A Monday KPI brief that explains last week's numbers instead of just charting them

Five tools, four revenue numbers, one Monday meeting

Picture a direct-to-consumer apparel brand doing around EUR 8M a year, mostly through Shopify, with a small wholesale side in HubSpot. Three people own the numbers: the founder, an operations lead, and a freelance performance marketer. Every Monday morning the ops lead exports CSVs from Shopify, Meta and Google Ads into a master sheet, redoes the same lookups, and is the only person who can.

Then the meeting starts with reconciling. Shopify says revenue fell 9 percent, Meta reports its best week of the quarter, and GA4 shows conversion rate collapsing on Saturday. Twenty minutes go on deciding which number is real, and the meeting ends before anyone discusses what to do. The Looker Studio dashboard built last year is accurate, complete, and opened about once a month.

What the founder wants is three paragraphs: what changed, why, and whether any of it is a problem. That is an arithmetic job followed by a writing job, and the order matters. I compute every figure in SQL, check the data is complete, and only then let a model explain the week from what the code found. As one engineer put it in an n8n community thread about factory reporting, 'a wrong production report is worse than no report, because someone acts on it.'

A SaaS company has the same shape with different sources: MRR from Stripe or Chargebee, pipeline from HubSpot, activation from the product database, and a monthly investor update that otherwise means Stripe, then Sheets, then screenshots. Follow-up questions ('split that by channel') belong to a plain-English question box over the database.

Last Monday's brief, with the queries underneath

An illustrative brief for the brand above. The tool lines ran before the model wrote anything; the message is what the team read.

The metric dictionary the brief is written from

Every figure comes from a definition agreed before the first run and implemented as a SQL model. The founder signs this table off, and the model never sees a metric outside it.

From Sunday midnight to the 07:00 post

Most of these steps are scheduled jobs and SQL. The model appears once, to write, and only sees what the earlier steps hand over.

Six ways a weekly number misleads

None of these is unusual. Each turns up within the first few months of any weekly report, and each is handled in code before the model sees anything.

Who owns which part of the brief

Two weeks from kickoff to the first real Monday

One of the smallest builds on this site, and a good first project: useful from its first Monday, and low risk, because it reads data and changes none.

1. Agree the metric dictionary with the founder: what counts as revenue, which orders are excluded, what a promotion is, and which week the business uses.
2. Connect the sources with read-only credentials the business owns, and backfill two years so every week has a same-week-last-year.
3. Write the SQL models and reconcile three past weeks against Shopify's own reports, until every remaining difference is one the dictionary explains.
4. Break things on purpose: delete a day of GA4 data, duplicate a sync, shift a timezone, and confirm the brief says incomplete each time.
5. Generate briefs for eight past weeks and have the ops lead mark each sentence right, wrong or pointless. Those marks become the evaluation set every later prompt or model change must pass.
6. Go live with the ops lead approving four Mondays in a row, then let it post on its own.

Credentials, SQL and prompts live in the business's own accounts and repository, with a runbook, so the brief keeps running whether or not I stay involved. The LLM evaluation guide covers how an evaluation set like this is scored.

Is Triple Whale or Polar enough, or is your own brief worth building?

For a Shopify brand whose data sits in Shopify, Meta, Google and Klaviyo, look at the DTC analytics platforms first. Triple Whale and Polar Analytics connect those sources in a few clicks, each takes its own approach to attribution, and both now include AI features for questioning the data; Polar can also push insights into Slack. If your KPIs are the standard ones, a subscription will be cheaper and faster than a build, and I will tell you so.

The other cheap route is Looker Studio on a connector such as Supermetrics, plus someone who writes the commentary each Monday. It works while that person is careful and not on holiday. Its weak spot is the one this page is about: nobody checks that the data was complete before the story was written.

A custom brief earns its place when your sources or definitions are unusual: wholesale in HubSpot beside DTC in Shopify, a 3PL's stock feed, a SaaS product database, margin that depends on your accountant's cost of goods, or several stores in different currencies. It often runs next to a DTC platform rather than replacing it. Built as AI workflow automation, it is a small project, and the SQL, prompts and history stay yours.

How you would know it is working

A blueprint has no results to report, so here is what I would measure from the first week instead, on your own data.

Minutes spent reconciling

For four Mondays before launch and four after, time how long the meeting spends agreeing on numbers before it talks about actions.

Corrections per brief

Sentences the ops lead corrects in the approval month, each traced to a definition, a data problem or the prompt.

Tracking problems caught

Health-check alerts that turned out to be real, and how many days earlier they were found than the team would have noticed.

Drill-down clicks

Clicks on the brief's links, logged through a redirect. None at all means the brief answers everything, or the wrong questions.

Validator rejections

Drafts regenerated because a figure or cause did not match. A rise after a model or prompt change is a reason to roll that change back.

What a build like this costs

This is built as AI Workflow Automation, which runs $3.5K - $60K overall. A build like this one usually lands in the single-step flow tier: $3.5K - $12K, 1-2 weeks. The first working version runs on your real data well before the end of that window.

What it costs to run

Small. The model writes a few hundred words a week, so model costs are cents a month. The warehouse and any connector subscription cost more than the AI does, and at this data size BigQuery often costs little or nothing.

What moves the price

• How many sources, and how clean: Shopify plus two ad accounts is quick, a 3PL portal without an API is not
• Whether margin is in scope, which means reconciling cost of goods and payment fees with the accountant
• Several stores or currencies, each needing a conversion rule and one reporting currency
• How much history to backfill, and whether a warehouse already exists
• More than one audience: founder, marketing and board versions are three outlines and three validators

Who this is for

• Shopify brands with a few million to a few tens of millions in revenue and no analyst on staff
• Founders who get four different revenue numbers from four tools every Monday
• Ops leads who spend Monday morning exporting CSVs into a master sheet only they can maintain
• SaaS teams with MRR in Stripe or Chargebee, pipeline in HubSpot and activation in the product database
• Brands selling direct and wholesale at once, where no off-the-shelf dashboard shows both

Questions people ask about this