Five things worth knowing this morning.
Including the one I got wrong.
Every weekday, a short brief assembled from your own
data — what changed, what's drifting, what needs a
person today. Each item shows the evidence behind it.
And it publishes its own hit rate, because an
advisor that never admits a miss isn't an
advisor.
It does not give you business strategy. It
tells you facts about your own operation that you
would have found eventually, sooner and with the
working shown. Everything else on this page follows
from that limit.
Tuesday briefNorthgate Supply · 06:40 localSample company · not live data
Quiet week so far, with one exception.
Nothing in cash or fulfilment needs you today. The
billing thread below does, and it's been building
since Thursday.
Needs a person todayOwner: Priya N.
Billing complaints are up 4.6× and all
trace to one invoice batch
Since Thursday,
218 billing tickets
against a baseline of 47. Every one
references an invoice from the
14 Sept batch. Same
wording in 190 of them: customers
charged for a plan tier they say they
didn't select. No other ticket category
is unusual.
Suggested: pull the 14 Sept
batch and check the tier assignment
before more invoices go out on the
28th. I can't tell you whether the
batch is wrong — only that 190
customers describe the same
thing.
Three of your ten largest accounts have
gone quiet
No inbound contact from
Halloway, Fenn Group or
Ardent
in 41, 38 and 36 days respectively.
Their normal interval is 9–14 days. All
three renew in Q1. Two had support
tickets closed as resolved in
the last quiet window.
Suggested: a check-in call
before the renewal conversation
starts. Silence isn't churn — it's
just not information, and these
three are worth having information
about.
Warehouse pick times held through the
volume increase
Order volume up
23% against last month;
average pick time unchanged at
7.2 minutes. The
September layout change looks like it
absorbed the growth. First month it's
been tested at this volume.
Suggested: nothing. Included
because someone made a decision in
September and it worked, and that
usually goes unremarked.
Evidence
WMS — pick time logsOrder volume, 60 days
Confidence0.89
AcknowledgeNot useful
I was wrong
Last Wednesday I flagged a supplier
delay risk. It didn't happen.
I flagged
Kestrel Components as a
delivery risk based on three late
shipments in a row. All subsequent
deliveries were on time. Looking again,
the three late ones shared a single
cause — a public holiday in their region
— which I treated as a trend rather than
a one-off.
Adjusted: regional holiday
calendars now excluded from
delay-pattern detection. This one
cost your ops lead about twenty
minutes of unnecessary chasing, and
it's on the scorecard below.
Evidence
Delivery records — 11
shipmentsRegional holiday calendar
Was0.66
Noted
Can't answer
You asked whether to raise prices in the
Nordics. I'm not going to tell you.
That's a strategy decision requiring
competitive positioning, customer
willingness to pay and appetite for risk
— none of which is in your systems. What
I can put in front of you:
Nordic gross margin is
4.1 points below your
average, and three of the last five
Nordic deals closed at a discount above
20%.
Suggested: those two facts
are the input to your decision, not
the decision. Anything that told you
to raise prices from this data alone
would be guessing in a confident
voice.
Evidence
Margin by region — definition
layerCRM — 5 Nordic deals
On the facts0.92
Noted
Five items, two of which are not advice. One
admits a mistake, one refuses the question asked. A
brief where every item is a confident recommendation
is a brief nobody should trust by the third week.
The scorecard
It grades itself, monthly, and you see the grade
Every flagged item is checked afterwards against
what actually happened. Not by me — automatically,
against the outcome in your own systems, with the
ones nobody acted on marked as unknown rather than
quietly dropped.
94
Items raised in the last 90 days
71%
Confirmed useful by the owner who acted on
them
9
Wrong, retracted and explained
17
Never acted on — outcome unknown, counted as
neither
Kestrel Components — false delivery
risk18 Sep
Cause: a regional holiday read as a
supplier trend. Detection now excludes
holiday calendars. Cost: roughly twenty
minutes of chasing.
Flagged a churn risk that was a billing
address change2 Sep
Cause: account activity dropped
because the account was being migrated
internally, which looks identical to
disengagement from the outside. Now
cross-checked against account-change
records.
Missed a stock issue entirely for four
days21 Aug
Cause: the warehouse feed had been
failing silently since the 17th and nothing
checked that the source was alive.
The worst kind of miss — no signal at
all rather than a wrong one.
Feed health monitoring added.
A 71% useful rate is not a marketing number. It's
what an honest brief looks like — and if a vendor
shows you 98%, ask what they counted.
Design limits
What it refuses to do, on purpose
This category is full of products that promise a machine
will run your business. The limits below are what
separate a useful brief from an expensive horoscope.
01
No strategy advice
Pricing, hiring, market entry, whether to fire
someone. These need context that isn't in any
system, and a confident answer from data alone
is worse than silence because it sounds
researched.
02
No causation claims
It reports that billing tickets rose 4.6× and
that they reference one batch. It won't tell you
the batch caused it — that's a human checking a
hypothesis, and the difference matters
enormously.
03
Nothing acts by itself
No emails sent, no records changed, no refunds
issued. It observes and suggests; a named person
decides. Automation is a separate conversation
with a separate risk profile.
04
Silence when there's nothing
A quiet week produces a short brief, sometimes
two items. Manufacturing five insights a day to
justify a subscription is exactly how these
tools train people to stop reading them.
05
Never about individuals
It watches processes and accounts, never
people's performance. The moment it becomes a
surveillance tool, the humans feeding it start
behaving differently and the data stops being
true.
06
It says when it can't see
A dead data feed produces a visible gap, not a
quieter brief. The August miss on the scorecard
is exactly this failure, which is why feed
health is now monitored as carefully as the data
itself.
