An AI agent for preparing weekly business reports—and daily summaries

An AI agent for preparing weekly business reports starts with named inputs

Before you pick a template or a schedule, decide exactly what the agent is allowed to look at. A report you can trust has five parts, and this is the whole recipe:

• Named inputs — the specific sources each number comes from, listed by name.
• A fixed template — the same blocks, in the same order, every time.
• Success criteria — what a good report contains, written down before the first run.
• A source behind every number — so any figure can be traced back and checked.
• A readable cadence — a daily summary, a weekly report, or both, on a schedule you can predict.

Start with the inputs, because they set the ceiling on everything else. Instead of “our data,” write the list: last week’s orders, the support inbox, the two ad accounts, the subscriptions sheet, and the refunds log. Six named things, and nothing beyond them.

Why it matters: a bounded list is testable. You can point at each source and confirm the agent used it; “whatever is relevant” gives you nothing to point at. How to judge it: read the report and ask, for every figure, “which named source did this come from?” If even one answer is “not sure,” the input list is too loose. The failure case is familiar—asked to summarize the whole business, the agent quietly drops the source it could not reach and never mentions the gap, so a blank week reads as a calm one. The trade-off is honest maintenance: a named list must be updated when you add a channel or retire a spreadsheet. That upkeep is the price of a report you can actually verify.

Fix the template so you scan instead of re-read

Once the inputs are named, give the report a shape that never changes. Decide the blocks in advance—for example: the one headline number, what moved and by how much, what needs a decision this week, and anything unusual worth flagging. Same blocks, same order, every single time.

Why it matters: a stable shape lets you read in seconds and, more importantly, notice what is missing. When the “decisions needed” block is always third, an empty third block is a signal, not a mystery. How to judge it: put two weeks side by side; if the structure looks identical and only the numbers differ, the template is working. The failure case is free-form prose that reshuffles itself each week—fluent, pleasant, and perfectly capable of hiding that last week’s comparison simply is not there. The trade-off is rigidity: a fixed template struggles with the genuinely strange week, so keep one open block, literally titled “anything else worth flagging,” as the escape valve. Everything structured lives in the fixed blocks; everything surprising lives in the one that is free.

Write the success criteria before the first run

Decide what “good” means while you are calm, not while you are grading a draft at 7 a.m. Write it down as a short checklist the report either passes or fails: every number names its source, no fixed block is empty, week-over-week change is shown, and anything the agent could not confirm is marked clearly rather than smoothed over.

Why it matters: without written criteria you grade on tone, and a confident, well-worded report reads as a correct one. Google’s People + AI team frames this as defining success before you build, so you are measuring a result you agreed on rather than reacting to whatever the system produces. How to judge it: run the checklist against the first draft out loud; if you cannot answer each item with a plain yes, the report is not ready. The failure case is the fluent miss—a polished report that quietly omits the churn line, which you approve because nothing looked wrong. The trade-off is that strict criteria reject more early drafts and feel slow in week one; that rejection is exactly how the report earns the right to be skimmed in week six.

Review before you trust—every number points to a source

For the first few weeks, do not skim. Open the report and spot-check figures against the named source they claim to come from. The goal is not suspicion; it is calibration—learning where this agent is reliable and where it is not, before you start acting on its numbers without looking.

Why it matters: a confidently wrong number is worse than a blank one, because it feels finished. This is why the report should carry its sources with it—so a figure can be traced in one step instead of investigated across three tabs. Research on human-AI interaction is consistent on the point: outputs people can inspect and correct build appropriate trust, while opaque ones invite either blind acceptance or blanket doubt. How to judge it: pick any figure at random and try to trace it to its source in under a minute; if you can, the report is reviewable. The failure case compounds quietly—an agent that averages the wrong column produces a wrong trend that looks more convincing each week it repeats. The trade-off is plain: review costs you minutes now, and those minutes are precisely what let you stop reviewing later.

Run two cadences: a daily summary and a weekly report

The same inputs and template can feed two schedules that answer different questions. An AI agent for daily business summaries catches surprises—a refund spike, a support thread going sideways—while the weekly report shows the trend and the decisions. Schedule both from the dashboard; the agent’s persistent memory holds last week’s numbers so the weekly report can show change instead of restating totals, and the pre-configured Telegram gateway can deliver either one where you will actually read it.

Why it matters: a daily summary and a weekly report are not the same document at different lengths. The daily should be three lines and changes only; the weekly should be the fuller five blocks. How to judge it: the daily is working if you can read it in under a minute and act on nothing most days; the weekly is working if five minutes tells you what to do next. The failure case is a daily summary that repeats everything—it becomes noise, and noise is something you stop reading, which quietly defeats the point. The trade-off is real: every cadence you add is more to read, so add a daily only if you would act on it the same day. Two field notes go deeper on the timing itself—how often your agent should run for matching cadence to how fast the answer goes stale, and the dashboard signals that tell you an agent is healthy for confirming the scheduled run actually happened and finished clean.