How to Build an Automated Monthly Reporting Workflow with AI

6 min read

You build an automated monthly reporting workflow with AI in five stages: a schedule trigger fires on the first of the month, a data step pulls the raw numbers from your database or spreadsheets, a code step computes every metric deterministically, a language model writes the narrative using those locked numbers, and a final step assembles the document and sends it. The single rule that makes this trustworthy: the model never calculates anything. It only writes sentences around figures your code has already computed and frozen.

Key takeaways

Compute every number in code and lock it. The model writes prose around fixed figures, it never does the arithmetic.
Use a schedule trigger plus a data pull plus a deterministic metrics step plus an LLM narrative plus assembly and delivery.
Add a validation pass that checks the report contains only numbers your code produced, and fails loudly if not.
Keep the prompt boring: feed it labelled metrics and ask for plain sentences, not analysis it might invent.
A monthly report is cheap to generate. The expensive failure is one wrong number that a stakeholder acts on.

The shape of the workflow

Every reliable reporting workflow I have built has the same skeleton. The temptation is to hand the model a pile of raw data and ask it to "write the monthly report". Resist it. That is exactly the setup where a model confidently reports growth that did not happen.

Schedule trigger. Fire on the first of the month, early, before anyone is awake to ask where the report is.
Pull the data. Query your database, Stripe, analytics, or a Sheet. Get the raw rows, nothing computed yet.
Compute the metrics in code. Totals, deltas, percentages, all in a code step. These numbers are now frozen.
Write the narrative. Hand the locked metrics to the model and ask for plain sentences that reference them.
Validate, assemble, and send. Check the prose uses only your numbers, render the document, deliver it.

Monthly metrics dashboard with charts feeding into a written summary

The one rule: code owns the math

Here is the metrics step as a code node. It computes everything, rounds once, and produces a clean object of labelled numbers. Notice that nothing downstream is allowed to recompute. The model receives these exact strings and is told, in the prompt, to use them verbatim.

javascript

// n8n Code node: compute and freeze every reporting metric
const rows = items.map((i) => i.json);

const revenue = rows.reduce((s, r) => s + r.amount, 0);
const lastMonth = $('previousMonth').first().json.revenue;
const deltaPct = ((revenue - lastMonth) / lastMonth) * 100;

// Round ONCE, here, then never touch these again
const metrics = {
  revenue: `$${revenue.toLocaleString()}`,
  newCustomers: String(rows.filter((r) => r.isNew).length),
  growth: `${deltaPct.toFixed(1)}%`,
  direction: deltaPct >= 0 ? 'up' : 'down',
};

return [{ json: { metrics } }];

The prompt then looks like this: "Write a four-sentence summary. Use only these figures, exactly as written: revenue {{revenue}}, new customers {{newCustomers}}, growth {{growth}} {{direction}}. Do not compute or estimate any other numbers." The model is reduced to doing the one thing it is genuinely good at, turning structured facts into readable prose, and blocked from the one thing it is bad at, arithmetic under pressure.

The month the AI rounded its way into a wrong number

February 2025, a 10-person agency I was advising sent monthly performance reports to a dozen clients. Marco, who ran delivery, had wired up a tidy workflow that handed the model the raw lead data and asked it to write the summary. For three months it was fine. The fourth month, one client report said "leads up 20%" when the real figure was 12%. The model had not pulled from a metrics object, because there was not one; it had eyeballed the raw numbers and rounded toward a story. The client noticed before Marco did. We rebuilt it the boring way: a code step computed every figure, the model got only the locked numbers, and a validation step rejected any stray digit. The wrong numbers stopped that month and never came back. The lesson was not "AI is unreliable". The lesson was "do not ask it to do arithmetic you can do in code".

Running this costs almost nothing. A monthly narrative is a few hundred tokens, and at GPT-4o-mini’s 0.15 dollars per million input and 0.60 dollars per million output (OpenAI 2024 pricing page), a dozen client reports cost a rounding error. The value is not in saving the model cost. It is in nobody spending a Friday afternoon assembling a deck by hand.

My strong opinion on AI and numbers

I will say it plainly: never let a language model compute a number that ends up in front of a stakeholder. Models are better at arithmetic than they used to be, and people will tell you the newest one can be trusted with a sum. Maybe it can, most of the time. "Most of the time" is exactly the failure mode that burns you, because the one wrong figure arrives wearing the same confident tone as the ninety-nine right ones, and you cannot tell them apart without checking, at which point you have done the work anyway. Compute in code, lock it, let the model write. Where I will soften: for a throwaway internal scratch summary nobody acts on, who cares, let the model freestyle. For anything a client or an executive reads, the math belongs to your code. For the tooling side of this, see the best AI tools for monthly reporting workflows and the fully local version in automating monthly financial reports with a local LLM and n8n.

“Let the model write the sentence. Never let it do the sum. The difference is the entire reason anyone trusts the report.”

Frequently asked questions

Can the AI just generate the whole report from raw data?

It can, but do not let it. Handing raw data to a model and asking for a report invites it to compute and round numbers itself, which is where wrong figures creep in. Compute the metrics in code first, then let the model write prose around them.

What tools do I need to build this?

A scheduler and a workflow tool such as n8n or Make, a data source like your database or Stripe, a code step for the metrics, and any LLM API for the narrative. No custom infrastructure is required for a basic monthly report.

How do I stop the AI from inventing numbers?

Two safeguards. Feed it only the locked metrics and instruct it to use them verbatim, then add a validation step that scans the output for any number not in your metrics object and fails the run if it finds one.

How much does an automated monthly report cost to run?

Very little. The narrative is a few hundred tokens, so even a dozen reports cost a fraction of a cent in model fees. The real saving is the human hours you no longer spend assembling reports by hand.

Should the report send automatically or go to a human first?

For internal reports, automatic is fine once the validation step is in place. For client-facing or executive reports, route the first few months to a human for a quick read before sending, then automate once you trust the pipeline.