AI & MLAI transformation: how do you get from strategy to measurable business results?
What AI transformation is, how to plan it step by step, where to start, how to measure impact, and why companies stall between the pilot and scale.
Author: Łukasz SomniakPublished Updated 8 min read
Summary
- AI transformation means redesigning core business processes so that AI does a real share of the work and people oversee the result. Simply rolling out a tool is only deployment.
- AI is everywhere, value is not: in McKinsey's survey 88% of organizations use AI, but only about 6% attribute more than 5% of EBIT to it.
- According to BCG, about 70% of AI implementation challenges come from people and processes, 20% from technology, and only 10% from algorithms. Your transformation plan should follow the same proportions.
- Start with 2-3 pilots, each with a business metric and an owner, and decide on scaling based on measurement, not slide decks.
- Build an honest business case on data from your own pilot. An average "AI ROI" figure from the internet is no substitute for measuring your process.
Quick answer
AI transformation is redesigning a company's core processes, from customer service through operations to management decisions, so that AI does a real share of the work and people oversee the result. It starts with a strategy and 2-3 pilots tied to business metrics, and it scales only what measurably works. Most companies stall in the middle: according to McKinsey, 88% of organizations regularly use AI, but only about 6% attribute more than 5% of EBIT to it (McKinsey, 2025).
Why is AI no longer optional?
Because it has become a standard work tool, and the advantage no longer comes from access to AI but from the ability to get value out of it. In the same McKinsey survey, 39% of respondents report any EBIT impact from AI at the enterprise level, and in most cases it is less than 5% of EBIT (McKinsey, 2025).
BCG paints a similar picture: 74% of companies have yet to show tangible value from AI, and only 26% have built the capabilities to move beyond proof of concept (BCG, 2024). Over three years, the leaders in that study achieved 1.5 times higher revenue growth than other companies.
Adoption is also uneven across markets. In 2025, 20% of EU enterprises with 10 or more employees used AI technologies, and in Poland, where we are based, only 8.4% (Eurostat, 2025). Many companies are still choosing their first use cases, which is a good moment to do it methodically.
How is AI transformation different from deploying a tool?
Deployment gives people a tool. Transformation changes the process the tool works in. A company-wide license for an AI chat is deployment. A process where a photographed document fills in a form by itself and a person only checks the result is transformation.
The data shows the difference. McKinsey's "AI high performers" are nearly three times more likely to fundamentally redesign workflows when deploying AI instead of adding a tool to the old process (McKinsey, 2025).
BCG frames it as a 10-20-70 rule: roughly 10% of AI implementation challenges are about algorithms, 20% about technology and data, and 70% about people and processes (BCG, 2024). If 90% of your transformation plan is choosing models and infrastructure, the proportions are upside down.
| Deploying an AI tool | AI transformation | |
|---|---|---|
| Starting point | A new tool on the market | A problem in a specific process |
| What changes | Access to the tool | How the work is done, and roles |
| Success metric | Licenses, logins | Process time, quality, cost before and after |
| Who owns it | IT | The process owner, with leadership |
| Typical risk | A tool nobody uses | Too much scope at the start |
What does the AI transformation path look like?
A two-phase path works best: first a process-based strategy, then pilots and scaling with decision points.
Phase 1: a strategy that fits your company
An AI strategy starts with processes, not technology. We talk to the people who do the work every day, map the data and tools, and only then choose the solution: an application, an integration, automation, and AI where it genuinely helps.
Steps in this phase:
- Readiness assessment: what data is available, in what quality, who owns it, and what legal constraints apply (GDPR, trade secrets, the EU AI Act).
- Use case selection: a list of candidates scored on value, feasibility, and risk. Pick 2-3 for pilots.
- A business case per use case: measure what the process costs today (time, errors, delays) and agree on what pilot result would justify scaling.
- Roadmap: the order of use cases, dependencies (data, integrations, infrastructure decisions), and review dates.
Data is easy to forget, and it most often decides the outcome. Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data (Gartner, 2025). Checking this before you start means auditing the quality, availability, and access rights for the data behind the process you picked for the pilot.
Phase 2: pilots, scaling, and adoption
A pilot should prove value in a real process, not impress in a presentation. Design it like a small production system: with an owner, a business metric, system integration, and data access rules from day one.
This is where most projects drop out. In an IDC study, for every 33 AI pilots only 4 reached production (IDC for Lenovo, 2025). What changes that outcome includes a clear business metric, system integration, data access rules, and an engaged process owner from day one.
Elements of this phase:
- Pilots: short, with decision points: scale, pivot, or stop.
- Infrastructure and security: whether the model runs through an API or on your own servers (see our comparison of on-premise vs. cloud AI), permissions, logs, data rules.
- Change management: communication, training on your own examples, champions in each team, without which even a good tool goes unused.
- Scaling: more processes and teams, but only after a confirmed result.
- Oversight and optimization: regular reviews of quality, cost, and risk, with clear ownership of who runs them.
What does it look like in practice? The Expander example
AI transformation at Expander, a financial advisory firm, is a series of concrete solutions, each addressing one problem in advisors' daily work. We have been working there since 2025.
