AI Implementation in Business: How to Design a Process, Not Just Buy a Tool

August 24, 2026 · Bartłomiej Spasowski, Founder, modulla · EN

AI Implementation in Business: What It Is and Why Technology Alone Is Not Enough

AI implementation in a business is a multi-dimensional process of transforming an operational model, not a one-time software purchase. Technology is merely an execution tool. Real value emerges when an organization rebuilds its workflows, changes its decision-making habits, and redefines human roles in the processes that AI takes over or supports.

Why Most AI Implementations End in Disappointment

The data is unambiguous: buying a tool alone is not enough. According to McKinsey (2025) [11], as many as 95% of generative AI pilots produce no measurable impact on a company's financial results. BCG reports that 70% of AI-based transformations fail to deliver the expected value [17]. RAND Corporation reports that AI projects fail twice as often as traditional IT projects [21].

This is no accident. Let's be direct: companies treat AI like a new version of an ERP system. They buy a license, launch a pilot, and wait for results. The results never come, because not a single internal process has changed.

S&P Global Market Intelligence reports that by mid-2025, as many as 42% of companies had abandoned most of their AI initiatives [21]. On average, organizations abandon 46% of their proofs-of-concept before they reach production [21].

The barriers are predictable and mostly non-engineering in nature:

Automating chaotic, poorly documented processes (result: faster chaos) [19]
Lack of quality data for training and feeding models [19]
Employee resistance to changing habits (Harvard Business School calls this the "innovation curse") [22]
Unmanaged proliferation of AI tools (AI sprawl): 78% of employees use unauthorized AI applications [4]
Lack of a change management plan, not a lack of technology budget [22]

In the Polish context, this picture is even clearer. According to the 2024 Human Capital Balance (BKL) study published by PARP and the Jagiellonian University, only 23% of Polish companies use AI in their business processes, and most implementations are improvised, with no long-term strategic plan [58].

Three Strategic AI Implementation Modes According to McKinsey

Rather than treating all AI initiatives the same, companies achieving the best results categorize projects by their transformational depth. McKinsey calls this model DRI: Deploy, Reshape, Invent [9].

Deploy: Quick Wins on Existing Processes

Deploying ready-made SaaS tools onto existing workflows without modifying them. The goal is to build AI proficiency within the team and gather the first measurable results. Operational risk: low. Return: moderate [9].

Reshape: Redesigning Operational Functions

This is where the greatest impact on financial results is concentrated [9]. The company does not merely add AI to an existing process; it redesigns the entire workflow around what AI can do autonomously or support. Examples: customer service, finance, sales, HR. This mode is precisely what distinguishes companies building competitive advantage from those merely buying tools.

Invent: New Business Models

Creating products and revenue streams that would not exist without AI. Highest risk, highest potential. The right stage for organizations with a mature data layer and proven competencies [9].

The core issue is something else entirely: most companies try to jump straight to Invent, skipping the solid process redesign at the Reshape level. That is a direct path to "pilot purgatory."

How to Structure AI Implementation Step by Step

Step 1: Redesign Processes First, Then Choose Technology

McKinsey shows that companies achieving the highest returns on investment are nearly three times more likely to thoroughly redesign their operational processes before choosing AI tools and models [11]. Only 21% of organizations have actually attempted to redesign any workflow [11].

In practice, this means: before asking "which AI model should we use," ask "which task do we want to change and what should it look like after the change."

Step 2: Ensure Data Quality Before Scaling

AI models require structured, consistent, and complete data. When an organization lacks a mature data ecosystem, Data Science (DS) specialists spend 60% to 80% of their working time on manual data cleaning and preparation [3]. That is wasting potential before the project even starts.

The minimum entry point: an integrated data repository (CRM, ERP, or knowledge base), a quality validation process, and a clearly defined single source of truth for key metrics.

Step 3: Rebuild the Budget According to the 10/20/70 Rule

BCG has established that effective AI transformations require a complete reversal of typical IT budgeting models [17]:


Area	Typical allocation (wrong)	AI leaders' allocation (BCG)
Algorithms and models	~40%	10%
Data technology and infrastructure	~53%	20%
People, processes, change management	~7%	70%

Deloitte reports that the average organization spends 93% of its AI budget on technology and only 7% on people and process change [3]. Leaders do exactly the opposite.

Step 4: Introduce Change Management Before Resistance Appears

31% of employees admit to actively resisting new AI tools: refusing to use them, deliberately entering bad data [3]. Companies that ignore this buy expensive tools used by a minority of the team.

A proven approach is the ADKAR model (Prosci): sequentially building Awareness of the need for change, Desire to participate, Knowledge and Ability to act, and finally Reinforcement of new work standards [3]. Research indicates that achieving 30% active adoption among employees makes the change self-sustaining [3].

Step 5: Build the Human into the Decision Loop

AI handles tactics; the human protects strategy and brand. This is not a slogan — it is process architecture. Every AI output, before it reaches a customer or influences an irreversible decision, should pass through a defined quality control checkpoint: a stop condition built into the pipeline, not added at the end as an option.

Three Implementations That Show How Process Beats Tool

Klarna: Lessons from Trial and Model Correction

In early 2024, Klarna launched an OpenAI-powered customer service assistant that in its first month handled 2.3 million conversations, reducing resolution time from 11 minutes to under 2 [16]. Overall, the company reported savings of several tens of percentage points in operational costs in customer service and marketing, where visual content production was shortened from 6 weeks to 7 days [12].

