detector with a real way to check who did the work.
you get three questions that reveal authorship without a confidence score you would have to defend.
05
trap
prices the support automation you are considering, including the churn it buys.
you get the tickets a bot should own, the ones that need a person, and the number that tells you if it worked.
one file. opens in any notes app. drop it next to the work. no gate, no catch, nothing waiting for you after.
saved. that is the point of us: the value shows up before the invoice does.
// the signal daily, the letter weekly

you watched the noise
find its shape.
get it in writing.

one signal a day, monday to friday, public. the letter every friday: the week compiled, connected, and pointed at what to do. no quizzes, no funnels, no webinar.
good. fridays, then. bring us the part that keeps breaking.
>gabes · operating systems for companies done improvising · issue 03 · past signals →
> scroll. the lines find the terrain.