How much groundwork AI needs depends on how many systems your data lives in. A small firm on one accounting system and one CRM can connect an AI tool directly and get good answers, because there is little to reconcile. Once your data spans an ERP, a CRM, a ticketing system, a timekeeping tool, and a few hundred spreadsheets, you need AI-ready data before the answers hold up, and every acquisition that brings in another ERP or CRM adds to that work.
In our recent webinar, Brick Thompson and Landon Ochs talked through what that preparation looks like, drawing on how we prepared our own data for AI.
Where BI and AI each fit
Your dashboards handle the questions people ask every day. When twenty people check the same numbers each morning, a dashboard gives all of them the same validated figures, controls who sees which data, and avoids running a fresh AI query every time someone opens it. For the people running daily operations, BI works as the operating system of the business.
AI handles the question nobody built a dashboard for. An executive who spots an unusual number can ask what is behind it, and an analyst can pull margin by customer or trailing twelve-month margin by service line in minutes, then turn the result into a slide for the next board meeting. Reports that once took a controller hours in Excel, like a quarterly profitability review or the weekly operating spreadsheet everyone waits on, can be ready on demand once the data underneath is prepared. Both tools draw on the same data, and both depend on that data being right.
Preparing AI-ready data
An AI model answers every question with the same confidence, whether it has the facts it needs or is filling gaps with its best guess, so its answers look equally polished either way. In a company with several systems, the gaps show up in familiar places. Your CRM and your ERP each assign their own ID to the same customer, and the names rarely match exactly, so the model has to guess which records belong together. Your VP of sales usually means booked business when they say revenue, while your controller means the top line on the P&L for a given period. The spreadsheet someone uploaded last spring has been revised several times since.
The work that makes BI reliable closes most of those gaps, since your reporting model already matches customers across systems and settles each definition. On top of that model, we build a set of tables written for AI to read, with plain English column names, the details a question needs kept in one place, and the most common calculations already done. Each of those choices leaves the model fewer decisions to make on its own, so the same question returns the same answer more consistently, and because we store them as finished tables, a question comes back in seconds.
The AI also needs to understand your business, so we write a guide it can read, covering what the company does, the acronyms your team uses, which data answers which kind of question, and details like a 13-period fiscal calendar. Early on, some of our own answers came back wrong for no obvious reason, until one of our engineers asked Claude whether it had read the guide and it said it hadn’t. We now set up the connection so the model has to read the guide before it can run a query, and it has read it every time since.
Getting started safely
Security stays with your data platform, so when the AI queries on behalf of the person asking, it can only reach the data that person is permitted to see. A regional manager asking about sales gets their own region’s numbers, and a request for broader access ends with a suggestion to contact an admin. Live AI dashboards, which run a fresh query each time someone opens them, follow the same rules.
You can start with whichever part of the business already has validated reporting, such as sales or finance, and add others as each one is ready. Each area gets its own AI tables and its own section of the guide, along with rules about which questions are safe to answer so far. The definitions your team settled while building those reports are the same ones the AI will rely on, which makes any reporting you have already validated foundation.