AI reporting: ask for the report your ERP doesn’t produce

AI reporting means asking for a report in plain language and getting back a table built from your own records. You type a question such as “which customers bought less this quarter than in the same quarter last year?” The AI writes the database query, runs it against data from your ERP and accounting system, and shows the query and the rows it used. A person checks the result before it is shared. It suits the reports your ERP doesn’t produce, the ones that otherwise take an export to a spreadsheet and hours of someone’s time. This guide covers how AI reporting works, how to check its accuracy, who should see what, scheduled reports and where to start.

Derik Lawlis Founder, ThriveAI

What AI reporting is

A report answers a business question with numbers drawn from records: sales by customer, late orders by supplier, margin by product line. AI reporting lets the person with the question produce the report. They type the question, and the system returns a table with the evidence behind it. Nobody has to know which tables hold the data or how they join.

The questions it suits cross the boundaries of a standard report. Here are four from businesses that make, move or sell goods.

Each question combines records from several parts of the ERP, and some need data held outside it, such as a rebate list in a spreadsheet. Ad hoc reporting covers how one-off questions like these differ from the standard reports an ERP runs every week.

How custom reports get made today

A report your ERP doesn’t produce has to be built by someone. Without AI, it takes one of three routes.

• An export and a spreadsheet. Someone exports two or three lists from the ERP, matches them with lookups and pivots the result. The answer is right for that export. Next month the steps are repeated by hand, and small differences creep in, such as a filter left on or a new customer missing from a lookup range.
• A report writer. ERPs include tools for building reports and analyzing lists without code, covered under ERP reporting below. Someone still has to know which records hold the data.
• A consultant or IT. The ERP reseller or a developer writes a custom report. It runs the same way every time, and every change to the question goes back into their queue.

Each route depends on someone who knows where the data sits and how the tables join. When that person is busy or away, the question waits.

How AI reporting works

An AI reporting system worth trusting follows six steps. The technique at its centre is text-to-SQL: a language model translates a question into SQL, the query language that databases use.

1. You ask in plain language. “Which customers bought less this quarter than in the same quarter last year?”
2. The AI reads your definitions. Before writing anything, it looks up what your business means by “bought”: invoiced or shipped, before or after credit notes. Snowflake explains why this step exists in its documentation for Cortex Analyst: database schemas “lack critical knowledge like business process definitions and metrics handling.”
3. It writes the query against the tables that hold your invoices, credit notes and customers.
4. It runs the query with read-only access, on a copy of the data or through a read-only connection, so a badly written query cannot change a record.
5. It shows its work. The answer comes back with the query, the rows it used and a restatement of the question. Snowflake’s sample response begins “We interpreted your question as” and then gives the SQL. Microsoft says that with its Fabric IQ option turned on, Copilot in Power BI can “provide the underlying DAX query and data table used to generate a response,” where DAX is the query language Power BI uses.
6. A person checks it before it is shared. The manager who asked reviews the restated question, the filters and a few rows, then shares the report or saves it.

Steps 2 and 5 matter most. Written definitions mean two people who ask the same question get the same answer. The visible query shows exactly what was counted, so a wrong filter is easy to spot.

Text-to-SQL is a different tool from retrieval-augmented generation (RAG), which finds the passages in your documents that answer a question. RAG suits contracts and manuals. A report needs exact totals over every matching row, and a database query provides them.

AI reporting tools come in four forms: assistants built into BI tools such as Power BI and Tableau, AI features of data platforms such as Snowflake and Databricks, analysis features inside ERPs, and custom systems built on a reporting database. AI business intelligence compares them and covers the data they need.

How accurate AI reports are, and how to check them

Text-to-SQL has improved quickly, and it still makes more mistakes than a careful analyst. Two public research benchmarks show where it stands.

• BIRD pairs more than 12,000 questions with SQL answers over 95 databases. On its test set, data engineers and database students answered 92.96% correctly. The best AI system on the BIRD leaderboard, posted September 26, 2026, answered 82.95%.
• Spider 2.0 uses databases from real enterprise systems, often with more than 1,000 columns. On its Spider 2.0-Lite set of 547 questions, the top entry on the Spider 2.0 leaderboard completed 76.23% of the tasks when we checked on October 4, 2026.

If your ERP is full of custom fields and codes that only make sense inside your company, expect results closer to the harder benchmark. Vendors give the same warning about their own products. Microsoft’s overview of Copilot in Power BI says that without preparing the data, “Copilot can misinterpret the data and return generic or inaccurate results.”

Check every new kind of question before you rely on the answer. These five checks cover the common errors.

The repeat test is the one that scales. Databricks lets a team store up to 500 benchmark questions with known answers for its Genie tool, and compares each generated result with the correct one. It recommends two to four phrasings of the same question, because people word the same request differently. Microsoft gives similar advice for Power BI: define the expected measure, grouping, date context, filters and result for each test question, then compare. AI evals covers building that test set from the reports you already produce.

For the questions your team asks every week, a fixed answer is more dependable than a new query each time. Power BI’s verified answers work that way: a “human-approved” visual that Copilot returns “instead of generating a new response” when a question matches.

