Guide: AI in business › AI in sales and business departments
Artificial intelligence in manufacturing delivers results fastest in four areas: image-based quality control using cameras, predicting machine failures, demand forecasting and planning, and document handling in the ERP system – provided the company has well-organised data from its machines and systems. In Poland this is still rare. According to Eurostat, in 2025 8.4% of companies with at least 10 employees used AI, compared with 20% across the EU. This guide shows where to start, what is already built into popular ERP systems and which regulations apply to machinery with AI.
In 60 seconds AI on the shop floor and in ERP in brief
- What it involves: models learn from machine and quality control data, sales data and the ERP in order to forecast, detect errors and automate decisions.
- Where it works: in manufacturing plants and distribution companies that have machine data or several years of ERP history and recurring problems such as shortages, downtime or complaints.
- Cost range: from features included in the ERP licence to projects with sensors and an integrator; in SAP and Microsoft some AI features are billed based on consumption.
- First project: one problem with a measurable cost (e.g. downtime on one line) and checking whether data exists that describes it.
- Date to remember: from 20 January 2027, the new EU Machinery Regulation covers machinery with self-learning systems, so machine manufacturers need to prepare now.
What is artificial intelligence in manufacturing and in ERP?
Artificial intelligence in manufacturing means software that learns from data from machine sensors, cameras, the quality system or order history, and on that basis predicts, detects or suggests a decision. Robots that “think” are not the subject here. It is about forecasts and classifications such as: this machine will probably stop within a week, this part is defective, orders for this product will rise in November.
ERP (Enterprise Resource Planning) is the system in which a company manages orders and warehousing, production, purchasing and finance. Artificial intelligence in ERP has two faces. The first is classic analytics – forecasts, detection of unusual transactions and order suggestions – and the second, newer one is assistants and agenci AI, which you instruct in plain language (“show orders delayed by supplier X”), and they look up the data or carry out several steps in the system.
Manufacturing is just one of the departments where AI is changing day-to-day work today. For a fuller picture, from sales to marketing and PR, see our guide to AI in business.
Six applications where AI in manufacturing delivers the most
The best results come from applications where the problem is frequent and costly, and at the same time well described in the data. Here are the six most common, along with what each one requires. Demand forecasting works in much the same way in retail, as you can see from the example of AI in e-commerce.
| Application | What AI does | What data is needed | Typical pitfall |
|---|---|---|---|
| AI quality control | Detects defects in images from cameras on the line | Thousands of images of good and defective parts, labelled by people | Too few examples of rare defects |
| Predictive maintenance | Predicts failure from vibration, temperature and current | Sensor history and records of past failures | No record of breakdowns, so the model has nothing to learn from |
| AI demand forecasting | Predicts orders, taking seasonality and promotions into account | Several years of sales history from the ERP | Sudden market shifts that never appeared in the history |
| AI production planning | Suggests the order of jobs to cut changeover times | Operation and changeover times, availability of machines and people | ERP data that does not match what happens on the shop floor |
| Purchasing and inventory | Suggests what to order and when | Stock levels, delivery times, consumption | Manual stock adjustments outside the system |
| Documents and finance | Reads invoices and warehouse documents, matches payments | Documents in electronic form, e.g. from KSeF | Treating the extracted data as infallible |
What do SAP, Microsoft and Comarch offer in artificial intelligence?
The largest ERP vendors are now building AI into their systems. They differ in scope and in how they bill for it, and when it comes to budget the latter is often more important than the former. Below is the situation as of September 2026, based on vendor documentation and descriptions.
Jak z tego wybierać? Jeśli pracujesz na SAP, SAP Joule jest pierwszym miejscem do sprawdzenia, bo działa na danych, które już masz w systemie. Firmy na systemie Microsoftu zaczynają zwykle od funkcji Copilot w Business Central, a użytkownicy systemów polskiego dostawcy od narzędzia ChatERP Comarch. W każdym przypadku najpierw sprawdź, które funkcje masz w obecnej licencji, a za które zapłacisz według zużycia.
| System | AI assistant | What it can do | How it is billed |
|---|---|---|---|
| SAP (cloud, incl. S/4HANA Cloud) | Joule | Answers questions about data, guides users through transactions, drafts documents; agents for tasks such as maintenance planning | Core features included in the cloud subscription; premium features consume SAP AI units purchased separately (ERP Research, July 2026) |
| Microsoft Dynamics 365 Business Central | Copilot and agents | Suggestions, summaries, agents handling selected processes | Some features have no consumption charges, others are billed by Microsoft based on consumption; the billing type is shown in the system (Microsoft Learn, August 2026) |
| Comarch ERP | ChatERP and AI agents | Searching and filtering data with text commands, invoice reading (OCR), verifying business partners in KSeF, anomaly detection | Depends on the product and contract; details from the vendor |
When do AI in ERP and in manufacturing make sense?
