What we shipped last quarter, and what it changed

One of our clients, a consumer brand, came to us because their founder was starting and finishing every day in the inbox. Seven mailboxes, plus WhatsApp, social messages and marketplace chat, and around fifty customer emails a day in peak season.

What they asked for was not an AI that answers customers. It was a way to stop reading everything. So we built one triage layer across all of it, and the first thing it does is not AI at all. Plain rules, which they edit themselves, drop the newsletters and the notifications before the model ever sees them.

What survives lands in a named queue with the reason it landed there, a confidence score, and what a person should do next. Replies are drafted and wait. Refunds over an agreed threshold, anything legal, and diet-critical questions never become drafts at all. They ping a human.

Another client, an Australian property investment advisory, had the opposite problem. Plenty of automation, no way to see whether it had actually run. Deals moved, forms were supposed to fire, folders were supposed to appear, and nobody could say for certain that any of it happened on a given deal.

They asked for something their leadership could open on a Monday and trust. We built a dashboard over their live sales data with a tab per question. Where are deals sitting. Did the paperwork fire. Who is performing. How long is each stage taking.

The half that surprised people is the adoption tab. It does not report sales. It reports on us. Of 314 deals in the window, 313 have their form attached and 313 have a Drive folder, so the automation is either working or visibly not, on every single deal.

Our third client, a multi-region retailer, was reordering on instinct. Someone pulled exports, reconciled regions, and by the time the picture was assembled it was old enough that most reorders got made without it.

What they needed was not a prettier report. It was the same picture every week without anyone assembling it. So a daily sync keeps the numbers current, a tracker projects each product forward week by week, and a report lands in their chat every Monday.

The planner is the part their team actually argues over now. Set a revenue target, and it works out how many units each product has to move per week to hit it, whether current stock covers that, and the date you would have to order by given a 25-day lead time.

Three builds ended the quarter deliberately switched off. One creator follow-up flow was finished and tested and then held, because the team that would own it was not ready, and shipping into an unready team is how good systems die.

One alerting rule went back for rework after its first week. It worked, but not accurately enough, and a wrong alert costs more trust than a missing one. And we built a monthly reporting automation for ourselves, watched it run, and switched it off within days, once we saw that the manual review it replaced was where the judgement lived.