Implementing AI That Actually Works: Small Experiments Instead of a Big Rollout
Why 95% of AI pilots in companies yield no return, and what I do differently to make AI genuinely contribute to the bottom line.
Who it's for: For founders of small and medium-sized organizations and senior managers facing the decision of how to approach AI in their company.
Why you should read it: You'll find hard data from reports by MIT, McKinsey, and BCG, a reference to today's most talked-about book on working with AI, and my own specific example of what worked and why small steps triumphed over a grand plan.
Reading time: approx. 8 minutes.
I'll start with a number that should change how you think about implementing AI
95% of generative AI pilots in companies yield no measurable financial return. This isn't my opinion. It's the conclusion of the MIT report "The GenAI Divide: State of AI in Business 2025," based on an analysis of 300 implementations, over 150 interviews with leaders, and a survey of 350 employees [1]. The authors call this phenomenon the "GenAI Divide," the gap between how widely companies use AI and how rarely they actually profit from it [1][2].
Interestingly, MIT flatly rejects the simplest explanation. The report's authors state that this divide doesn't stem from the quality of the models or regulations, but from the implementation approach [1]. The problem isn't AI. The problem is how companies try to implement it.
A similar picture is painted by the latest McKinsey report: 88% of organizations worldwide now use AI in at least one area of their business, but only about 5.5% of companies, defined as "high performers," actually feel AI's impact on their financial results, exceeding 5% of EBIT [3][4]. The rest are stuck in what McKinsey calls "pilot purgatory," testing and experimenting but failing to scale [4].
My journey with automation before ChatGPT appeared
I started back when the term "AI" wasn't yet used in this context. My primary tools were Excel and Google Sheets, and one of my favorite functions was importrange. It allowed tables owned by different people, with different access levels, to automatically populate with data entered once, in a single location.
This shaped a principle I adhere to today: data input in an organization should happen only once. Then, that data, in various formats, should automatically appear wherever it's needed. This is how analytics and finance, in particular, should work. I followed the same principle in sales, where results entered into a CRM automatically generated reports that visualized key KPIs in real time.
Today, these capabilities are exponentially greater. Data can be pulled directly from email communication or video conferences, and some models are evolving to anticipate our needs before we even express them. On top of that, there's a whole suite of tools I use daily: ChatGPT and Claude for analytical and text-based work, Lovable for rapid website prototyping, Envato AI and Opus Clip for video content production, and for graphics, I experiment with MidJourney and Nano Banana.
The effect in numbers: automating reporting gave us a significant increase in the quality and speed of decision-making. It suddenly became clear in real time how profitable operations were in a given area, or the current effectiveness of new client acquisition, instead of finding out with a delay during a periodic summary. When creating graphics in a specific visual style, the difference is even more tangible: what used to take several hours of work for a single graphic now takes just a few minutes.
Two different paths that should be separated
Here we arrive at a distinction I consider crucial and one that I believe appears too rarely in guides on AI implementation.
The first path is inexpensive, quick experiments with off-the-shelf tools. Where organizational data is not sensitive, it's worth simply testing what's available on the market without extensive upfront planning. This is exactly what Ethan Mollick, a Wharton professor and author of the bestselling book Co-Intelligence, describes, formulating one of the four fundamental principles of working with AI: always invite AI to the table and treat this work as a disciplined experiment [5]. Mollick calls this the "jagged frontier" phenomenon, where a model can surprisingly handle a difficult task yet fail at something seemingly simple [5][6]. The only way to discover this frontier in the context of your own company is to test, not theorize. Research cited by Mollick also shows that entrepreneurs who use disciplined, hypothesis-driven testing achieve significantly higher revenues than those who do not [6].
The second path is costly, dedicated implementations. Very often, the challenge is data confidentiality, which requires organizations to have specific, dedicated solutions running on their own cloud infrastructure. This is expensive and time-consuming, so the most important starting point is to precisely map the process and information flows, and to check two things: how much time this work actually takes people (or, as programmers jokingly say, "protein-based systems") today, and what the real ROI of such automation will be. It's equally important to determine whether the process is a permanent part of the organization or if it will undergo dynamic changes in the coming years, or even months.