AI will not fix a process that does not work without it. It makes sense where the process works but is expensive, slow or dependent on the experience of a handful of people, as in these situations:
- A planner who builds the schedule in a spreadsheet because the ERP “doesn't understand” changeovers – and nobody else can do it while they are on holiday.
- A quality department that inspects hundreds of parts per shift by hand and lets defects slip through when tired.
- Maintenance that reacts to breakdowns instead of preventing them.
- A sales team that forecasts “by eye” every month, after which production catches up with overtime (here, help also comes from CRM with AI).
- Accounting and purchasing retyping data from documents, which we cover in more detail in the article on AI in company finance.
Jak Commwork®AI pokazuje, gdzie AI w zakładzie da wynik
W firmie produkcyjnej dane leżą w ERP, w systemie jakości i w arkuszach planisty, więc trudno ocenić, gdzie AI się opłaci. Commwork®AI łączy programy, które już masz, i buduje jeden kokpit z właścicielem przy każdym wskaźniku. Zespół Commplace® przygotowuje warianty z wyliczeniem, a decyzję podejmuje człowiek.
- Data integration 38%
- Sensors and cameras 22%
- Plant staff time 18%
- Partner wdrożeniowy 14%
- AI features in ERP 8%
- 24.07.2026Omnibus cyfrowy przesuwa część obowiązków AI Act
- Week 3Mapa danych gotowa, koszt problemu policzony
- Tydzień 5Pilotaż na jednej maszynie, porównanie z bliźniaczą
- Week 12Pomiar skutku i decyzja o kolejnych maszynach
- 20.01.2027Stosowanie rozporządzenia maszynowego UE 2023/1230
- 2.08.2028Obowiązki wysokiego ryzyka dla AI w produktach
- 01Mapa danych i programów
- 02Wybór jednego problemu
- 03Pilotaż z porównaniem
- 04Measuring the impact
- 05Scaling
- Łączy ERP, system jakości i arkusze, które firma już ma, bez kupowania kolejnego programu, i porządkuje dane, zanim ktokolwiek zacznie na nich budować model.
- Przy decyzji o pilotażu przygotowuje symulację wariantów i naradę pięciu modeli z zaznaczonym sporem; gdy modele się nie zgadzają, decyzja wraca do człowieka.
- Zapisuje w rejestrze decyzji, kto, kiedy i na jakich danych uruchomił pilotaż, a po kwartale pokazuje jego skutek. Jak takie wdrożenie zaplanować, opisujemy w 90-day AI implementation plan.
The view reproduced is modeled after the dashboards provided to customers. The names, numbers, and terms are examples and show the way the system works, not the results of a specific company.
How to implement AI in manufacturing step by step
Projekty AI w przemyśle rzadko upadają przez algorytm. Upadają przez złe dane, brak właściciela po stronie zakładu i brak porównania z tym, co było wcześniej. Porównanie zaplanuj od pierwszego dnia: wybierz dwa albo trzy mierniki z listy, którą podajemy w tekście z przykładami KPI, i zapisz ich wartość sprzed pilotażu.
- STEP 01Choose one problem with a quantifiable cost
Downtime on one line, the complaint rate for one product, excess stock of one group of materials. Write down what it costs today, because that figure will be your baseline.
- STEP 02Check the data before you choose a tool
Are there failure records? Do ERP stock levels match the warehouse? Are defect images labelled? Without data, even the best model will predict nothing.
- STEP 03Start with what you already have in your ERP
Check which AI features your system has and how they are billed, because sometimes the first step does not require a new purchase, just enabling and configuring a feature the company is already paying for.
- STEP 04Appoint an owner on the shop floor
A maintenance, quality or planning manager who will assess the AI’s suggestions and decide what to do with them. Without them, the model becomes a report nobody reads.
- STEP 05A pilot with a control group
Several months on one line or product group, compared with a line without AI or with the same months of the previous year. Correlation is not impact. A drop in downtime in summer may simply be due to a lower workload.
