From Excel to AI — How to Implement Artificial Intelligence When You Don’t Have Big Data

Most mid-sized companies in Poland have exactly the same problem. They read about the AI revolution, see case studies of corporations operating on petabytes of data, then look at their spreadsheets and think: “this isn’t for us.” That is one of the most expensive technology myths of 2026 — and this article takes it apart.

The truth is that most practical AI applications in SMEs do not require big data. They require data — yes. But “data” is not a synonym for “millions of records in a data lake.” Data is your sales spreadsheets, order history in Excel, a mailbox with three years of correspondence, or a CRM with 2,000 customer records.

This guide shows, step by step, how to go from a spreadsheet to a production AI system — realistically, without marketing hype, and with concrete numbers on minimum data requirements.

The article is for boards and COOs of SMEs — people who make budget decisions and need hard arguments rather than promises. You will find tables of minimum data requirements, an implementation roadmap with a timeline, a case study with concrete ROI figures, and an honest look at situations where AI genuinely does not make sense.

The Big-Data Myth — Why Companies Think They Cannot Afford AI

Myth: AI requires millions of records and a data lake

The technology industry has a communication problem. Conferences, webinars, articles — everywhere the talk is of “big data,” “data lakes,” “petabytes of training data.” OpenAI trains GPT on trillions of tokens. Tesla collects data from millions of cars. Netflix analyses billions of user interactions.

And suddenly the owner of a 60-person manufacturing firm, who has four years of order history in Excel (about 8,000 rows), concludes: “AI is a technology for the big players. We don’t have enough data.”

That is a classic cognitive error — survivorship bias in technology communication. The media write about giants because that impresses. Nobody will write an article titled “A company in Radom deployed a predictive model on 1,500 records and lifted retention by 12%.” And that is exactly the story we tell in the case-study section below.

Where does the myth come from? Three sources. First — the marketing of large cloud platforms (AWS, Azure, GCP), which sell large-scale data-processing services. The more data you process, the more you pay — so their educational materials naturally emphasise the need for “more data.” Second — technology media, which cover AI breakthroughs that require huge datasets because that is simply more interesting than “a company used Random Forest on 2,000 records.” Third — consultants themselves, who, when selling large data-transformation projects, have an interest in building the belief that “first a data lake, then AI.”

Truth: many AI applications work on hundreds or thousands of records

Let us distinguish two worlds of AI:

World 1 — Foundation models and generative AI. Yes, here you need petabytes. But you are not training GPT from scratch. You use a ready-made model and adapt it to your needs (fine-tuning, RAG, prompt engineering). Fine-tuning needs 200–500 examples. RAG needs your existing knowledge base.

World 2 — Classical ML and prediction. Regression, classification, clustering, time-series forecasting. Here the absolute minimum is often 200–500 records for simple models, and a comfortable level starts at 1,000–2,000 records. Most SMEs that have been operating for 3+ years have many times more data than they need.

The mental problem: “we only have Excels” = self-exclusion

The phrase “we only have Excels” carries the assumption that Excel is something inferior — a format “for the poor.” From a data-science perspective, a well-structured Excel spreadsheet is an excellent source of training data.

In our consulting practice we see this pattern at least three times a month. A CEO or COO says: “We would like to implement AI, but we don’t have the right data infrastructure.” We ask: “What do you have?” The answer: “Well… Excels, a CRM, some data in the ERP.” And then comes the sentence that changes the perspective: “That’s more than enough.”

Format does not matter. CSV, XLSX, Google Sheets, an ERP export — to the algorithm it is all the same. What matters:

Are the data structured (columns = features, rows = observations)?
Do you have enough history (at least 12 months for time-series forecasts)?
Do the data contain a target variable (what you want to predict)?

If the answer to those three questions is “yes” — you have a sufficient foundation to implement AI. You do not need Snowflake, Databricks, or a data lake for PLN 500,000.

Paradoxically, companies with “only Excels” have one advantage over corporations with data lakes: simplicity. Their data sit in one place, in a comprehensible format, maintained by people who know the business context. You do not need weeks mapping hundreds of tables in a warehouse or resolving conflicts between sources. A sales spreadsheet is a sales spreadsheet — it needs no translation.

How Much Data Is Enough to Implement AI?

The table below is the minimum data requirement for the most popular AI applications in SMEs. Note — we are not talking about millions of records. We are talking about volumes that most companies already have in their existing systems.