Why human flexibility still wins where exceptions matter
People and their work are more flexible and better adapted to changing business needs than a rigidly programmed AI-based process. AI tools often struggle with exceptional cases and strategic or systemic changes.
And the data directly confirms this. The same MIT report indicates that current generative AI systems have problems with persistent contextual memory, don't learn from feedback on the fly, and repeat the same mistakes in subsequent sessions [2]. One manager quoted in the report described it bluntly: the tool doesn't remember customer preferences or learn from previous corrections, but repeats the same errors in every new session [2]. This is the exact same observation that underlies my approach: before you lock a process into automation, make sure that process is actually stable enough to be worth locking down.
A common mistake highlighted by data
The MIT report points to something that sounds counterintuitive: companies building their own internal AI solutions from scratch are about three times less likely to succeed than those who buy off-the-shelf tools from specialized vendors and build partnerships with them [2]. Over half of AI budgets today go to sales and marketing, whereas MIT indicates that the largest, most measurable return usually comes from automating internal processes, where AI replaces tedious, repetitive operational work [2].
BCG also confirms this with its "10-20-70" rule: only 10% of the effort in implementing AI should go into the algorithms themselves, 20% into technology and data, and a full 70%—the vast majority—into people and processes [7][8]. Companies that reverse these proportions and invest primarily in technology regularly fail to achieve a return, regardless of how good the AI model itself is [8][9].
What to do with this in practice
A few principles that I apply myself and recommend to the founders and managers I work with:
- Separate experiments from implementations. Where data isn't sensitive, test quickly and cheaply with off-the-shelf tools. Where data confidentiality is a concern, start by thoroughly mapping the process.
- Before investing in costly automation, calculate the real ROI. Check how much time a given process actually takes people today before you decide to lock it into a system.
- Check the process's stability. If a given area is set to change dynamically in the coming months or years, human flexibility might be worth more than rigid automation.
- Don't build everything from scratch yourself. The data shows that using ready-made tools and partnerships is more effective than building internal solutions in isolation.
- Invest primarily in people and processes, not just technology. This is 70% of the total effort, not an afterthought.
Contact me
What I've described above isn't theory. It's an approach I use daily in running my own business and working with founders and managers on their AI implementation strategy, alongside topics like team management and leadership.
If you want to talk about what my approach to AI, content, and client work looks like in practice, feel free to get in touch.
Sources
[1] MIT NANDA, The GenAI Divide: State of AI in Business 2025 — https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
[2] Legal.io, MIT Report Finds 95% of AI Pilots Fail to Deliver ROI, Exposing "GenAI Divide" — https://www.legal.io/blog/5719519/MIT-Report-Finds-95-of-AI-Pilots-Fail-to-Deliver-ROI-Exposing-GenAI-Divide
[3] McKinsey & Company, The state of AI in 2025: Agents, innovation, and transformation — https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
[4] Winsome Marketing, McKinsey's State of AI Report: 88% Adoption, But Only 6% Are Actually Winning — https://winsomemarketing.com/ai-in-marketing/mckinseys-state-of-ai-report-88-adoption-but-only-6-are-actually-winning
[5] Ethan Mollick, Co-Intelligence: Living and Working with AI, Portfolio/Penguin, 2024 — https://www.penguinrandomhouse.com/books/741805/co-intelligence-by-ethan-mollick/
[6] Forbes, Ethan Mollick On The Four Rules Of Co-Intelligence With AI — https://www.forbes.com/sites/peterhigh/2024/05/07/ethan-mollick-on-the-four-rules-of-co-intelligence-with-ai/
[7] BCG, The Leader's Guide to Transforming with AI — https://www.bcg.com/featured-insights/the-leaders-guide-to-transforming-with-ai
[8] BCG, How AI Leaders Create Competitive Advantage — https://www.bcg.com/publications/2026/how-ai-leaders-create-competitive-advantage
[9] BCG, AI Transformation Is a Workforce Transformation — https://www.bcg.com/publications/2026/ai-transformation-is-a-workforce-transformation