Three systems to install before Q4 planning

Every week at one Australian financial-services client closes with the same status report. It has run that way since early 2025. There were 71 dated reports in the shared folder the day we took these screenshots.

The week opens the same way. A scheduled email carrying the week's meetings lands before anyone has asked. The assembly is automated. The words about what the week meant are not, and that split is deliberate.

The Saturday before one August week, the founder replied to the monthly lead report. The leads-assigned figure said 462 out of roughly two thousand leads. That looked wrong to him.

His question, the screenshot above, was posted into the shared Slack channel on the Monday. Five replies later the calculation was explained and the tick went on. That is what a scheduled report is for. The challenge lands on a number everyone can see. It never becomes a meeting about what the numbers even are.

Requests travel the same road. A change asked for in chat becomes a tracked ticket the same evening. The work then shows up in the Friday report. The fourth item under Focus This Week below was raised in that channel eighteen days earlier. Raised in chat, tracked on the board, reported on Friday. The loop closes in writing.

Freckleberry is a family chocolate factory. Seven mailboxes plus WhatsApp, Meta DMs and Shopify Inbox. Around 50 customer emails a day in peak season. A team member cleared the backlog between 8pm and 1am.

We built one triage layer across all of it, and the first thing it does is not AI. Plain rules, editable by the client from a dashboard, drop newsletters and notifications before the model sees anything. On 16 June that filter removed 81 messages in a single day. Every drop is logged with its reason.

Then the mistakes, because there were some. An early version read only the newest message in a thread. It politely re-asked questions the customer had already answered. In one case it drafted a gift-card reply keyed off a stray phrase, missing what the thread was actually about. The fix was a sharper rule, not a smarter model. Read the whole thread.

It also guessed at prices on quotes. Once it quoted a school discount inline that should have gone through the approval flow. So pricing was removed from every AI reply. Drafts now link to the website instead of quoting numbers.

What survived is a system that prepares the work and keeps people in charge. Every reply waits in the mailbox's own Drafts folder with an [AI DRAFT] prefix. Refunds over an agreed threshold, legal or regulator mentions and diet-critical questions never become drafts at all. They ping a named human. The founder works two queues and a 7am digest instead of a midnight inbox.

The third system is ours. Every sales and client call at DashboardLim lands in one table with its full transcript. The recordings arrive from tl;dv, Gemini, GHL and Loom. AI types the call and writes the structured brief.

The brief waits for a person's approval in Slack. Only then do the CRM note, the follow-up draft and the next step land. Nine workflows run the pipeline.

Counted against the live database this month, the memory held 841 calls, 551 structured summaries and 4,027 searchable records. Those numbers are floors. They only grow.

What that buys is easiest to show with the client from System 02. The discovery call happened in April. The proposal quoted that call back to the founder and was approved two weeks later. Every check-in since has started from the last conversation instead of a re-introduction. The most recent one's decisions, training the agents on gifting keywords and routing wholesale enquiries to a human, are recorded as next steps. Below is a real question asked of the live memory. The client and commercial details are masked.

Look across the three and the shape repeats. Machines do the collecting, routing and drafting on a schedule. People own the judgement. The explanation in the thread. The edit to the draft. The approval before the CRM write.

And every exception has an address. A challenged number becomes a thread. A strange message becomes a queue row. A summary becomes an approval with a name on it.

None of these systems is exotic, and none is finished. The reporting cadence gets challenged. The inbox gets corrected. The memory gets queried and edited. That is the point. The systems that survive are the ones built to be questioned.