Nine years running B2B demand-gen delivery by hand — lists, asset QA, returns, pacing. Two years turning each manual loop into a system that the same team now runs without me.
Twenty to forty hours a week, one client at a time. Remote from Germany, on site in the Netherlands when it helps.
Counted from production repositories, September 2026. No client names, no client data.
Somewhere on your campaign team there is a spreadsheet that turned into software about a year ago. It has formulas only one person can read, and it decides which leads go back to the client. Someone else pastes client data into a chatbot every afternoon because it saves two hours. Neither of them is doing anything wrong. They found the fastest route to the work.
What they do not have is the second half: a written standard, a review before anything touches a client deliverable, a record of what was decided and why. Tool vendors sell the first half — your people already found it without you. The second half is not a product. It is a discipline, and somebody has to install it.
Context: 81% of employees use AI tools their employer has not approved — Retool, State of AI Governance 2026 report.
The numbers at the top are the ones a managing director cares about. These are the ones that tell you the systems keep working when I am not looking.
Each one replaced a separate manual loop inside a B2B demand-gen operation — the same loops your delivery team runs for clients every week. All seven are in production, listed here in the order the work flows through them.
Client campaign criteria in, normalised targeting out — geography, industry, job function and level, company size, suppression references — resolved against a taxonomy and handed to everything downstream, with a review screen where a manager corrects the interpretation inline.
A value with no match in the taxonomy becomes a flag to verify, never a silent substitution and never a silent drop.
Twelve-tab workbooks and suppression files in; one verified, deduplicated account list out, with the origin of every row kept.
When a name could mean several companies it writes no domain at all, and says which ones it could not choose between.
The accounts automatic matching cannot resolve get researched against real sources and answered with proof links.
A domain the model did not actually find in the search results is thrown away as fabricated.
Checks whether the asset actually matches what the targeting promised, before the campaign runs rather than after the rejection.
The client's criteria are authoritative. The asset may flag a misfit; it may never overrule them.
Browser agents fetch gated assets and attribute each one to the contract that paid for it.
Attribution is the product. Each job gets its own address, and the match is on the full tag.
Returns, quality checks, lead verification and campaign pacing in one place, with the reason behind every rejection, fed from the spreadsheets, mailboxes and CRM the data used to be scattered across.
A manual correction survives the next sync. The source systems never overwrite a person's fix.
Every task and every change across all of it, on one board that outlives whoever was on shift.
The journal is append-only, and it can never break the operation it records.
Built inside an operation of about a thousand people. Seven systems, each built alone; an eighth was built with a team and is not in this list. They plug into what you already run — your CRM, your ESP, your spreadsheets, your portal. The systems are under NDA. The mechanism is not.
No platform rollout, no pilot committee. One process that costs you money, removed first. The weeks below assume twenty hours a week; at forty they halve.
I follow the work, not the org chart. We pick the first process by what it costs you now, and write down the number it will be judged by.
One manual process, replaced in production, measured before and after. The people who used to do it by hand are the ones running it afterwards. That is the test of whether it was built right.
How that system was built gets written down: template, review rubric, who approves what, what a tool must prove before it ships. Tool number two is built against the standard instead of from scratch.
Every change is reviewed against the standard before it reaches a client. I keep the review loop running and the standard current. That is the only part I stay for — and only as long as you want me to.
The agent cannot merge its own verdict. That separation is the point.
How a tool gets built. What it must prove before it ships. Who approves what. On paper, in your repository.
An agent reads every change against the standard and writes a verdict. A person decides. Nothing merges itself.
Every change, every approval, every accepted risk and why. When a client asks what your AI is doing, you hand them a document.
Most layers already run on open-weight models, and the proprietary one sits behind a key you can pull — the system still boots without it. The model is a configuration value, not a dependency.
Everything I build runs in accounts you own: your GitHub organisation, your model keys, a European instance that is yours alone — not shared, not offshore. The same pipeline can point at a model hosted on that instance, so a client’s account list is processed without leaving the EU and without reaching a US provider at all. Your clients are beginning to ask that question; this is an answer you can put in a pitch. Stop at the end of any month and it all stays where it is. Nothing to extract, no export to negotiate, no seat licences to renew.
I will not automate a process that should be deleted. A good share of what looks like an automation problem is a decision nobody has made yet. I will say so, and saying so costs you nothing.
I will not show you the internal tools of the operation I work in. Not blurred, not with invented data, not privately. The discipline that protects them is the discipline that will protect you.
I will not take a second client while yours is running. The hours you book are the hours you get, and I will turn down the next call rather than thin them out.
I will not leave you dependent on me. A standard that only works while I am in the room was never a standard.
Thirty minutes. I will tell you whether it can be automated, roughly what that takes at the hours we agree, and which parts I would leave alone. The third answer is the one nobody else gives you.
Germany · remote · working language English
No client names and no client data appear on this page. Every figure is counted from production repositories, September 2026.