In many businesses, somebody starts each day by pulling numbers. Yesterday’s sales from the point of sale system, collections from the finance system, orders from the website, stock from a spreadsheet. They copy the figures into a message or a sheet, format it, and send it to the owner or the management group. It takes forty minutes on a good day and much longer when a system is slow. It is also the kind of task that stops when that person is on leave. A morning report agent does this job every day at the same time, without being asked.
What goes into a useful morning report
The mistake most people make is to automate the report they already have. Before building anything, ask the people who read it what they actually look at. Usually it is a handful of things:
- Yesterday’s main numbers: sales, orders, collections, cash position.
- Comparison with the same day last week and the month so far against target.
- Exceptions: a branch with no sales, a large refund, stock that has run out, an unusually big order.
- Anything waiting on a decision: approvals pending, complaints open more than two days.
A good morning report fits on one phone screen. If people need to scroll through three pages to find the one figure that matters, they stop reading, and the report has failed even if every number is correct.
How the agent builds it
The work splits cleanly into two parts, and it is important to keep them separate.
- Gathering the numbers is ordinary software. At a set time, say 6:30 in the morning, a scheduled job connects to each system, runs fixed queries, and collects the figures. There is no AI in this part, and there should not be. A sales total should come from a database query, not from a model’s reading of a screenshot.
- Writing the report is where the model helps. Given the figures and the comparison, the model writes the short commentary, for example: “Sales were well below last Tuesday, mainly because the Kandy branch recorded nothing after 2pm.” It picks out the exceptions and orders them by importance.
This split matters. If the model is asked to calculate totals itself, it will occasionally get the arithmetic wrong in a way that looks perfectly plausible. Let the database count and let the model describe.
The finished report is delivered where people already look. In Sri Lanka that is often a WhatsApp group or an email, sometimes both. The delivery channel is a small detail that decides whether the report gets read.
Making it trustworthy
A report that is wrong once loses trust for months. Build in these protections from the start.
- Say when data is missing. If the finance system did not respond, the report should say “Collections: not available, system did not respond at 6:30” rather than show zero. A silent zero is worse than no report.
- Show the freshness. Every figure should carry the time it was taken, so readers know whether late entries have been counted.
- Keep the numbers and the words apart. Put the table of figures first, exactly as the systems gave them, and the commentary after. Anyone can check the words against the numbers.
- Log each run. Keep a record of what was fetched, when and from where. When a manager questions a figure, you can trace it in minutes.
- Tell someone when it fails. If the report does not go out by 7:00, a named person should get an alert. Scheduled reports usually die quietly, and nobody notices for weeks.
What it does not do well
The agent is good at spotting that a number moved. It is much weaker at knowing why. It may notice that a branch’s sales dropped, but it will not know that the branch was closed for a funeral or that the power was off for half the day. Its commentary should describe, not diagnose, and readers should treat any “because” as a hint to check rather than a finding.
It also depends completely on the systems it reads from. If sales are entered late, or two systems disagree about stock, the report will faithfully repeat the confusion. In some businesses, the first benefit of building a morning report is discovering how untidy the underlying data is. Fixing that is often more valuable than the report itself.
And if your business runs on one system and the owner can open a dashboard in ten seconds, you may not need an agent at all. The report earns its place when numbers live in several places and someone is spending real time stitching them together.
A sensible way to begin
Start with the report someone already sends by hand. Collect the last month of those reports, list the figures they contain and where each one comes from, and ask the readers which lines they would miss. Build only those lines first. Run the automated version alongside the manual one for two weeks and compare them every day. When they match for two weeks straight, let the person who used to write it spend those forty minutes on something that needs a human.