I spent four years as a business intelligence architect (a time I reflect on separately), and I can tell you what the job mostly was. It was not insight. It was assembly. Every Monday someone pulled the same exports from the same three systems, pasted them into the same workbook, fixed the same formatting problem, and sent round the same pack. Then, all week, people asked variations of the same questions. Can I see that by region? What about excluding returns? Same thing but for last year?
Each of those was a small queue. Each answer took a day to arrive, by which time the person had stopped caring or, worse, had made the decision without it.
What changes
An AI analyst does the assembly and answers the variations. In practice that means three things.
The pack builds itself. The weekly report is generated from the sources on a schedule, runs alongside the manual version until you trust it, and then replaces it. It is ready before the meeting rather than during it.
Questions get answered when they are asked. “Sales by channel, last six weeks, excluding the clearance lines” is a sentence, not a ticket. The people closest to the decisions, who were never going to learn a BI tool, can simply ask.
Exceptions come to you. Instead of scanning forty rows for the one that matters, the analyst tells you: this store is down twelve percent against trend, this product’s margin moved, stock cover on this line is under two weeks.
By hours, I would put that at seventy to eighty percent of what an analyst in a small or mid-sized business does. The part that remains is the valuable part: deciding what is worth measuring, noticing what the numbers cannot tell you, and telling the story to people who have to act on it.
The rule that stops it causing chaos
When numbers become cheap, you get more of them, and they disagree. Someone builds a weekly summary one way; someone else asks the same question slightly differently; both arrive at the meeting with confidence and different answers. A consultant friend of mine was once called in to counsel a leadership team whose chief executive had discovered he could generate reports in an hour, and had used them to tear into his people. After that, nobody on the team wanted to show him how anything was done.
So, one rule, agreed before anything is built: every headline metric has one official source and one owner. Everyone is encouraged to explore. An exploratory number is labelled as one when it leaves your desk. And AI-generated analysis of a colleague’s area is shared with them, not sprung on them.
How it is built
The sources stay where they are: your point of sale, your e-commerce platform, your accounting system. A thin layer reads from each, by API where you have one and by scheduled export where you do not, and turns them into plain, consistent rows in a small database you own. The report and the questions both run on that layer.
That has a useful side effect. When your software vendor announces its own AI reporting feature, or you decide to change BI tools, nothing above the layer has to change. The official number stays the official number whichever system is feeding it. It also makes the old question, “do we still need that BI subscription?”, one you can answer calmly.
Where to start
With the pack you already produce. Have the person who builds it record a fifteen-minute walkthrough of where every number comes from. That recording is the specification. Build it, run it in parallel, compare, cut over. Then add the questions people actually ask, one at a time.
If someone in your business loses a day a week to assembling a report, tell me what it is built from. I will tell you honestly how much of it a machine can take, and it is usually most of it.