77% of companies in Poland haven’t implemented AI, and 80% of those that started are stuck at the pilot stage. The reason isn’t technology. It’s that most companies confuse buying a license with implementing AI, skip the part where someone owns the project, and try to boil the ocean instead of automating one repetitive process that runs ten times a week.
In this article you’ll learn:
- Why buying a ChatGPT or Copilot license is not the same as implementing AI
- What “shadow AI” looks like in your company right now and what risks it creates
- Why 95% of AI pilots never reach production and what role makes the difference
- How to choose your first automation using three simple criteria
- What realistic budget, timeline and ROI look like for a first implementation
- What to do tonight to start, without an audit or a strategy deck
What “implementing AI” actually means (and what it doesn’t)
Start with an uncomfortable statistic. According to PwC’s 2026 AI Performance Study, 74% of AI’s economic value is captured by just 20% of organisations — leaving the vast majority of businesses stuck in pilot mode, generating activity without measurable financial return.
The pattern I see most often looks like this. A company buys Microsoft Copilot or ChatGPT Enterprise seats, sends a memo saying “use AI”, and considers the topic closed. Six months later there’s no measurable change in any process, no one can name a single workflow that runs differently, and the only certainty is the invoice.
„If it’s just a license thrown at employees with the slogan “use AI”, then you don’t have AI implemented in the company. You just have an invoice for AI.”
That’s the starting point worth being honest about. Everything that follows assumes you want the second outcome, not the first.
Shadow AI: what your team is already doing without you
Even if you’ve officially done nothing with AI, 75% of your people are already using it in some form — and 78% of them brought their own tools to work without waiting for company policy.
Sales writes offers in ChatGPT. Marketing generates LinkedIn posts. Customer service drafts complaint replies. Someone’s assistant runs meeting transcriptions through a free tool nobody approved. Each person has their own prompts, their own habits, their own preferred model. From the outside, the company looks the same. From the inside, it’s ten chefs in one restaurant, each cooking the same dish from a different recipe.
This creates three concrete risks.
Data leakage. Offers, financial analyses, client lists end up on personal accounts and free tiers you have no visibility into. Where exactly that data is stored and what it’s used for is, in practice, unknown.
External inconsistency. Two salespeople reply to the same type of inquiry with different arguments, different tone, sometimes different pricing logic. The client experience becomes a lottery.
Knowledge loss. One person on the team has cracked a really good prompt for generating proposals. Nobody else knows about it. When that person leaves, the prompt leaves with them.
„If everyone uses AI differently, you don’t have an AI strategy. You have random experiments nobody controls.”
Why 95% of AI pilots fail: the missing AI owner
MIT NANDA research shows that 95% of AI pilots never reach production. Only 5% make it through.
In companies seeing real results, nearly half of respondents confirm that senior leaders show clear personal ownership of AI — compared with only 16% at companies that don’t.
Think of it like building a house. You can have a great architect, top-tier contractors, the best suppliers. If nobody runs the site, each team does its own thing and you end up with five disconnected fragments instead of a house. AI implementation works the same way. Someone has to walk the site, make calls in real time, coordinate between business and technology.
This person doesn’t need to be technical. AI implementation is roughly 80% non-technical work, processes, procedures, organisation, and 20% technical. What they need is two things: real decision-making authority, and someone competent beside them to execute the technical side, whether in-house or external.
In a 50 to 100 person company this is usually the CEO or a deputy. In a larger company it might be the COO. What it almost never is, in my experience, is a freshly invented Chief AI Officer title without budget, without authority, and without a seat at the leadership table. That role is a wish, not an initiative.
Process before tools: organise first, then automate
You can’t automate a process that doesn’t exist as a process. If the current way of handling client inquiries lives in five people’s heads and varies by who picks up the email, automating it will simply make the inconsistency faster.
This is where most companies want to skip a step. They want to jump straight to “what tool should we buy”. The honest answer is: it doesn’t matter yet. First, write down what currently happens. Who does what, in what order, with what inputs and outputs. Even a rough version is enough. Without it, you’re automating chaos and you won’t even know whether the result is better or worse than the baseline, because you never measured the baseline.
This part isn’t exciting. It’s also the part that separates the 5% from the 95%.
How to choose your first automation: three criteria
Counter-intuitively, your first AI automation should be small. Not strategic, not transformative. Small. The goal is to prove to your team and to yourself that this works, not to remake the company in six months.
