AI reporting in manufacturing: From Excel to a report that explains everything on its own

Most manufacturers keep their data in an ERP and their reports in Excel. In between sit people who copy, join and check that data every month. Jan Papoušek, Data & AI Solutions Architect at FLO, walks through our project with Povrly Copper Industries (PCI) and shows how we replaced that work with an agentic system that builds the report itself and stands ready for any follow-up question.

In brief
  • We were looking for a way to let AI take over the slow, manual preparation of an existing report.

  • We started with one report that hits the bottom line directly: material consumption in production against the technological plan.

  • We had to gather the right data in good quality. The harder and more important part was explaining the company’s context: everything an in-house analyst could recite in their sleep. Only then could we get output that was accurate and usable.

  • The numbers in the reports come from a fixed, deterministic calculation. AI proposes how the report looks, writes the commentary and interpretation, and answers questions. Every number in an AI comment is checked against the calculated values.

  • You can ask any chart what its numbers mean. And if you don’t like how a chart looks, you can simply ask the AI to change it.

Manufacturing reporting is still manual work

Production data is there in most companies. It sits in the ERP, in master data and in all kinds of exports. What’s usually missing is an efficient way to get from that data to a decision.

In many companies, the regular report is still made the same way. Someone in controlling exports data from the ERP, copies it into Excel, matches it by hand against the product and machine master data, calculates the KPIs on a separate sheet and checks the result. Copying and joining tables is where most errors creep in. And because the job takes several days, management often hears about a problem only at the meeting after month-end close. By then the plant has used up another month of expensive material.

How the report looks matters as much as the numbers. A production manager needs to see at a glance whether the plan is being met and, if not, where and why. At the start of the project, PCI’s head of controlling put it precisely:

“Nobody on the shop floor will read a dashboard full of paragraphs, or understand it if they do.

Head of Controlling, PCI

The third condition is trust. When AI writes that a variance has grown, a controller will rightly ask which numbers that’s based on, for which period, and whether it matches the management report. Until those questions get answers, AI won’t take hold in reporting.

Standard business reporting tools (Power BI, Tableau and others) can already produce charts that refresh themselves. AI adds two things. The report gets from a business brief to the screen much faster, and the user gets a plain-language explanation of any data selection, plus an answer to any follow-up question.

Povrly Copper Industries (PCI) is a Czech manufacturer that processes copper. We have worked together since autumn 2025, and the project has run in two phases so far.

Phase one delivered an assistant for management

The first brief came from the company’s management, who wanted an AI-based personal assistant for managers. The assistant drew on the monthly management report and on company documents. It answered questions and linked every answer to its source. We built this classic “chat with your data” system on top of our SmartScout solution. From the monthly figures, the assistant also put together an interactive overview that opened with a management summary and showed the main KPIs and the most significant adverse trends, ranked by how much they cost the company. The client accepted the results of phase one, and management got to see its own numbers from a new angle.

The assistant, however, worked with a finished report. It saved no work for the controlling team who put that report together every month, and putting it together was what took them the most time. At the kick-off meeting for phase two, the client sharpened the brief. They wanted reports that fill themselves from the data, so controlling wouldn’t have to touch them. AI was to add what a person would otherwise have to dig out of the data: where a trend has turned, what the likely cause is, and a warning before the monthly meeting comes round.

Phase two focused on a single report

At the kick-off meeting, we proposed automating several reports at once. PCI’s finance manager narrowed our proposal down to one:

“I’d rather have one report fine-tuned down to the last detail than four done halfway.”

Finance Manager, PCI

Several areas were on the table: material charge, energy, non-conformities and material input ratios. The client chose the input ratio report. The input ratio expresses how much material has to go into a production step to get one kilogram of product out of it. Each step has a planned input ratio set by its technology. When the actual ratio is higher, the company is losing material, and with copper that’s serious money. The client picked this report because it could be automated with reasonable effort and promised the biggest improvement to the bottom line.

The report was to show the actual and planned input ratios, the gap between them and its financial impact. The figures had to be viewable by day, week, month and year, with drill-down from machine to product to an individual batch. Until then, the report was made by hand. Batch data and master data were copied from the ERP into Excel, the KPIs were calculated on a separate sheet and the result was checked. Most of the errors and delays came from exactly that copying and from joining the master data to the batch data.

Our approach rests on four principles

None of them is specific to input ratios. They work just as well for any regular report.

1. A brief that both the client’s people and the AI understand

Every report we build starts with a formal brief in six parts. We ask first what the report is for, and only then about the data:

  • Who works with the report and what decisions it informs.

  • How the report is made today, and where mistakes or lost time happen most often.

  • Which KPIs it contains and exactly what they mean.

  • Which data the KPIs are calculated from, and how each one maps to that data.

  • What the output should look like, and how far down the drill-down should go.

  • What role the AI should play, and what stays in human hands.

What matters most is that for every KPI it’s clear which data it’s calculated from and what that data means on the shop floor. That’s exactly where an AI would most easily fill a gap with a guess, so we want it in black and white before we start.

Because PCI knew very precisely what report it wanted, the controlling team gave us the brief, and our team used AI to turn it into the form the agents work with. The brief doesn’t have to be written entirely by hand, though. If a company isn’t yet sure what report it wants, anyone can put the brief together with the AI, right in the interface, even from a one-sentence request. From the available data, the AI proposes KPIs, views and the links between them, asks about whatever it can’t read from the data, and people then correct and complete the proposal. If you have the data and know what it means, you can have a relevant, verifiable and interactive report in minutes instead of weeks.

