I have calls with founders who want AI. Most of the conversation looks the same: big expectations, vague brief, systems that don't talk to each other. And I have to say something nobody wants to hear — not yet.
Not because AI doesn't work. But because AI is a multiplier. It multiplies what you already have. If you have chaos, AI will speed up your chaos and make it more expensive. If you have order, AI will multiply that.
This article is a checklist. Go through it before you decide on your first vendor or first chatbot. You'll save yourself months of disappointment and hundreds of thousands of CZK.
Where does your company stand right now?
Based on dozens of advisory engagements, companies are usually in one of three states.
Phase 1 — Foundations are missing. Processes live in people's heads and Excel. CRM doesn't exist or nobody uses it. Data is scattered. AI doesn't make sense yet — the basics need to be digitized first.
Phase 2 — Systems don't work together. CRM, ERP, email, and accounting each live their own life. Data exists, but nobody knows which is correct. AI here would solve the wrong problem.
Phase 3 — Systems work, data is clean. Now is the time for an AI layer. Processes are stable, data is available, the company is ready for automation and augmentation.
If you're in Phase 3, you can go straight to AI Operating System Design. If you're in Phase 1 or 2 — read on.
Five prerequisites for a successful AI deployment
1. Data exists, is structured, and is understandable
This is the most common blocker. Companies that come in saying "we have data" usually mean emails, spreadsheets, and folders on a server. That isn't a data foundation for AI — it's an archive.
AI needs data that is:
Complete. Records missing key fields confuse AI more than they help. If you have 3,000 contacts in CRM and 60% are missing the industry or company size, AI can't work with that meaningfully. Empty fields aren't neutral — they're noise.
Consistent. The same thing has to be named the same way everywhere. Customer / client / buyer in one system is a problem. Praha / Prague / PRG too. Before deploying AI, the key datasets need to be reviewed and the terminology unified.
In a format understandable to humans and machines. No merged cells in Excel, no abbreviations without explanations, no values like "see attachment" in a field that should hold a number. If you can't explain to a new employee in an hour what every field in CRM means, AI won't get it either.
Up to date. Data three years old without updates is more noise than signal for AI. A knowledge base for AI needs an ongoing maintenance process, not a one-off import.
The three categories of data you need to get right first:
- Customer data — who the customer is, what they buy from you, their lifetime value, when they last contacted support.
- Operational data — how processes run, who does what, where the bottlenecks are. If it lives only in emails and meetings, it isn't data.
- Product or service data — what you sell, how it differs, what variants exist, pricing, terms. The AI knowledge base must have a source of truth to draw from.
2. Baseline systems are in place and actually used
AI plugs into existing systems. If the systems aren't there or aren't used, there's nothing to plug into. The key phrase is actually used — a system nobody works in is worse than no system at all.
The baseline categories differ by segment:
B2B services (advisory, agencies, IT companies): a CRM for managing opportunities (Pipedrive, HubSpot), project management, invoicing connected to accounting.
E-commerce: an e-shop platform (Shoptet, Shopify), an ERP for orders and stock, a warehouse system with real stock levels, payments automation, marketing automation. Without these layers connected, AI works with incomplete information about the customer.
Manufacturing: an ERP as the foundation (POHODA, Money S4, SAP), production planning, inventory management for both materials and finished goods. AI in industry typically works with data from machines and sensors — that requires an additional integration layer.
SaaS and software companies: a CRM for the sales pipeline, a helpdesk for customer support (Intercom, Zendesk), product analytics, a billing system.
Retail networks: a POS system with central oversight, central inventory, a loyalty program as a data source for customer behavior.
Start with the systems that generate or manage the most data about customers and processes. Those go first.
3. Processes are documented and stable
AI learns from how things work. If a process exists only in one person's head, or changes every month, AI won't manage it.
A documented process doesn't mean a 50-page manual. It's enough to answer five questions for every key process: Who does it? When and why? What are the inputs? What are the outputs? What happens if it fails?
I recommend validating processes with a firm specialized in process optimization before you start thinking about AI. A process consultant sees inefficiencies the internal team has stopped noticing because they got used to them. AI deployed onto a bad process won't optimize it — it'll automate the mistake. At the end of the article I can connect you with such a partner.
4. There's a person inside the company who understands AI and drives it
AI projects pushed by one person in their spare time die. You need an internal AI champion with a mandate from leadership, real time, and the ability to inspire others — not just convince them.
AI deployment changes other people's work. People naturally resist change, especially if they see it as a threat to their position. Companies that communicate AI as a tool that makes work easier for people have dramatically higher adoption rates than those that simply announce the change. Bring people from different departments into the pilot from day one.
What exactly an internal AI champion needs to know and be able to do — that's a topic for its own article: What your internal AI champion needs to know →
5. Leadership actively backs it and has realistic expectations
Without sponsorship at the leadership level, AI projects stall at the first problem. And problems will come — always.
Realistic expectations are critical. The first 6 months are an investment — into setup, integration, training, iteration. Leadership that expects immediate ROI will end the project early, right when it's just getting started.
A good test: ask leadership what success will look like in a year. If the answer is concrete (save 20 hours a week on support, shorten customer onboarding from a week to two days), the project has a chance. If the answer is "we'll be an AI company" — clarify the goals first.
How to start: one pilot, not the whole company
AI can't be deployed instantly across everything. Every deployment should start with one specific use case — a pilot. It validates the approach, shows real results, and teaches the team to work with AI before you commit a big budget.
How to pick the right pilot, and why this phase matters even for new processes and apps, I cover in a separate article: Why every AI project should start with a pilot →
What to take away
A closing checklist:
✓ Data is complete, consistent, and in a structured format
✓ Baseline systems matching your segment are in place and being used
✓ Key processes are documented
✓ You have an internal AI champion with a mandate and time
✓ Leadership actively backs the project and has realistic expectations
✓ You have one specific pilot picked
If every box is checked, you're ready. If not — you know what to do next.
If you're not sure where your company stands, or you want an independent view of what deploying AI actually involves, I'd be glad to help you map it out. And if your processes need work before AI — I can connect you with a partner specialized in process optimization.
/ Terms in this article /