Pick one process that meets all three of these criteria:
- It happens at least 10 times a week. Below that frequency, the return on investment gets very hard to justify.
- It’s measurable in time today. You need to know how many people spend how many minutes on it, otherwise you can’t prove anything after.
- It’s reversible. A mistake by the automation shouldn’t lose you a client. Drafting an email for human review is fine. Auto-sending offers to clients is not, at least not as your first project.
That’s it. Three criteria, applied honestly, will eliminate most of the “wouldn’t it be cool if” ideas and leave you with the candidates that actually pay back.
Three first automations that pay back
Three patterns work reliably as a first project for mid-sized B2B companies. None of them are exotic.
Inbox triage and CRM enrichment. Imagine five salespeople, ten inquiries per day each, 15 to 20 minutes per inquiry spent on classification, entering data into the CRM, and looking up basic information about the company that wrote in. That’s a conservative estimate. After automation, classification takes 30 seconds to two minutes, the CRM entry happens automatically, and enrichment data arrives ready to use. On a monthly scale, this frees up close to one third of a full-time role, roughly 3,500 to 4,000 PLN (about $900) in direct labour costs. The bigger win is that salespeople use that time to serve more clients or work existing ones more deeply.
Meeting transcription and action points. For any company that runs a lot of online meetings, automating transcription, structured notes and action-point extraction is a clean, low-risk first project. The quantifiable savings depend on meeting volume, but more importantly it builds the foundation for downstream automations: pulling action points into a task system, summarising client calls into CRM notes, generating follow-up drafts.
Offer generation from a brief. Salespeople typically spend 30 minutes to two hours preparing a proposal. A well-designed automation can generate the first draft in the company’s style, tailored to the client’s brief. The human stays in the loop for the final check and adjustments. Time savings vary by complexity, but the bigger value is consistency, the offer always reflects the current product, pricing and positioning, not what each salesperson remembers.
Budget, timeline and realistic ROI
For the kind of first automation described above, working with an external partner, you should expect a one-off implementation cost in the range of $5,000 to $8,000, plus around $150 per month in tool subscriptions. These are indicative figures — the exact scope depends on process complexity and existing infrastructure. Most subscription tools carry over to subsequent automations without costs scaling linearly.
The timeline from kickoff to a working, tested automation is measured in weeks, not months. The output is tangible: roughly a third of a full-time employee’s capacity freed up each month, and savings in the region of $900. And because most subscription tools carry over to your next automation, the cost curve flattens as you scale — payback comes faster than most organisations expect.
You can also build internally if you have someone technically competent on the team. Either way, the C-level owner requirement doesn’t change. Without that role, the cost numbers above stop meaning anything because the project doesn’t finish.
What to do tonight
You don’t need an audit, a strategy deck or a six-month roadmap to start. You need two lists.
First, count the repetitive, manual processes in your company that run on email, Excel, and moving data from one tool to another. Don’t analyse them, just list them.
Second, go to your three best employees and ask one question: what do you waste the most time on, repeatedly, every day? The answers will surprise you, and they’ll usually point at exactly the processes worth automating first.
That’s the starting list. Everything else, who owns AI, which process to pick, what tool to use, falls into place once you know where you’re starting from.
Your competitor is already counting these processes. The only real question is whether they finish counting before you, or after.
Key takeaways
- Buying AI licenses is procurement, not implementation. Implementation means a redesigned process with a measurable, repeatable output and a person accountable for it.
- Shadow AI is already happening in roughly 75% of companies and creates three real risks: data leakage, external inconsistency, and knowledge loss when people leave.
- 95% of AI pilots fail to reach production. The biggest single predictor of success is a C-level owner with real decision-making authority, McKinsey shows a three-fold gap in C-level involvement between successful and unsuccessful companies.
- AI implementation is roughly 80% process and organisation, 20% technology. Skipping the process step means automating chaos.
- Your first automation should be small, frequent (10+ times per week), measurable in time, and reversible. Inbox triage with CRM enrichment is the most reliable starting point.
- Realistic numbers for a first project: 6 weeks, $5,000 to $8,000 implementation cost, ~$150/month in tools, payback in 6 to 9 months.
- You don’t need an audit to start. You need a list of repetitive processes and a conversation with three of your best employees.
Wondering whether AI implementation makes sense for your company, and which process you should automate first? Book a free expert consultation – we’ll walk through your specific case and identify two or three concrete starting points, no commitment required.
Want to see how this plays out in practice? Watch the full episode on the Nejman AI channel.
