Field notes
Working notes on digital operations, automation, and the unglamorous structure that makes transformation actually run.
The operating system behind a +492% campaign lift
How structure, GenAI content, and disciplined targeting turned one webinar campaign into a registration record, and what carried over to everything after.
Read
The port nobody touched
A firewall change was verified correctly, port by port, and took the checkout offline through a port that was never part of the change.
Your contact form sends mail
Every field that ends up in an outgoing email belongs to whoever fills it in, and a contact form is a mail server with a friendly face.
The order is the decision
A platform project moved validation ahead of the investment decision, and that single reordering changed what the business case was even for.
The layer that rewrote my evidence
A verification that runs through the same layer it's supposed to check isn't a verification, a count returned the opposite of the truth and nothing stopped to say so.
Every number went up. Nobody booked.
A campaign improved on every metric it reports, and still produced zero of the only outcome it exists for.
A rule you dictate three times is not a rule
The line between an expensive and a cheap model isn't a cost optimization, it's a judgment boundary, and a boundary that only gets spoken gets renegotiated every session.
My tests knew the language better than the platform did
A local test that evaluates an expression the way the language should work will pass the exact code the platform runs differently, so the only test that counts for a platform's quirks is one that runs on the platform.
Volume spends the market
AI outreach needs more emails and collects more spam reports per booked meeting than a person does, and in a market you can count, those are contacts you cannot write to again.
The cap was money, the damage was mail
Seven autonomous agents ran real companies for 72 hours under a spending cap. The cap held. Invoices still went out to the wrong people.
The scarcest resource is the decision-maker's attention
The most effective automation step of the quarter was switching three notifications off and taking one word out of a meeting title.
The 38.8 percent that never showed up in the lab
A benchmark built from private codebases turned a familiar success story into a number nobody expected.
The watchdog died with what it watched
An operating-system update took down a monitoring job and the job it monitored, in the same instant, for the same reason.
Whose face carries the channel
An outreach programme moved off the best-known profile in the company, and the reasoning had nothing to do with reach.
Who actually answers your call
Two unrelated allegations from the same week point at the same gap: the model named in your stack may not be the model that actually processes your data.
The default branch was an accusation
Twenty-nine people said yes, one attendance record existed, and the automation told the rest they had missed it.
Four confirmations, none of them real
On day one of running my own AI assistant, the architecture held and the confirmations lied.
The people who research you are not the people who sign
A project team recently refused a target group that looked obvious on paper — and the market data that seems to contradict them actually explains why they were right.
What self-verification cannot find
Five agent phases verified their own work end to end. The two real bugs were still found by the first human who used the thing.
Thirty agents, one branch name
Running five copies of the same model is not redundancy. It is one judgement call, repeated five times.
Five green test suites, zero deploys
I built a counter that nags me about finished systems I never shipped. It has run flawlessly for weeks and changed nothing.
Half of them went through
An import moved about half the records it should have. Nothing failed, nobody was alerted, and a person quietly fixed it by hand.
Your policy file is not a control
Frontier models pass under 25 percent of tasks when the rules sit in a long handbook, including the clauses that just say stop. Written policy is guidance. Control has to run somewhere else.
The pipeline that went dark for three weeks
A weekly automation ran every Monday without a single error and produced nothing. The thing that caught it was not a test.
The tool is not the bottleneck. The channel is.
A team rated itself 7 out of 10 on doing the right things and 4 out of 10 on doing them efficiently. None of the fixes were software.
Build vs. buy for internal ops tools: a decision framework from the field
A recurring internal process, a modest per-use SaaS bill, and hours of manual glue work in between: the build decision came down to one question, not a feature comparison.
118 clicks, zero conversions: why attribution is the first automation step, not the last
A LinkedIn outreach funnel for a DACH leadership-development firm logged 118 clicks on a webinar page and not one attributable signup, because organic and outreach traffic shared a single UTM-tagged link.
Open-weight as cost architecture: when GLM-5.2 belongs in your stack — and when it doesn't
GLM-5.2 beats GPT-5.5 on coding benchmarks at a sixth of the cost, and the labs are building their own inference chips to push token prices down further. Neither fact tells you where to route a given task. A decision framework does.
The harness beats the model
A paper on externalization in LLM agents names exactly the stack most practitioners already run, and confirms the model is the least reliable lever in it.
Structure before tooling: why most automation stalls
Teams buy the platform, then wonder why nothing runs. The order is backwards, and fixing it is mostly a question of sequence.
Deliverability is a discipline, not a setting
Inbox placement is earned over weeks, not toggled in an afternoon. A field guide to IP warming and sender reputation.
From C-level vision to a running system in 90 days
A practical cadence for turning a strategy deck into operational reality: audit, architecture, build, hand-off.