AI application	Minimum data volume	Source format	Preparation time
Sales forecast	12 months of history (min. 365 data points)	Excel/CSV with dates and values	1–2 days
Document classification	200+ labelled examples (min. 50 per category)	Files + labels in a spreadsheet	3–5 days
Internal FAQ chatbot	50+ question–answer pairs	Google Docs, wiki, knowledge base	2–3 days
Customer churn prediction	1,000+ customer records with history	CRM/Excel with customer features	3–7 days
Anomaly detection	500+ normal observations	System logs, sensor data	1–3 days
Report automation	Templates + data sources (API/DB)	ERP, CRM, spreadsheets	2–5 days
Lead scoring	500+ historical leads with outcomes	CRM with conversion history	3–5 days

Concrete examples from practice

Sales forecast — a distribution company in Kraków had three years of sales data in Excel. 1,095 daily entries × 12 product categories. That was enough to build an ARIMA + XGBoost model that forecasts demand with 87% accuracy 30 days ahead. Result: a 23% reduction in over-ordering.

Document classification — a Warsaw law firm (35 people) had 4,000 documents in topical folders on a network drive. The folder names themselves served as labels. 4,000 documents × 8 categories = more than 500 examples per category. The model classifying new documents reached 91% accuracy.

FAQ chatbot — a Wrocław SaaS company exported 180 of the most frequent questions from Freshdesk together with agent answers. After cleaning and deduplication, 120 Q&A pairs remained. That was enough for a RAG chatbot that handles 65% of Tier 1 queries without a human.

5 AI Applications That Work on Small Datasets

1. Demand forecasting (Excel with orders)

What you need: Order history of at least 12 months — dates, quantities, optionally product categories and seasonality data.

How it works: Time-series algorithms (Prophet, ARIMA, ETS) and gradient boosting (XGBoost, LightGBM) handle “date → value” data extremely well. They do not need millions of points — they need enough history to capture seasonal patterns.

Minimum data: 365 daily observations or 52 weekly ones for seasonal models. If you have 2–3 years of history — that is a luxury situation.

Typical ROI: 15–30% reduction in warehousing costs, 20–40% reduction in stockouts. For a company with PLN 10 million/year in turnover, that is PLN 150–300k saved annually.

Real-life example: A Poznań building-materials wholesaler (70 employees) exported order data from the last 28 months — 840 daily records × 45 product categories. A Prophet model with added regressors (weather, construction season, holidays) achieved 11% MAPE on a 14-day horizon. Result: supplier orders generated automatically, warehouse overstock down 27%, stockouts down 34%.

2. Email and document classification (inbox + labels)

What you need: A set of documents or emails grouped into categories. Minimum 50 examples per category, ideally 200+.

How it works: NLP models (both classical TF-IDF + SVM and fine-tuned transformers) learn from labelled examples to recognise new documents. In 2026, fine-tuning a small language model on 200 examples delivers results that five years ago would have required 10,000 samples.

Minimum data: 200 documents in total (min. 50 per category with 4 categories). Source: folders on disk, email labels, tags in a DMS.

Typical ROI: 2–4 hours of administrative work saved per day. At an employee cost of PLN 8,000/month that is PLN 40,000–80,000/year.

Advanced variant: Automatic routing of customer emails to the right departments (complaint → quality, RFQ → sales, invoice → accounting). A Gdańsk logistics company processes 400 emails a day this way — previously one employee sorted them for 3 hours; now AI does it in 2 seconds at 94% accuracy.

3. Internal chatbot (knowledge base in Google Drive)

What you need: The company’s knowledge base — procedures, FAQs, process documentation, instructions. Format does not matter: Google Docs, Confluence, SharePoint, PDFs, even recorded and transcribed conversations.

How it works: RAG (Retrieval-Augmented Generation) indexes your knowledge base and then answers employee questions, citing specific source documents. It does not train a new model — it uses an existing LLM enriched with your data.

Minimum data: 50 documents or 50 Q&A pairs. The more content, the better the answers, but even 30 pages of documentation is a solid start.

Typical ROI: 40% reduction in onboarding time for new employees, 60% reduction in HR/IT load from repetitive questions. For a 50+ person company with 15% annual turnover — savings of PLN 50,000–100,000/year.

4. Report automation (data from ERP/CRM)

What you need: Access to data sources (API or export) plus the report templates you currently build by hand.