2. Code calculates the numbers, AI explains them

If I had to pick one decision from the whole project to recommend to everyone, it would be this division of roles:

  • A fixed calculation, meaning plain code that never changes by itself, computes all the numbers, charts, filters and comparisons. The same data always gives the same result.

  • AI works out from the brief which data the report needs and how it will be laid out. The proposal is then checked to make sure it uses only approved chart types and calculations.

  • AI writes the commentary. It works from the calculated summaries, extreme values and trends over time. Every number it quotes in a comment is automatically compared with the calculated values. If they don’t match, the comment is rewritten.

Once the report is fine-tuned, it is locked. At each daily refresh, the fresh data is simply slotted in and the AI updates its commentary to match.

3. Commentary on what’s in front of you right now

A typical report comes with one AI comment on the whole thing, explaining everything in it. But a plant manager looks at their own machine, and a controller at a specific material. So with one click the commentary rewrites itself for whatever the user has filtered, whether that’s a machine, a product, a material or a period.

You can also ask any chart a question directly. The AI sees what’s on the screen and knows the report’s brief, its rules and its data. When it needs more, it looks into the source data, read-only and within set limits. The answer therefore always rests on the same numbers the user has in front of them. And if you don’t like how a chart looks, you ask the AI to change it. This is how PCI’s head of controlling drilled all the way down to individual batches.

4. Every number can be checked

A controller only trusts a number they can check themselves. So these rules apply:

  • The report drills down from the overall view through machine and product to an individual batch.

  • Every chart comes with a description of how it’s calculated.

  • Data can be exported to Excel, either as displayed or as the underlying rows. (Yes, even AI products can’t escape the Excel export.)

  • Data is taken from the ERP unchanged and recalculated only as specified in the brief.

The client ran the toughest test themselves. Controlling compared the new report with the management report using the same filter settings, and the numbers matched. The only difference was the financial valuation of variances, which was more accurate in the new report.

Results
  • The input ratio report replaced manual data assembly in Excel. The data refreshes every day without controlling having to step in.

  • The client signed off both stages of phase two in writing.

  • Controlling introduced the tool to the plant managers themselves. They send users a round-up of what’s new and have gradually opened access to more colleagues in production and controlling.

  • The new report has become the yardstick other reports are measured against. PCI’s head of controlling explained it to colleagues in production like this:

“With the same filters, the numbers match. The only thing that doesn’t is the financial valuation of variances, and that’s more accurate in the online report. From next month we’ll align it so it matches the online report one to one.”

Head of Controlling, PCI

The project gave us the foundation of a system where we only need to add further data areas, and each one becomes available to everyone as genuine self-service BI. In this kind of self-service, all you need to know is what you want to show, and how to describe it. You don’t need to understand relational databases, aggregation levels or the rules about which data belongs in which type of chart, as you do with classic BI tools.

What we learned
  1. One report finished properly beats five half-built ones. A pilot on a single report with a clear financial impact gave us room to fine-tune the details that trust depends on: sign conventions, calculation method and financial valuation.

  2. Trusted data first, AI second. The finance manager put it plainly: the data has to be right first. So we started with data the company already trusts, and compared the new report with the existing management report.

  3. A regular report has to be stable. In the first version, the AI rebuilt the whole report at every refresh. For a report people open every day, it matters more that the same data always gives the same result. So we handed the calculation to fixed code and left the AI with the proposal, the explanations and answering questions.

  4. Don’t wait for perfect data access. Connecting to the ERP takes time and IT effort. Because we started on a sample of the data, the client saw the first report before the connection was finished.

  5. A report needs its own champion at the client. The turning point came when controlling presented the report to production themselves and started developing it as their own tool.

Where next

The same approach fits anywhere a company regularly builds a report from data it trusts:

  • Inventory and warehouses. Where money is tied up in excess stock, which items risk running out before the next delivery arrives, and what the company buys again only because it already has the same thing under a different code.

  • Financial and accounting statements that still run on Excel.

  • Production data, especially downtime and production times.

  • Order profitability.

This kind of AI reporting answers the questions a company knows how to ask, in the areas it knows it wants or needs to watch.

The next step, the advanced course if you like, is to have the system proactively look for areas the company hasn’t thought of yet. That’s what our other tool, Thesis, is for. Working on the same data, it proposes hypotheses about what needs improving, ranked by estimated financial impact. It attaches the supporting evidence to each hypothesis, and the decision to confirm or reject it is yours. AI Reporting and Thesis both run on our AI agent platform, which we deploy into a separate environment for each client.

Let’s talk about it at Smart Industry in Pilsen

If your regular report still gets built by, say, copying from the ERP into Excel, or you can’t find the time to analyse your data and look for the value it’s leaving on the table, come and see us at the FLO stand at the Smart Industry conference in Pilsen on Thursday 12 November 2026. We’ll show you what our solution looks like, explain in detail how it works, and work out how we can help you in this area.

Written by Jan Papoušek, Data & AI Solutions Architect at FLO.

Go top

Our growth and expansion are backed by Rockaway, one of the largest investment groups in the CEE region.

© 2026 FLO Group s.r.o., VAT: CZ11983311, Rohanské nábřeží 678/27, 186 00, Prague, Registered in the Commercial Register maintained at the Municipal Court in Prague, File No: C 357494