AI business intelligence: BI for a business without a data team

AI business intelligence is business intelligence (BI), the reports and analysis a company builds from its own records, with AI doing work that used to need an analyst: answering questions asked in plain language, flagging unusual numbers, forecasting from history and writing summaries of what changed. For a business without a data team, it lets the owner and managers ask their own questions of the data. It needs a data foundation underneath: a data warehouse, or a smaller reporting database, fed from the ERP and the books. This guide covers what AI adds, the data it needs, the categories of AI BI tools and how to start small.

Derik Lawlis Founder, ThriveAI

What business intelligence is, and what AI adds

Business intelligence is the reports, dashboards and analysis a company builds from its own records to see how it is doing and decide what to do next. AI adds four capabilities on top of the same data.

Adoption is rising from a low base. In Statistics Canada’s survey for the second quarter of 2026, 19.2% of businesses had used AI to produce goods or deliver services in the previous 12 months, up from 6.1% two years earlier. Among those businesses, data analytics was the most common use, at 36.6%.

How BI works without a data team

Without a data team, BI work falls to the controller, the owner or the one person who knows the ERP. Take a distributor with 60 employees as an example. At month end, the controller exports sales and costs, updates a spreadsheet that produces the management pack and sends it a week after the books close. The questions the pack raises wait for the next month, because answering them means another round of exports.

Some businesses pay a consultant to build dashboards in a BI tool. The dashboards then need someone to change them whenever the ERP or the business changes. In both cases the limit is the same: the number of questions one busy person can answer in a week.

AI raises that limit without a new hire. It writes queries and drafts commentary, and its alerts watch the numbers between meetings. A person still has to decide what each number means and keep the data clean.

The four capabilities, in practice

Plain-language questions

A manager types a question and gets a table or chart. Behind it, the AI writes a query against a model of your data, and the answers depend on that model: if it defines revenue, every answer uses the same definition. Looker’s conversational analytics, for example, uses “the LookML definitions of your data” as its source of truth, which is how “revenue” stays consistent from one answer to the next. AI reporting covers how these answers are produced and checked, and ad hoc reporting covers the one-off questions they suit.

Anomaly alerts

Anomaly detection learns the normal range of a number over time and flags values outside it. In Power BI, anomaly detection adds an expected range to a line chart, marks the values outside it and offers possible explanations drawn from other fields in the data model. Microsoft notes the trade-off: at higher sensitivity, “even a slight deviation is marked as an anomaly.” Tune the sensitivity on past data, so the alerts match the events your team would have wanted to hear about.

For a distributor, a useful alert is a customer’s orders falling below their usual range, or a supplier’s prices climbing across several invoices. For a plant, it might be scrap on one line rising above its normal level. Each alert should go to the person who owns that customer, supplier or line, with the rows that triggered it.

Check which feature does which job. Microsoft’s documentation for Copilot in Power BI, updated September 29, 2026, says Copilot can’t answer questions that require anomaly detection, forecasting or finding key influencers. In that product, anomaly detection is a separate feature of line charts.

Forecasts

A forecast projects a number forward from its history. Excel’s forecast sheet shows the basics: Microsoft says it uses “the AAA version of the Exponential Smoothing (ETS) algorithm” and draws a confidence interval, the range in which 95% of future points are expected to fall at the default setting. Two habits carry over to any forecasting tool. Show the range, because a single line hides how uncertain the forecast is. And give seasonal patterns enough history: Microsoft advises against setting a seasonality with fewer than two cycles of data, so a yearly pattern in monthly sales needs at least 24 months. AI demand forecasting covers forecasting demand from your own order history.

Written summaries

A written summary tells a manager which numbers moved, by how much and what drove the change, so nobody has to read every chart to find the ones that changed. The safe design computes the numbers first and lets the language model only describe them. Tableau describes its Pulse insight summaries that way: statistical analysis generates facts about each metric, and the summaries use “natural language grounded in statistical truths.” A summary that a language model writes straight from raw data can state a wrong number fluently, so ask each vendor how its summaries are produced.

