Automation

A read on the month, in five sentences

A grade, a short summary and one to four things to do next, written from a fact sheet the database produced. The AI writes the sentence; the database writes the number.A grade, a short summary and one to four things to do next, written from a fact sheet the database produced rather than from raw rows handed to a model. The AI writes the sentence. The database writes the number.

Pacing 9% above goal· Main Performance
Paid carries 78% of leads· Sources
58% of leads qualified· Funnel
Plovdiv at 2.4× expected· Geo map
No Answer up 6 points on last month· Leads Status
AI Insights
4/5
Twenty-three days in, leads sit at 1,284 of the 1,400 goal and project to 1,527, roughly 9% clear. Qualified is on the same curve at 742 of 800, so the pace is not coming from thinner leads. Spend is the pressure point: €18,420 of a €22,000 budget, tracking to about €24,800 by the 31st. CPL is holding at €14.35 under the €16.00 ceiling, which is what keeps the lead goal affordable. Response time averages 1h 24m against a 1h 30m target. Inside it, but weekends run long.
Recommended Actions
Cap Meta daily budget to land under €22,000
Hold Google Search spend: it carries the 1,400 goal
Add weekend cover in Plovdiv to protect 1h 30m
The architecture

Why this one can be trusted with a number

Everyone has seen an AI state something false with total confidence. The answer is structural.Everyone has now seen an AI feature state something false with total confidence. The answer is structural.

The model never sees raw rows

It receives a fact sheet: a fixed set of figures, already computed, each one labelled. It cannot do arithmetic on your database because it is never given your database.

The figures match the screen

The fact sheet is built from the same live aggregators the dashboard renders. Deliberately, over faster rollup tables, because the two once disagreed and a summary that contradicts the chart above it is worse than no summary.

The deterministic half is always on

The short notes under each chart are computed on the server with no model involved: a divergence, a weekday effect, a change point. They are there whether or not you use the written summary.

What you get

Grade, summary, actions

01

A grade

One letter for the period, derived from pacing against your own goals rather than against strangers.One letter for the period, so the state of the month survives being glanced at. It is derived from pacing against your own goals, not from a comparison with strangers.

02

At most five sentences

Long enough to say what moved and where, short enough to read. If nothing happened, it says so.Long enough to say what moved and where, short enough to read. If nothing meaningful happened, it says so instead of padding.

03

One to four actions

Specific and traceable: which campaign to change, and why it is the one.Specific and traceable. Each one points at a figure in the fact sheet and names the campaign to change, and why it is the one.

Questions

Insights, in detail

How often does it run?

On demand today. You generate a read for a brand and a period when you want one. Scheduled runs are built but not switched on, and we would rather say that than let a “fresh every morning” claim sit on the page.

Can it be wrong?

It can emphasise the wrong thing, the way any analyst can. What it cannot do is state a figure that is not in your data, because it has no route to invent one. Every number in the summary is a line of the fact sheet.

Does my data leave Airdec?

A fact sheet does: a page of already-aggregated figures, with no lead records, no contact details and no credentials. The service that writes the summary has no access to the database at all.

Can I see what it was given?

Yes. The fact sheet behind a run is inspectable, which is the point: an insight you cannot check is an opinion.