Who sees what: permissions

A report can show anything the system behind it can read. If the AI connects to your ERP with an administrator’s account, any employee who asks can see every customer’s margin and every supplier’s prices.

The safe design runs each query as the person asking. Microsoft’s MCP server for Business Central performs all operations “with your user identity and permissions,” and Snowflake says the queries Cortex Analyst generates “adhere to all established access controls.” An MCP server is one way to give AI tools this kind of access. Ask any vendor how its tool applies the permissions in your ERP, then test it by asking questions from a junior employee’s account.

Check the exceptions as well. Power BI’s row-level security limits the rows each user sees, but Microsoft notes that it “only restricts data access for users with Viewer permissions.” Workspace Admins, Members and Contributors see every row, so a sales rep given Member access to share a report sees every territory.

Reports also leave the system. A scheduled report goes to whoever is on its distribution list, so review the list when the report is set up and whenever someone changes roles.

ERP reporting: what your ERP gives you, and where it stops

ERP reporting is the reporting built into the ERP: standard reports, saved searches, list views with filters and a report writer for custom layouts. Vendors keep extending it. Oracle groups NetSuite’s tools under SuiteAnalytics, which includes NetSuite Analytics Warehouse, Workbook, searches, reports and dashboards. Business Central’s analysis mode lets users summarize and pivot list data “directly from the page, without having to run a report,” and its analysis assist feature turns a plain-language description into “a suggested layout as a starting point.”

These tools work well for questions inside one system. A question reaches their limit when it needs one of these:

• Several modules at once, such as the quote, the order and the production hours for the same job.
• Data outside the ERP, such as rebate agreements in a spreadsheet or freight bills in a separate accounting system.
• History from an old system, kept after a migration in a database nobody reports on.
• More than one company, when each legal entity has its own ERP database.

Without a reporting database, reports that need these get rebuilt by hand each month. AI reporting reaches them by querying one database that holds every source. AI for ERP covers the ways to connect to an ERP, and legacy ERP automation covers older systems with no API.

Report automation: from a checked answer to a scheduled report

Report automation means running a report on a schedule and sending it to the people who need it, with nobody rebuilding it. ERPs and BI tools already do this for standard reports. In Business Central, the job queue runs reports once or on a recurring schedule, and Microsoft’s example is a weekly sales report by salesperson. Power BI subscriptions email a snapshot of a report or dashboard “on a schedule you set.”

AI adds a step before the schedule. Once a manager has asked a question and checked the answer, the system saves the query itself, with its definitions and filters. The scheduled report reruns that saved query each week without asking the model again. It runs the same way every time, and anyone can read what it counts.

Two rules keep scheduled reports healthy:

• Alert on the run as well as the result. A report that returns no rows, or far fewer than last week, can mean a broken connection or a renamed field. Send that alert to the report’s owner before the report goes to the distribution list.
• Give every report an owner and a review date. Each quarter, retire the reports nobody opens.

AI dashboards apply the same idea to the numbers people watch every day, and AI in finance covers the month-end reporting a finance team produces.

What data AI reporting needs

AI reporting runs on data you already keep. The questions in this guide draw on four sources:

• The ERP: orders, shipments, inventory, purchasing and production.
• The accounting system, if it is separate from the ERP: invoices, credit notes, costs and the general ledger.
• Spreadsheets that hold price lists, rebate agreements, budgets and targets.
• Written definitions of the terms people use: what counts as a sale, how margin is calculated, which date makes an order late.

The data should sit in one reporting database that the AI can query without touching the live systems, refreshed on a schedule. The definitions should sit beside it, where every report reads them. dbt, a tool for building that layer of definitions, says AI tools connected to it answer with “your governed metrics instead of guessing at raw tables.”

Where to start

The safest way to start using AI for reporting is with questions you already know the answers to.

1. Collect ten questions. Ask managers which reports they build by hand or wait for, and keep the ten that come up most.
2. Keep last month’s answers. Save the spreadsheet behind each one. Together they become your test set.
3. Write the definitions those ten questions depend on, and agree on them with whoever closes the books.
4. Copy the data read-only. Load the ERP and accounting tables those questions need into a reporting database, refreshed nightly.
5. Test before you trust. Ask the ten questions, compare the answers with last month’s, and fix definitions until they match.
6. Pilot with two managers for a month, then schedule the reports they keep asking for.

The same reporting database supports the next projects. Margin analysis uses the same sales and cost lines, and AI demand forecasting uses the same order history.

Risks and limits

• Wrong answers that look right. A wrong filter produces a clean table with wrong numbers. The checks above, and a person reviewing before anything is shared, are the defence. Human in the loop covers how to design a review that people actually carry out.
• Definitions that drift. When the business changes how it counts something, such as moving freight out of cost of goods sold, update the definition and rerun the test set.
• Stale data. A report is only as current as the last refresh. Print the refresh time on every report.
• Where the data goes. The question and the rows it returns pass through an AI model, which may run in another country. Ask every vendor where prompts and results are processed and stored. Data sovereignty and private AI for business cover the options.

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 every answer shows where it came from: each number links to the document or ERP record behind it. When the data cannot answer a question, the system says what is missing instead of guessing.

ThriveAI’s systems read your ERP as it is, 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.