How it works: AI does not only automate extraction and formatting — it generates analytical commentary, highlights anomalies, compares with previous periods and suggests actions. That is a step beyond classical BI — the report “thinks” instead of merely presenting numbers.

Minimum data: Here it is not about the volume of training data, but about access to sources. Defining 3–5 pilot reports with clear data sources is enough.

Typical ROI: Elimination of 8–15 person-hours per week of manual reporting. At the average cost of an analyst that is PLN 80,000–150,000/year in savings plus faster reaction to business change.

5. Anomaly detection (system logs)

What you need: Data describing the “normal” state of a system, process or operation. Anomaly-detection algorithms learn the pattern of normality and flag deviations.

How it works: Unsupervised models (Isolation Forest, Autoencoder, DBSCAN) do not need “anomaly / no anomaly” labels. They need only a sufficient sample of normal behaviour — a minimum of 500 observations. Anything that significantly departs from the pattern is reported as an anomaly.

Minimum data: 500+ “normal” observations (e.g. 500 logs from a fault-free period). Zero required labels — that is the method’s greatest advantage.

Typical ROI: Early detection of production problems, fraud, infrastructure failures. In a manufacturing company — 25–40% reduction in downtime, which at a downtime cost of PLN 5,000–20,000 per hour yields savings of PLN 200,000–500,000/year.

Roadmap: From Excel to Production AI in 4 Steps

Step 1: Data inventory (what you have, where, in what format)

Before you think about algorithms — map your data. That is the foundation without which every AI project is built on sand.

What you map:

What data does your company generate (sales, orders, customers, production, HR, finance)?
Where is it stored (Excel, Google Sheets, ERP, CRM, mailboxes, network drive)?
In what format (structured tables vs unstructured documents)?
How deep is the history (how many months/years back)?
Who owns the data (who updates it, who has access)?

Duration: 3–5 working days.

Output: A “Data Inventory” document — a map of all data sources with an assessment of their usefulness for AI. In our audit process (Process Intelligence Audit) this is part of the Discovery phase.

Typical findings: Companies usually have 3–5× more useful data than they think. Data “hidden” in mailboxes, CRM notes, Slack history or system logs is a gold mine the team had forgotten.

Most common inventory omissions:

Emails with customers (communication history, sentiment, contact frequency)
Salespeople’s notes in the CRM (free text — NLP features can be extracted)
Spreadsheet change history (Google Sheets tracks edits — that is process data)
System logs (ERP, WMS, e-commerce — every operation leaves a trace)
IoT device data (sensors, gates, terminals — even simple data on operating times)

Step 2: Cleaning and structuring (ETL lite)

Raw data rarely go straight into modelling. They need “ETL lite” — a lightweight extract, transform and load process.

Typical problems to solve:

Missing values (empty cells in Excel)
Inconsistent formats (dates recorded in 5 different ways)
Duplicates (the same customer under 3 different names)
Typos in categories
Ambiguous labels

Tools: Python (pandas), OpenRefine, or even advanced Excel formulas to start. The goal is not perfection — it is “good enough” for the first model.

Duration: 5–10 working days (depends on the state of the data).

Golden rule: Do not clean all the data at once. Clean only what is needed for the first use case. The rest you do iteratively.

Practical tip: Do not chase perfection. In our audit practice we see companies that spend 6 months “tidying data” before starting an AI project — and never reach modelling. The 80/20 rule works perfectly: 80% of the value comes from the first 20% of work on data quality. The rest is optimisation you do iteratively once you have a working model and know which features have the greatest impact on the result.

Step 3: PoC on 1 use case (2–4 weeks)

The Proof of Concept is the moment of truth. You pick one business problem, build a model on cleaned data and measure the results.

Criteria for choosing the first use case:

High ROI potential (at least PLN 100,000/year in savings or revenue)
Available data (from step 1 you know what you have)
Measurable outcome (a clear success metric: accuracy, time, cost, conversion)
Stakeholder support (someone on the board wants this result)

Duration: 2–4 weeks for a PoC with clear results. If after 4 weeks you do not have a “it works / it doesn’t” answer — something is wrong with the scope.

Output: A PoC report with metrics, a comparison to baseline, and a go/no-go recommendation for production.