- STEP 06Scaling and maintenance
After a successful pilot, you scale up the project and decide who will monitor the quality of the model when machines, raw materials or the product range change.
How much does artificial intelligence in manufacturing cost and what drives the budget?
There is no single price for “AI for the factory”. The budget is made up of several items, and their proportions depend on whether you use features built into your ERP or build your own solution.
Próg decyzji ustal przed pilotażem. Przykład: model utrzymania ruchu przechodzi dalej, jeśli jego ostrzeżenia potwierdzają się na hali w większości przypadków, a liczba nieplanowanych postojów na maszynie z pilotażem spada wyraźniej niż na podobnej maszynie bez niego. Jeśli ostrzeżeń jest dużo, a potwierdza się mało, ludzie przestaną je czytać po kilku tygodniach.
Przy rozmowie z integratorem albo dostawcą ERP zadaj pytania, które wyciągają koszty ukryte za słowem „wdrożenie”.
Gdy nie wiesz, który problem policzyć najpierw, zacznij od diagnozy: Commplace AI agency wskaże proces i wyliczy jego koszt, zanim zamówisz czujniki czy licencje.
- Które funkcje AI mam w obecnej licencji, a które kosztują dodatkowo?
- Ile będzie kosztowało zużycie przy naszej liczbie użytkowników i transakcji?
- Kto po wdrożeniu utrzymuje model, gdy zmieni się asortyment albo maszyna?
- Jak porównamy wynik z maszyną lub linią bez pilotażu?
- Jakie dane zostają u nas, jeśli zakończymy współpracę?
- Ile godzin ludzi z zakładu potrzebujesz i w którym tygodniu?
| Item | What it depends on | Notes |
|---|---|---|
| AI features in ERP | System version, licence edition, billing model | In SAP some features consume AI units; in Business Central some are billed by consumption |
| Sensors and cameras | Number of machines and workstations, age of the machine fleet | Older machines often need retrofitting |
| Data integration | Number of systems (ERP, manufacturing execution system, quality, warehouse) | Usually the largest item in the project |
| Integrator or implementation partner | Scope of work, industry experience | Compare offers by how they measure results |
| Plant staff time | Labelling data, assessing suggestions, training | Often left out of the budget |
| Model maintenance | Changes to the product range, machines, raw materials | An unattended model loses accuracy |
New rules for manufacturers of machinery with AI
For machine manufacturers the key date is 20 January 2027. From that day the EU Machinery Regulation 2023/1230 applies, replacing the Machinery Directive. In its September 2024 overview, PARP (the Polish Agency for Enterprise Development) points out that the new rules introduce the concept of machinery with evolving behaviour, i.e. equipped with self-learning systems. The manufacturer must assess the risks arising from changes the system may introduce into how the machine operates, and safety-function software with machine learning requires new conformity assessment procedures.
The second act is the AI Act. AI systems built into products covered by sector-specific legislation may be classified as high-risk systems, and the amending Regulation (EU) 2026/1744, the so-called digital omnibus, published on 24 July 2026, postponed these obligations to 2 August 2028. For machinery the change goes further: according to a Eurogip overview from August 2026, AI performing safety functions in a machine is assessed primarily under the Machinery Regulation, and AI requirements are to be added to it by a delegated act. The 20 January 2027 deadline has not changed. We track the current state of the law in our overview of The AI Act in Poland. Machine manufacturers should plan for compliance now, because designing a new machine takes longer than the time that is left.
Sztuczna inteligencja w produkcji maszyn to więc nie tylko temat dla działu badań i rozwoju, ale też dla działu prawnego i dokumentacji. Rozporządzenie maszynowe a AI to w praktyce trzy pytania do zespołu konstrukcyjnego: czy maszyna uczy się w trakcie pracy, czy ta nauka może zmienić funkcje bezpieczeństwa i kto oceni zgodność. Jeśli choć na jedno odpowiedź brzmi „tak” albo „nie wiemy”, zaplanuj przegląd przed 20 stycznia 2027 roku. Czasu jest mało.