This is where the difference begins: an implementation optimized purely for cost-cutting and headcount reduction hit a quality barrier with complex interactions [16]. Klarna had to step back from the aggressive model and shift to a hybrid co-pilot model: AI handles high-volume routine queries (return statuses, payment schedules), while complex, emotional, or high-risk cases are immediately escalated to a human agent with full conversation context in hand [7].

Octopus Energy: Gradual Authorization

Rather than a one-time revolution, Octopus Energy introduced AI in stages: first AI generated email drafts edited by agents, then summaries, and finally authorized actions [7]. The result: customer satisfaction at 80% (versus 65% with human-only service), handling the volume of 250 full-time employees, zero layoffs, a 30-fold increase in customer count to 7.8 million [7]. The process ran deeper than the tool.

Żabka: AI Built into the Work Environment

Żabka Polska deployed the Żabka Assistant application across more than 10,000 stores [66]. The system processes 450 TB of data daily in the cloud and generates real-time prioritized tasks for cashiers on a tablet: when to prepare hot dogs, when to apply markdowns on near-expiry products, when to report a refrigerator fault [25]. AI is not a chatbot. It is embedded in the physical work environment and reduces employees' cognitive load instead of adding a new application to their tool stack.

Common Mistakes and How to Avoid Them

Mistake 1: Automating Chaos

Deploying AI into a poorly documented, inconsistent process generates faster chaos. The rule: standardize and document first, then automate. If a process depends on the informal knowledge of specific individuals, AI will not improve it [19].

Mistake 2: Parallel AI Tool Sprawl

78% of employees admit to using unauthorized AI tools at work [4]. Each operates in a separate context. The result: fragmentation of organizational knowledge, security risk, productivity losses. Companies that do not prevent this pay twice: for the tools and for the time lost switching between them [4].

Mistake 3: Skipping Change Management

A tool installed without working with the team will be ignored or sabotaged. Deloitte reports that 42% of organizations cite employee resistance as the main challenge in AI implementation [3]. The solution is not technical.

Mistake 4: Ignoring the "Decision Window"

Organizational AI maturity is measured by how quickly a company can act on a model's recommendation before the window to change the outcome closes [3]. Companies that deploy predictive models without redesigning the decision-making process have the data and no way to respond to it.

Traditional Approach vs. Process-Based Approach


Dimension	Buying a tool	Process-based implementation
Starting point	Technology selection	Workflow diagnosis
Budget	93% on technology [3]	70% on people and processes [17]
Change management	Tool usage training	ADKAR, AI Champions, staged adoption
Data quality	Discovered along the way	Entry requirement
Human in the loop	Optional, added after the fact	Built into pipeline architecture
Success metric	License activation	Measurable P&L impact
Typical outcome	Pilot without scaling [21]	Operational repeatability

What the Process-Based Approach Delivers

Companies that move away from the "buy a tool" logic and shift to the "design a process" logic achieve results qualitatively different from the market average. According to BCG, only 5% of organizations reach the level of full AI embedding in operations [17]. This elite group generates 3.6 to 4 times higher three-year shareholder returns than competitors [17].

This is not the effect of better models. It is the effect of better processes.

Polish market examples extend beyond Żabka. PZU has been deploying computer vision for vehicle damage assessment since 2018, integrating Semantic OCR for document processing, and using satellite analysis for agricultural damage estimation, reducing analysis time by 50% and accelerating payouts by 25% [38]. VeloBank deployed a low-code BPM platform in five months, automating 62 manual tasks and launching a lending service activated by photographing a product price tag, which has so far financed over $1,357,500 in loans [34].

The common denominator: process redesign preceded technology selection.

FAQ: AI Implementation in Business

Where to Start with AI Implementation in a Business?

Start with a process audit, not a tool selection. Identify three to five workflows that consume the most repetitive team time and have well-documented operating rules. These are the best processes to start with. Data quality in those areas is a prerequisite before launching any model.

How Much Does AI Implementation in a Business Cost?

The cost depends on the scale and depth of transformation, but the key mistake is concentrating the budget on licenses and models. According to BCG [17], the right ratio is 10% on algorithms, 20% on technical infrastructure, and 70% on people, process change, and change management. In practice: for every dollar spent on a model, three dollars are needed for operational implementation and team adoption.

How Long Does AI Implementation in a Business Take?

The first measurable results appear within 3 to 6 months in Deploy mode (ready-made SaaS tools on existing processes). Deeper redesign of operational functions (Reshape mode) requires 6 to 18 months, and full integration of AI as an element of the business model is a 2 to 3 year horizon. Companies that compress this timeline without staging join the 42% of organizations that abandon AI projects at the pilot stage [21].

How to Measure the Success of AI Implementation?

Measure what changes in P&L or in operational cost: task resolution time before and after implementation, volume handled without headcount growth, the percentage of active tool users on the team (the 30% threshold is the point after which adoption becomes self-sustaining [3]), and the measurable reduction in working time spent on repetitive tasks. License activation is not a success metric.

If you are wondering which process to start with in your company and how to distinguish AI projects with real potential from those that will end up as yet another pilot without follow-through, we would be glad to work through the analysis together. Write to us.