What to avoid at the PoC stage:

Choosing a use case “because it is technically interesting” instead of “because it delivers the highest ROI”
Building ideal infrastructure before validating the hypothesis
Testing on all the data without a train/test split (overfitting)
Reporting accuracy without business context (82% accuracy means nothing without a baseline comparison)
Involving too many stakeholders — a PoC is a 2–3 person team, not a committee

Step 4: Productionisation + monitoring

A PoC is not production. A model in a Jupyter Notebook is not an AI system. Productionisation means:

Infrastructure:

Deploying the model on a server/cloud (API endpoint)
Data pipeline (automatic feeding of new data)
Monitoring of accuracy and data drift
Alerting (when the model starts losing quality)

Processes:

Retraining schedule (weekly? monthly?)
Alert escalation (who reacts to a drop in accuracy?)
Technical and business documentation
SLA (response time, availability, expected quality)

Duration: 4–8 weeks from a completed PoC to full production.

This is the moment when many companies need operational support — they either build an internal MLOps team or use managed-operations services such as QCare.

The most important lesson from our deployments: Productionisation is not “deploy and forget.” An AI model is a living organism — it needs feeding with new data, quality monitoring and periodic retraining. Companies that treat AI like traditional software (you deploy once and it runs for years without change) systematically lose value from the investment. The maintenance plan should be created at the PoC stage, not after go-live.

How Much Data Do You Really Need? Minimum Viable Dataset Table

The table below is concrete numbers — the minimum datasets at which different types of AI models deliver valuable results. The “comfortable level” is the volume at which the model achieves stable, repeatable results.

Model type	Min. records	Min. features (variables)	Comfortable level	Example application
Linear regression	100	3–5	500+	Project cost forecast
Decision trees / Random Forest	300	5–10	1,000+	Customer scoring
XGBoost / LightGBM	500	5–20	2,000+	Churn prediction
Neural networks (tabular)	1,000	10–50	5,000+	Dynamic pricing
Time series (Prophet/ARIMA)	365 points	1–3	730+ (2 years)	Sales forecast
NLP — text classification	200 documents	N/A	1,000+	Email categorisation
NLP — LLM fine-tuning	200 examples	N/A	500+	Specialist assistant
Anomaly detection	500 (normal)	3–10	2,000+	Fraud detection
Clustering	200	3–10	1,000+	Customer segmentation
RAG (chatbot)	50 documents	N/A	200+ documents	FAQ chatbot

Key observation: Most SMEs operating for 3+ years have data well above the minimum level. The problem is not data volume — it is awareness that this data is sufficient, and the skill to prepare it.

How to read this table: If you run a company with 2,000 customers in the CRM, 14 columns of data on each customer and 3 years of history — you meet the requirements for XGBoost / LightGBM with a large margin. You do not need more data. You need someone who will turn those data into features and train a model.

A common mistake: Companies compare their data with the requirements for neural networks (deep learning) and conclude “too little.” But deep learning is not the only AI. Classical ML algorithms — random forest, XGBoost, logistic regression — are still the workhorses of the industry and work excellently on sets of 500–5,000 records. In many benchmarks, gradient boosting (XGBoost, LightGBM) beats neural networks on tabular data, regardless of dataset size.

Case Study: A 45-Person Services Firm — From Excel to Churn Prediction

Starting situation

A services company in Łódź, 45 employees, PLN 12 million/year in revenue. It operates in B2B — providing maintenance services to corporate clients. The problem: 20% annual churn among customers, each lost customer worth on average PLN 45,000 in lost annual revenue.

The board knew it was losing customers, but did not know why — or, more importantly, which customers it would lose next quarter. The reaction was always after the fact: a customer left, the sales team called with a retention offer, but it was already too late. They needed an early-warning system — and they had the data to build it. They just did not know that it was enough.

The data they had

Three Excel spreadsheets maintained since 2023:

Customer base — 2,000 records (active + former customers) with: industry, size, location, start-of-cooperation date, assigned account manager
Invoice history — 8,500 rows: customer, date, amount, service type, payment punctuality
Complaints register — 340 tickets: customer, date, problem category, resolution time, satisfaction

In total: 3 Excel files, 10,840 rows of data, zero data lake, zero cloud.