Data matters more than the model
Zarząd najczęściej pyta, który system AI wybrać. Lepsze pytanie brzmi: czy nasze dane są na tyle dobre, żeby jakikolwiek system coś z nich wyczytał? Sprawdź, czy stany i czasy w ERP zgadzają się z rzeczywistością, czy zdarzenia na hali (awarie, przezbrojenia, braki) trafiają do jednego miejsca i czy dane mają jednego właściciela. Gdy dane o zamówieniach i awariach leżą w osobnych programach, Commwork®AI cockpit łączy je w jeden widok, w którym każdy wskaźnik ma źródło, datę, próg i właściciela. Dopiero na takim widoku widać, czy model ma z czego się uczyć.
AI assistants in ERP, such as Joule or ChatERP, answer based on the data in the system. When the data is wrong, the answer sounds convincing but is still wrong – and we describe exactly this phenomenon when discussing AI hallucinations. Where knowledge from documents is needed (manuals, process sheets, procedures), what works well is the approach we describe in the article on AI on company knowledge. For everyday analysis outside the ERP, many teams still rely on spreadsheets – and you can find out what AI can do there today in the article on AI in Excel.
How a manufacturing company can talk about AI – and what not to promise
Implementing AI is also a communication matter, because B2B customers ask about quality and on-time delivery, candidates about a modern workplace, and the media look for concrete stories from industry. A story about AI works when it has measured figures and is honest about what does not work yet. It does harm when it promises a “factory of the future” without proof. In B2B, one untrue sentence in a proposal can cost you a contract.
Komunikację zmiany zaplanuj razem z projektem. Ustal, co powiesz załodze (AI nie zastępuje ludzi, tylko pomaga w konkretnych zadaniach), co klientom, a co mediom branżowym. Obawy ludzi na hali są naturalne, a sposoby prowadzenia ich przez nowe opisujemy w artykule o tym, jak zarządzać zmianą w organizacji. Podstawy budowania wiarygodnego wizerunku opisujemy w tekście o tym, how to build a brand image, and you can read about how Commplace® supports industrial companies on the page industrial marketing.
Typowy zakład w 90 dni i lista kontrolna przed pilotażem
Ten scenariusz to przykład poglądowy, a nie opis wdrożenia u konkretnego klienta. Załóżmy zakład z trzema liniami, systemem ERP od kilku lat i jedną maszyną, której postoje najbardziej blokują wysyłki.
Dni 1–30. Zespół spisuje, gdzie leżą dane: stany i zlecenia w ERP, awarie w zeszycie utrzymania ruchu, wady w arkuszu jakości. Wszystkie działy, od produkcji po finanse, uzgadniają jedną definicję przestoju. Ktoś liczy, ile kosztował przestój tej maszyny w ostatnim kwartale, bo bez tej liczby nie będzie z czym porównać wyniku. Brzmi nudno, ale oszczędza miesiące.
Dni 31–60. Na maszynie pojawiają się brakujące czujniki, a model zaczyna ostrzegać. Nikt jeszcze nie zmienia planu przeglądów. Tak ma być na starcie. Mistrz zmiany zapisuje przy każdym ostrzeżeniu, czy coś się faktycznie działo. Tylko tyle. Załoga słyszy wprost, po co jest pilotaż i czego nie zmienia; dobre zasady takiej rozmowy opisuje tekst o przepływie informacji w firmie.
Dni 61–90. Zespół porównuje maszynę z pilotażem z podobną maszyną bez niego i przegląda ostrzeżenia, które się nie sprawdziły. Decyzja ma trzy możliwe wyniki: rozszerzyć na kolejne maszyny, poprawić dane i przedłużyć pilotaż albo zamknąć temat. Każdy z nich jest dobrym wynikiem, jeśli opiera się na pomiarze, a nie na wrażeniu; różnicę między zbieżnością a wpływem wyjaśniamy na stronie o korelacji i wpływie.
Najdroższą pozycją w takim kwartale zwykle nie jest licencja, tylko integracja danych i czas ludzi z zakładu, dokładnie jak w tabeli kosztów wyżej. Prognozowanie popytu da się dołożyć później, gdy dane o zamówieniach będą spójne z planowaniem; od strony handlowej pomaga w tym tekst o sales planning.
- Czy mamy jeden problem z policzonym kosztem w złotówkach?
- Czy zapis awarii i wad z ostatniego roku istnieje w jednym miejscu?
- Czy wszystkie działy liczą przestój tak samo?
- Czy mamy maszynę lub linię do porównania, której pilotaż nie obejmie?
- Czy wiadomo, kto na hali odpowiada za pilotaż i ile ma na to godzin?