Prerequisites
This is the last thing to build, not the first
A daily brief is only as good as the definitions
underneath it. If nobody agrees what "active account"
means, the brief will be confidently wrong every morning
— which is worse than no brief.
Agreed definitions for your core metrics
What counts as an active customer, a late
delivery, a churn risk, a resolved ticket.
Written down once, owned by someone. Without
this the brief is generating opinions, not
observations.
At least three connected systems with real
history
Twelve months minimum. Patterns need a
baseline, and "unusual" is meaningless until
you know what usual looked like across a
full seasonal cycle.
Named owners for the areas it watches
Every item needs somebody it belongs to. A
brief that arrives to everybody arrives to
nobody, and that's the most common way these
die in month two.
Someone who will mark items useful or not
Thirty seconds a day. Without that feedback
there's no scorecard, no tuning, and within
a month it's noise that everyone archives
unread.
If you're missing the first one, start there instead — a
definition layer is worth building on its own, and it
makes this cheaper and better later. I'd rather sell you
that first and this second.
How we'd work
Read-only for a month before anyone gets a brief
The first month exists to find out what this system
would have said about things that already happened —
which is the only honest way to know whether it's worth
reading.
Weeks 1–3
Readiness and definitions
Which systems, what history exists, and what
your core terms actually mean. Sometimes this
ends with "build the definition layer first" —
and I'll say so.
Weeks 4–7
Backtest in silence
Run it over the last six months and see what it
would have flagged. You judge those against what
really happened, before anyone receives a live
brief.
Weeks 8–10
One team, live
The brief goes to a handful of named owners.
Feedback tunes the thresholds — mostly downward,
because the first version always flags too much.
Weeks 11–12
Scorecard and handover
Self-grading turned on, runbook written, and
your team trained to add new checks without me.
Pricing
Published, and the first step can end in "not yet"
Start hereReadiness check
Do you have the data, definitions and owners for
this to work?
The one number that decides whether to keep paying
After ninety days live, we look at the
useful rate — the proportion of items the
owner marked as worth their attention. If it's below
50%, the brief is noise and you should stop.
I'll say that out loud rather than let a
subscription drift.
Everything is built to make that number honest:
items nobody acted on are counted as unknown rather
than as successes, and every retraction stays on the
scorecard permanently.
Model and infrastructure costs sit in your own
accounts, typically $120–$450 a month. Not marked
up.
Questions
The sceptical ones, which are the right ones
How can it possibly know enough about my
business to advise me?
It can't, and it doesn't try. It knows
what's in your systems and what normal looks
like there. That's enough to notice that
billing tickets are 4.6× baseline and all
reference one invoice batch — which is a
fact, not advice.
Everything requiring judgement about your
market, your people or your strategy is
explicitly out of scope, and the brief says
so when you ask. The fifth card in the
sample is that refusal.
How is this different from alerts and dashboards
we already have?
Threshold alerts fire when one number
crosses a line, which produces either noise
or silence and rarely the thing you needed.
Dashboards require you to go and look, and
to already suspect what you're looking for.
The difference here is combination and
context: patterns across several systems,
compared to your own history, ranked by
whether a person needs to act today, and
delivered without you asking. The honest
overlap is real though — if your existing
alerting is good and people trust it, this
is a smaller improvement than the price
suggests.
Won't people just start ignoring it?
That's the default outcome for this whole
product category, and the design fights it
in three specific ways: it stays silent when
there's nothing worth saying, every item has
one named owner rather than going to a
distribution list, and the useful rate is
measured so decay is visible before it's
terminal.
If the useful rate drops below 50% at ninety
days, the honest answer is to stop. That's
in the pricing section for a reason.
What stops it telling us something confidently
wrong?
Every item shows its evidence and a
confidence score, and low-confidence items
are framed as questions rather than
findings. It reports what changed, not why —
no causation claims, because that's where
these systems usually embarrass themselves.
Then the scorecard: wrong items are
retracted publicly, with the cause explained
and the detection adjusted. Three of those
are on this page. A tool that hides its
misses is a tool you'll eventually stop
believing all at once.
Is this watching our staff?
No, and I won't build the version that does.
It watches processes, accounts and volumes —
never individual performance, never who said
what. If leadership's actual goal is
monitoring people, this is the wrong project
and I'd decline it.
Practical reason as much as an ethical one:
the moment people believe they're being
watched through it, the data feeding it
stops reflecting reality.
What if our data isn't good enough?
Then the readiness check says so in two
weeks, and you've spent $5,000 instead of
$40,000. It happens often — usually missing
definitions, or not enough history for a
baseline.
In that case the definition layer is the
thing to build first, and it's useful on its
own regardless of whether you ever come back
to this.
Who builds it
I've spent eight years on alerting that had to
be worth reading at 3 a.m.
I'm Bahman Shadmehr. Automated trading platforms
on the Texas wholesale power market and US
equities, a national payment gateway I helped
decompose while live, and the monitoring
underneath all of it.
Those systems taught the lesson this product is
built on:
an alert that fires too often is worse than
no alert, because people learn to dismiss it and then
miss the real one. Tuning thresholds so a page
means something, and measuring whether it did,
is most of what that work actually was.
A daily brief is the same discipline aimed at a
business instead of a pipeline. That's why the
scorecard exists, and why the limits section is
longer than the features.
2024 — 2025Automated energy trading, ERCOT
— health metrics and anomaly alerting
across the pipeline
2023 — 2024US equities trading platform —
anomaly detection on daily trading
jobs
Not a hypothetical — an actual one from this
year. I'll tell you within a day whether a brief
like this would have caught it, and I'll tell
you when it wouldn't have.