The data foundation: a warehouse fed from the ERP and the books

Every capability above reads from the same place: a store of clean, consistent data. For a smaller business, that store can be a small data warehouse. AWS defines a data warehouse as “a central repository of information that can be analyzed to make more informed decisions,” with data flowing in from transactional systems “typically on a regular cadence.” For a business that makes, moves or sells goods, it is fed from:

• The ERP: orders, shipments, inventory, purchasing and production.
• The books: the general ledger, invoices, payables and costs from the accounting system.
• Spreadsheets: budgets, price lists, rebate agreements and targets.
• Other systems where they exist, such as a warehouse management system, a CRM or a quality system.

A few design choices matter more than which database you pick:

• Copy the data on a schedule. Load it nightly or hourly, so heavy questions don’t slow the ERP that people are using to ship orders. AWS lists separating analytics from transactional databases as a benefit that “improves performance of both systems.”
• Keep snapshots. Store a copy of inventory and open orders each night, so questions about past positions don’t depend on rebuilding them from transaction history.
• Write the definitions. A semantic layer, the definition of each business term in a form software can read, is what lets an AI answer “what were sales?” the same way every time.
• Reconcile to the books. Monthly sales and margin in the warehouse should match the income statement before anyone relies on them.
• Show the refresh time on every report and dashboard.

If your ERP is older and has no API, the data can still reach a warehouse through a read-only database connection or scheduled exports. AI for ERP covers the routes, and legacy ERP automation covers systems installed on your own server years ago.

Categories of AI BI tools

AI BI tools fall into five categories. The examples are well-known products, listed to show what each category looks like. Features differ by edition and version, so check the release you would run.

Ask every vendor the same questions: how the tool applies your ERP’s permissions, whether each answer shows its query and rows, how summaries are produced, and where prompts and results are processed and stored.

How to start small

1. Pick one subject. Sales and margin make a good first subject: the data sits in the ERP and the books, and the questions come up every week.
2. Load two sources. Copy sales lines and item costs from the ERP, and the general ledger from the accounting system, into a reporting database each night.
3. Write ten definitions, starting with sales, cost, margin, customer and product family, and the date that decides which month a sale belongs to.
4. Reconcile. Make monthly sales and margin match the income statement before anyone uses the numbers.
5. Add one AI capability at a time. Start with plain-language questions, then a weekly written summary, then alerts on the few numbers that matter most.

Margin analysis and AI in finance use the same sales and cost data, and AI dashboards cover how to present it to the people who act on it.

Risks and limits

• Data quality. AI answers from the data it has. Duplicate customers, unposted invoices and stale standard costs flow into every answer, summary and alert. Fix them at the source and keep a record of the fixes.
• Alert fatigue. Too many alerts, or alerts at the wrong sensitivity, teach people to ignore them. Start with a few and tune them on past data.
• Forecasts read as facts. Show the range with every forecast, and compare each forecast with what happened.
• Skills. In the same Statistics Canada survey, a lack of skilled workers was a barrier to AI use for 12.0% of manufacturers. Even with AI writing the queries, someone has to own the data and the definitions, whether that is an employee or an outside partner.
• Where the data goes. Cybersecurity or privacy concerns were the barrier businesses cited most, at 13.4%. Microsoft’s requirements for Copilot in Power BI say it is disabled by default when a tenant or capacity is outside the United States or EU data boundary, and turning it on means allowing data to be processed outside the tenant’s geographic region. Data sovereignty and private AI for business cover where the data and the model can run.

How ThriveAI helps

ThriveAI is an AI engineering company in Ottawa that builds private AI systems on a company’s own data, for businesses that make, move or sell physical goods. It brings what your ERP, email, drawings and spreadsheets hold into one database that belongs to your company. Your team asks it questions in plain language, and each number links to the document or ERP record it came from. When the data cannot answer a question, the system says what is missing instead of guessing.

ThriveAI’s systems read your ERP and your other software as they are, including older versions installed on your own server, and every connection only reads data. The platform is designed to keep each client’s data on its own server in Canada. You choose a model on that server or a hosted model under a written zero data retention agreement, and a hosted model may process requests outside Canada. Derik Lawlis, the founder, leads every project and stays close to the build. About ThriveAI covers the company.