What they did (the process)

Weeks 1–2: Data preparation

Joining 3 spreadsheets into one dataset per customer (feature engineering)
Features: average invoice value, order frequency, value trend (rising/falling), number of complaints, complaint response time, length of cooperation, industry, size
Target variable: “did the customer leave within the next 6 months?” (binary)
Result: 2,000 records × 14 features

Week 3: Modelling

Test of 4 algorithms: logistic regression, random forest, XGBoost, LightGBM
Validation: stratified 5-fold cross-validation
Best model: XGBoost — 82% accuracy, 78% precision, 76% recall

Week 4: Business validation

Presentation of results to the board with interpretation (SHAP values)
Top churn predictors: falling order-value trend, rising number of complaints, lengthening time between orders
Decision: go for production

Weeks 5–8: Productionisation

Deploy on AWS (simple API endpoint)
Weekly scoring of all active customers
Dashboard in Google Data Studio with “high-risk customer” alerts
Business process: the account manager contacts a high-risk customer within 48 hours

Results after 6 months

Metric	Before AI	After AI	Change
Churn rate (annual)	20%	8%	−12 pp
Customer retention	80%	92%	+12 pp
Lost revenue/year	PLN 1,800,000	PLN 720,000	−PLN 1,080,000
Additional net revenue	—	+PLN 180,000/year	(after deducting implementation and maintenance costs)
Implementation cost (one-off)	—	PLN 85,000	—
Maintenance cost (annual)	—	PLN 24,000	—
ROI (first year)	—	—	112%

Key takeaway: 3 Excel files with 3 years of history were enough to build a system that generates PLN 180,000 of additional revenue per year. No data lake, no in-house data-science team, no GPU budget — an 8-week project with an external partner was enough.

What was critical to this project’s success:

A clearly defined target variable (left / did not leave within 6 months)
Sufficient history — 3 years of data meant they had both “leaving” and “loyal” customers to train the model
Feature engineering — simply joining 3 spreadsheets and computing trends produced 14 informative features
Board support — the CEO personally sponsored the project and required weekly results reporting
Fast time to value — 8 weeks from the first meeting to a working system, with ROI appearing in month 5

When the Data Really Is NOT Enough (An Honest Assessment)

Not every situation is suitable for AI. It would be dishonest to claim that “Excel is always enough.” Here are situations in which data can be a real barrier:

Fewer than 50 examples in classification

If you have a classification problem (e.g. “is this document an invoice or an order?”) and fewer than 50 labelled examples per category — classical ML will probably fail. In 2026 there are workarounds (few-shot learning with LLMs, transfer learning), but results will be unstable and hard to validate.

Solution: Before you invest in a model — invest 2–3 weeks in labelling data. 50 examples per category is typically 2–4 hours of work for a person who knows the domain.

No history (greenfield)

A startup in its first month of operation, a new product with no sales history, entry into an entirely new market. If you have no historical data — predictive models have nothing to learn from.

Solution: Start collecting data from day 1. Define the structure (what you measure, how you record it) and after 3–6 months you will have a sufficient foundation. In the meantime you can use rule-based automation and generative AI, which do not require historical data.

Unstructured data without labels

You have 10,000 photos from a production line, but none is labelled “defect” / “OK.” You have 5,000 documents, but nobody knows which category they belong to. Raw data without labels is raw material — but not yet fuel for supervised learning.

Solution: Three paths — (1) invest in labelling (crowdsourcing or an expert), (2) semi-supervised / self-supervised learning techniques, (3) unsupervised models (clustering, anomaly detection), which do not require labels but give less precise results.

Data too homogeneous (no variance)

Sometimes the data exist in sufficient quantity but are too homogeneous — all customers look similar, all transactions have similar values, there are no clear patterns to discover. An AI model needs variance — differences between observations — to learn anything.

Solution: Add new data sources (external data: weather, macroeconomics, industry data) or feature engineering — creating new features from existing ones (e.g. trends, ratios, lag values). Sometimes it is enough to change granularity — use daily data instead of monthly.

Low-quality data (dirt, gaps, inconsistencies)

In theory you have 5,000 records. In practice — 40% have missing values in key fields, 20% contain obvious errors, and formats changed 3 times over 2 years. After cleaning you are left with 2,000 usable records — but that can still be enough.

Solution: ETL lite (Step 2 of the roadmap). A realistic assessment of “how much remains after cleaning?” is a key metric of a data audit.

FAQ — The Most Common Questions About Implementing AI with Small Data

Is Excel a sufficient data source for AI?