- Czy sprawdziliśmy, które funkcje AI są już w naszej licencji ERP?
- Czy dział konstrukcyjny wie, czy nasze maszyny podlegają nowym przepisom?
When to implement AI in manufacturing and when to wait
A good moment is when you are planning next year's budget and during a quieter production season, because a pilot needs time from plant staff, and at the peak of orders that time is not there. For machine manufacturers, the calendar is set by regulation. Only a few months remain until 20 January 2027.
It is a good time if you have a recurring, costly problem, several years of data and someone who will lead the project on the plant side. Wait if you are in the middle of replacing your ERP (new system first, then AI), if data sits in spreadsheets on private drives, or if nobody can say what the problem AI is supposed to solve costs today. When choosing a contractor, use the questions from the article AI software house or AI agency, and if you are considering an assistant in your office suite for administrative teams, take a look at the guide on Copilot for business. Jeśli chcesz, żeby ktoś z zewnątrz wskazał proces z największym potencjałem i policzył go przed pilotażem, zobacz, jak pracuje nasza AI agency.
Five ways to make an AI project at a plant stall
A company buys an AI platform “because the competition has one” without specifying what should improve. It pays for licences and integration, but has no result it could show the board.
A model trained on incorrect stock levels and incomplete breakdown records produces forecasts nobody trusts, so after a few months the team goes back to spreadsheets.
The project is run by IT alone, and the production manager finds out about it by email. AI suggestions are ignored because nobody is responsible for the decisions.
AI agents in the ERP are rolled out to all users without a cost simulation. The bill exceeds the entire saving from the pilot.
A machine manufacturer adds self-learning features without a new risk assessment and, after 20 January 2027, runs into problems with conformity assessment and CE marking.
What manufacturing company boards ask us
How is artificial intelligence used in manufacturing?
Most often for camera-based quality control, predicting machine failures, demand forecasting, production planning and reading documents in the ERP system. Each of these applications requires data the model can learn from.
What is artificial intelligence in ERP systems?
These are features built into business management systems that forecast, detect unusual data and respond to instructions written in plain language. Examples include Joule in SAP, Copilot in Microsoft Dynamics 365 Business Central and ChatERP in Comarch systems.
How many companies in Poland use AI?
According to Eurostat, in 2025 8.4% of Polish companies with at least 10 employees used AI technologies. The EU average was 20.0%, and Poland was among the countries with the lowest share.
How much does implementing AI in a factory cost?
It depends on whether you use features built into your ERP or build your own solution with sensors and data integration. The biggest costs are usually data integration, retrofitting machines and plant staff time, and some AI features in ERP are billed based on consumption.
Does the new Machinery Regulation apply to AI?
Yes. Regulation (EU) 2023/1230, applicable from 20 January 2027, covers machinery with self-learning systems and requires an assessment of the risks associated with changes in how it operates. Safety-function software with machine learning is subject to new conformity assessment procedures.
Where should you start with AI in manufacturing?
Start with one problem with a quantifiable cost, for example downtime on one line. Then check whether you have data that describes it, and run a pilot compared with a period or line without AI.
Will AI replace the production planner?
It will not. AI suggests the order of jobs and warns about problems. The decision is made by a person who knows the constraints of the shop floor, the people and the customers, so it is best treated as support for an experienced planner.
Sources
- SOURCEEurostat — 20% of EU enterprises use AI technologies · 11 December 2025
- SOURCEPARP — The new Machinery Regulation: key changes for businesses · 2 September 2024
- SOURCE
- SOURCEERP Research — SAP Joule explained: what it does, cost and limitations · reviewed 22 July 2026
- SOURCEMicrosoft Learn — Business Central: understand how each Copilot and agent capability is billed · updated 27 August 2026
- SOURCEComarch — Artificial intelligence (AI) in Comarch products · accessed 24 September 2026
- SOURCEEUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI), Polish text · of 8 July 2026, OJ EU of 24 July 2026, accessed 24 September 2026
- SOURCE
Read next
Zaczynamy od diagnozy: mapujemy dane w ERP i na hali, liczymy koszt jednego problemu i planujemy pilotaż z grupą porównawczą. Pomagamy też opowiedzieć o zmianie załodze i klientom.
Sebastian Kopiej, President of Commplace®. In public relations since 1996. Text prepared with the help of AI tools and fact-checked by the author. Data current as of September 29, 2026.