From lead to off-boarding: agency automation.
A client's board said five campaigns were still to do. All five were live and spending. Someone had shipped the work and forgotten the card. We don't send a report we can't stand behind, so we built the thing that checks every board against the real ad account every morning. Then eleven more like it, covering everything from capturing a lead to handing the whole account over on the way out. All of it before we had the client count to justify any of it.
One warehouse per client. Everything reads from it. Nothing gets its own private copy.
The whole arc, one warehouse.
Every client gets their own store, filled daily from the ad account, the shop, the subscription system and their own attribution tool. Every automation below reads from that one place, so two people can never quote two different numbers in the same meeting. Stage three is the one that repeats: it runs every week for as long as they stay.
Every test on one board.
This is what stage three looks like from the inside. Each card is one experiment with its own label in the ad account, so the spend can always be traced back to the card that authorised it. Nothing moves right without a person moving it.
Experiments9 open2Work14 open
Meetingsnext: thu
Warehousesynced 07:04Experiments
read due: 2Built paused, labelled XP-### · the run flip is the only thing that spends money
The handover at the end of the arc is the part most agencies do not have. A client who leaves takes the warehouse, the dashboard, the boards and the automations with them, under licence. We wrote that down before we had a client who could ask for it.
It only speaks when something is off.
The truth-check posts nothing on a normal day. Silence means the boards, the ad account and the invoice agree. When they don't, it says exactly what drifted and leaves the fix to a person.
Two boards drifted overnight. Nothing was changed for you.
Built before the scale that needs it.
One lesson cost real money before it became a rule. An early demand-mapping run was about four times more expensive than it needed to be and produced a headline number that did not survive checking, because the system was doing arithmetic in prose. Now a script does every calculation that touches a client's money, and the model only writes the reasoning around it.
Reproducible, or it doesn't ship.
Every client's boards, ids and settings live in one config file and a set of blueprints. That is what makes a handover possible: if a client leaves, they take the warehouse, the dashboard, the boards and the automations with them.
Google Adsbuild · measure
Shopifyorders · productsmonday.comboards · blueprintsCalendarmeeting intakePostgreswarehouse · learnings
Claudethe reasoning layer
Railwaycrons · services
Slackfour channels per client