1. Receiving goods without retyping
In FoxFix, supplier items used to be retyped into stock. Now AI automation prepares a draft stock booking and a person confirms it with one click. The retyping is gone. Lesson: don't leave everything to a single language model. A mix of specialised tools, firm rules and a human check is more reliable than any one of those steps on its own.
2. An AI assistant for staff and customers, with clear rules
It is the most requested feature and also the easiest one to get wrong. In FoxFix the assistant helps the team find answers and prepare material faster. In the ZEUS Core app it answers customers' everyday questions, so they don't need to call a sales rep. Lesson: rules first, chat second. It must be clear in advance what the assistant may do, where it runs and who is responsible for the result.
3. An energy offer in minutes instead of hours
A tailored energy offer means working out distribution tariffs, rates, regulated charges and VAT. In ZEUS Core, AI helps the sales rep pre-fill the calculation and the system works out the rest. The offer takes minutes and the rep has more time to sell. Lesson: AI occasionally fills in things it shouldn't. So every result passes validation rules (ranges, totals, mandatory tariffs), and anything outside them is checked by a person.
4. Content that grows every day
The Slevy a Akce app needs fresh deals and coupons every day. We built an automation in which AI prepares new offers for approval every night. For partner stores they go live by themselves. The client's team only approves content, which takes a few minutes a day instead of hours. Lesson: start with an approval queue and leave autopilot for later. Trust grows with data, so we switched on automatic publishing only where we had a verified relationship with the store.
5. AI that needs no internet
For Bodka, our own meeting transcription and summary product, we run speech recognition and the language model directly on the phone and on the Mac. Audio never leaves the device. That took our own speech recognition model fine-tuned for Czech and language models adapted for phones. The result is a product with no per-user cloud costs and no question of where the data goes. Lesson: AI that runs on the device itself is now a fully viable option for sensitive data.
One rule for all five
Every deployment that paid off had a clear purpose, a human in the loop where a mistake hurts, a measurable result and a cost ceiling. The ones we turned down usually lacked at least one of those. If you are wondering where to start with AI in your company, start with the most boring routine task someone does every day. That is where the investment pays back fastest.