- Reading documents from a phone photo. The system detects the page edges, corrects the perspective, reads the text, and moves values into form fields. A person checks the result instead of retyping it. Reading a note takes 12 seconds instead of more than 8 minutes.
- A voice-driven mobile app. Speech becomes structured data, and before saving, the app shows what it understood. Speech recognition and language models run on our own GPUs.
- Training recordings generated from text. Speech synthesis turns a script into a voice-over, so fixing the material means editing text, not booking another studio session.
- Simple apps without AI. Merging reports from several sources and matching bank transfers to accounting records, with doubtful cases routed to a person.
The last point matters. AI transformation does not mean every problem is solved by a language model. Sometimes a plain application delivers the best result, and AI comes in where it genuinely saves time.
How do you choose tools and technologies?
Match the technology to the process and the data, not the other way around. The model market changes every few months, so an architecture that lets you swap models matters more than any specific name.
| Area | Typical solutions | When to choose |
|---|---|---|
| AI assistants for teams | Business editions of AI chat tools and assistants built into office suites | Text work, drafts, summaries, when data may go to the vendor |
| Company knowledge | RAG, a model that answers from your documents, with citations | Questions about procedures, contracts, documentation |
| Documents | OCR and vision models, field extraction into your systems | Invoices, forms, notes, scans |
| Process automation | Integrations, RPA, AI agents with human approval | Repetitive tasks across several systems |
| Tailored models | Open models on your own infrastructure, fine-tuning | Data that cannot leave the company, specialized language |
The choice between an off-the-shelf tool and your own solution usually comes down to models tailored to your data, the subject of our guide to fine-tuning open-source models. We start every project from our own components (an agent core and a knowledge base engine), and we run models, embeddings, and speech recognition on our own GPU infrastructure when client data cannot go to external APIs.
What business impact can you expect?
The impact you measure in your own process. Published "AI ROI" averages come from different industries, definitions, and methods, so they do not belong in a specific company's budget. An honest business case is built on pilot data.
Value from AI usually shows up in three places:
- Time: shorter process cycles, less manual retyping and searching for information.
- Quality: fewer errors, more consistent documents, answers with a citation to the source.
- Scale: more cases handled by the same team, without proportional headcount growth.
How to measure so the result is credible:
| Stage | What to measure | How |
|---|---|---|
| Before the pilot | Baseline: task time, error count, cost | Measurement on a sample of real cases |
| During the pilot | The same metrics plus tool usage | Logs, comparison with the baseline |
| After rollout | Business result of the process and running cost | Quarterly review, cost per task |
Remember the AI-side costs: token fees or hardware, maintenance, quality monitoring, and the time of people who review outputs. ROI without these items is inflated.
Where should you start?
With the one process that hurts the most, and with measuring what it costs today. A simple plan for the first quarter:
- Pick 2-3 pilot candidates: repetitive tasks with available data and a clear owner.
- Measure the baseline and agree on what result would justify scaling.
- Check the data and legal constraints before you choose the technology.
- Run the pilot in a real process, with the people who do the work.
- Decide based on measurement: scale, pivot, or stop.
Organizations that win with AI don't experiment forever. They pick concrete problems, measure results, and scale what works.
Sources
- McKinsey: The state of AI in 2025: Agents, innovation, and transformation (PDF, November 2025)
- BCG: AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value
- Eurostat: 20% of EU enterprises use AI technologies (2025)
- Gartner: Lack of AI-Ready Data Puts AI Projects at Risk (2025)
- IDC for Lenovo: CIO Playbook 2025, It's Time for AI-nomics
Frequently asked questions
How long does AI transformation take?
A first pilot with a measurable result can usually be run in a few weeks if the data and the process are accessible. Transforming a whole company takes years and never formally ends, because models, tools, and processes keep changing. Plan in quarters with decision points rather than promising a single end date.
How much does AI transformation cost?
There is no honest average, because cost depends on the number of processes, the state of your data, integrations, and whether models run through an API or on your own infrastructure. The cost includes people (analysis, implementation, training), system integrations, model fees or GPU hardware, and maintenance. The safest path is a pilot with a fixed budget and success criterion, then estimating the cost of scale from pilot data.
Do I need my own AI department?
Not at the start. You need a business owner for each use case, someone responsible for the data, and a technical partner or small team to build and maintain the solution. An in-house AI team makes sense once AI runs in several processes and maintaining it becomes ongoing work.
What are the most common mistakes in AI implementation?
The most common ones are deploying AI without a concrete business problem, running a pilot without a metric or an owner, skipping data preparation, and treating changes to how people work as an add-on to a technical project. The common thread is investing in the model and too little in the process and the people.
Can a small company go through AI transformation?
Yes, and it is often easier: faster decisions, fewer legacy systems, one process that hurts the most. Start with a single use case with a measurable effect, such as reading documents or searching company documentation. In the EU, only 20% of enterprises with 10 or more employees used AI in 2025 (Eurostat), so many companies are still at the starting line.

