Task and reason
Task: build the company’s own Telegram mini app and measure how long the road from task to working version takes when AI writes the code and people set tasks and accept the result.
Reason: the industry has plenty of figures about speed-ups, but a client needs to understand something else: what they will have in an hour, in a week and by launch. Figures from other people’s examples do not show this, so we measured the road on our own product: on the night of 17 September 2026 the company needed a mini app that shows its lines of work, takes requests and calculates prices.
AI-assisted development has moved from experiments to everyday practice: according to the 2025 Stack Overflow survey, 84% of developers use AI tools or plan to start. Back in 2022 researchers from GitHub and Microsoft measured the effect on the task “write a web server”: the group with an AI assistant finished 55.8% faster.
At OM-Digital we build websites and apps together with AI and measure how long it takes. Below are the six steps our Telegram mini app went through, and what AI did along the way and what people did.
Short answer
Today a first working version is assembled in minutes: a page, a prototype, an app with the main screens and forms. A finished product still means testing on real phones, fixing bugs, launch and support, and that road is still measured in weeks. AI has shortened the first stretch and sped up the rest, but it has not removed them.
Step 1. Setting the task as a list of screens and features
On 17 September 2026 at 00:14 a task appeared in the log: build the company’s Telegram mini app. The brief listed ten lines of work, a request form, a service picker based on a few questions, a calculator and a demo assistant. The more specific the list, the fewer fixes later.
Step 2. A working draft and where it lives
At 00:30 the first version was already opening on the website. Sixteen minutes. The word “minutes” should be read precisely here: what appears is a working draft in which screens open, buttons respond and the request goes through. It is a basis for showing the idea, discussing it and making a decision. It becomes a finished product later.
Step 3. Testing the draft inside the messenger on a phone
The messenger’s built-in browser renders the design in its own way, so we open every version on a phone inside Telegram itself, screen by screen, and not only in a desktop browser.
Step 4. Adding the features the client sees
Seven minutes later, at 00:37, version 1.1 came out with one-tap contact sharing and data that syncs between devices. A person does not have to type their number by hand, and their data stays with them across devices.
Step 5. Collecting fixes in a batch based on feedback
By four in the morning nine more fixes had come out based on feedback: card sizes, photos, the order of sections, which blocks to keep on the main screen. Each fix took minutes, but the decision about what exactly to change was made by a person looking at the screen.
Step 6. Testing on real data and getting to launch
We finish every version the same way. AI makes the changes and runs the automated checks itself. Then a person opens the result on a phone and goes through the whole scenario: taps every new button live and tests forms on real input, not short samples.
This produced a short checklist for anyone starting to build apps with AI:
- keep each rule or limit in one place in the code and ask the AI to check that it is not duplicated anywhere;
- test on real data: long texts, large files, slow connections;
- before release, install the app on a phone and go through the scenario by hand.
The road to launch is not just code. AI does not test on the device by itself, does not prepare texts and materials for publication and does not make design decisions.
What this changes for business
First, an idea can be tested before large investments. A prototype is assembled in an evening, and it can be shown to clients, partners or the team before deciding what to invest in.
Second, the structure of work changes. Less and less time goes into writing code and more and more into setting the task, checking and finishing.
Third, the human role shifts towards decisions and acceptance. AI writes fast, but people decide what counts as done.
Key takeaways
- Problem: figures about faster development do not answer the client’s question of what they will have in an hour and by launch.
- A six-step solution: a list of screens, a draft, testing in the messenger, client-facing features, a batch of fixes, tests on real data.
- Measurable result: 16 minutes to the first version, 7 more minutes to the version with contact sharing, nine fixes in 2 hours 14 minutes.
- A first working version is assembled in minutes; the road to a finished product includes testing on the device, materials for publication and launch.
- A prototype in an evening lets you test an idea before investing in development.
Result
- The task appeared at 00:14, and at 00:30 the first version was already opening on the website: ten lines of work, a request form, a question-based service picker, a calculator and a demo assistant. That is 16 minutes.
- Seven minutes later, at 00:37, a version came out with one-tap contact sharing and data that syncs between devices.
- Between 01:48 and 04:02, in 2 hours 14 minutes, nine fixes came out based on feedback: card sizes, photos, the order of sections.