Yes — provided the data is structured (columns = features, rows = observations) and you have enough history. File format (XLSX, CSV, Google Sheets) is irrelevant — ML algorithms operate on numeric matrices, not files. Excel is one of the most common data sources in AI projects for SMEs, and there is nothing inferior about it compared with data from a dedicated warehouse. Moreover, Excel data has one advantage — it is maintained by people who understand the business, so columns have sensible names and values correspond to real processes. In corporate data warehouses that is often worse.

How much does AI implementation cost in a company that only has spreadsheets?

A typical SME budget for a first AI project (from data audit to a PoC in production) is PLN 50,000–150,000. That covers: data audit (PLN 5–15k), data preparation and modelling (PLN 20–60k), productionisation (PLN 20–60k), and the first 3 months of monitoring (PLN 5–15k). ROI typically appears within 6–12 months.

Do I need to hire a full-time data scientist?

Not at the start. The first AI project is better delivered with an external partner — faster, cheaper and less risky than building competence from scratch. A full-time data scientist makes sense when the company has 3+ models running in production and a clear roadmap of further deployments. Until then, managed operations (outsourcing AI maintenance) is better economics.

For context — hiring an experienced data scientist in Poland in 2026 costs PLN 18,000–28,000 gross/month (plus benefits, training, tools). That is PLN 250,000–400,000 per year. For that amount you can deliver 2–3 full AI implementations with an external partner and keep maintenance in a managed-operations model. A full-time hire only makes sense at a volume of 4+ AI projects per year — which in a 50–200 person company is rare in the first 2 years of transformation.

How long does AI implementation take from scratch in a company with no infrastructure?

From the first meeting to a working system in production: 8–16 weeks. Breakdown: data audit and use-case selection (1–2 weeks), data preparation (2–3 weeks), PoC and modelling (2–4 weeks), productionisation (3–6 weeks). This is not a year-long project — it is a quarter-long project.

Important caveat: 8–16 weeks is the timeline for one use case with ready data. If the data needs significant cleaning or the company first has to digitise paper processes — add 4–8 weeks of preparation. Even so — from decision to business value is at most 6 months, not years.

Is AI from small data less accurate than corporate AI?

Not necessarily. Model accuracy depends on data quality, feature selection and algorithm fit — not only on volume. An XGBoost model trained on 2,000 clean, well-chosen records often beats a neural network trained on 100,000 dirty records. In SME practice we achieve 75–90% accuracy depending on the problem — enough to generate real business value.

It is worth remembering that in a business context what matters is not absolute accuracy, but the lift versus the status quo. If today your salespeople pick at-risk customers with 40% accuracy (intuition-based guessing) and a model from 2,000 records delivers 80% — that is a doubling of effectiveness that translates into real money. You do not need 99% accuracy — you need a sufficient edge over the current way of making decisions.

Summary — Your Data Is Enough, Start Today

The main takeaway from this article is simple: if your company has been operating for 2+ years and keeps any digital records — you have enough data to implement AI. You do not need big data. You do not need a data lake. You do not need a team of 10 data scientists.

The big-data myth is convenient — it lets you postpone the decision to implement AI until “later, when we have more data.” But that “later” never arrives, and the competition that started from its Excels builds an advantage every month. The companies that implemented AI first in their niche did not have better data — they had a better decision.

You need:

Awareness — that your Excels, CRM and mailboxes are valuable data sources
An audit — a professional assessment of which data are suitable for AI and which use case will deliver the fastest ROI
A partner — who will take you from spreadsheet to production in 8–16 weeks, without building infrastructure for hundreds of thousands of zloty

Next step: Process Intelligence Audit

Don’t know where to start? Our Process Intelligence Audit is exactly that first step. Within 2–3 weeks we will:

Inventory your data and assess its quality
Score AI readiness (data readiness score) on a 1–10 scale
Identify 3–5 use cases with the highest ROI for your industry
Deliver an implementation roadmap with concrete numbers, a timeline and a budget
Indicate which data you need to clean further, and which are ready “as is”

Every company we have worked with had “only Excels.” Every one of them implemented AI.

Stop waiting for data you “don’t have yet.” Start with what you already have. Book a free consultation — we will show you what your spreadsheets can do for you.

Article updated: August 2026. All figures, amounts and examples reflect the state of the AI market for SMEs in Poland at the time of publication. If you are reading this more than 6 months after the publication date — contact us for current benchmarks and minimum data requirements for the latest models.