How a Claude-based Telegram assistant eased the CEO's load and took on part of the team's work

A small SaaS team embedded an AI assistant directly into their work Telegram chat and over several months transformed it from an experiment into a full-fledged operational tool. The bot answers questions from product managers, reviews code, creates tasks in GitHub, and takes on part of the workload that previously fell almost entirely on the CEO.

How the assistant works

The team describes a fairly typical workflow for a small SaaS business: five developers, four managers, one CEO, a stack built on PHP, Vue, MySQL, GitHub, and Telegram. The problem wasn't a lack of tools, but that almost all internal expertise converged in one person. The CEO manually answered repetitive questions, explained how the product worked, and constantly switched between sales, support, and development. Because of this, even simple customer clarifications would hang for hours waiting for a response.

They embedded the assistant directly in the general chat so it works where daily communication already happens. The foundation is a bot built on Claude, which receives not just an isolated message, but the context of the thread, access to the codebase, customer data, GitHub issues, and internal documents. The authors emphasize an important point: the decisive factor turned out to be not fine-tuning, but properly assembled context. A model is useful only when it sees real data and the current work situation.

Where it really helps

The most straightforward scenario is quick questions from managers when a customer needs an answer right now. Instead of the chain of "ask the CEO, wait for the developer, check the code, come back with an answer," the bot itself finds the needed part of the project and gives a practical tip. The article gives an example with a template where a customer tried to insert a link: the assistant checked text processing and explained that HTML links are stripped, but a plain URL is preserved without problems.

"90% of the magic isn't in the model, but in what context you pass to it."

No less useful turned out to be the task creation mode. When the CEO asks to "write up a task," the assistant doesn't copy the conversation into an issue, but transforms the discussion into a technical description with implementation steps. In the case of apartment photo analysis, the bot itself linked the existing hashing with the new comparison scenario, added Hamming distance, matching thresholds, combined score, and validation on a large sample. The developer gets a tech-lead level task in a minute, not raw chat transcription.

• Answers routine manager questions in 15–40 seconds
• Creates GitHub issues with assignees and labels
• Suggests code solutions and narrows the scope of bug hunting
• Writes changelogs after releases directly in the chat
• Helps prepare customer responses even late in the evening

The third working scenario is messages after deployment. The bot automatically writes to the chat what improvements made it into the release: from unread message logic and interface changes to blocking emails from unwanted contacts. For a small team, this isn't cosmetics but operational discipline: less time spent on manual reporting, the chat retains a clear trail of changes, and managers quickly understand what's been deployed and what they can promise customers.

Where it fails

The authors honestly acknowledge that the assistant doesn't handle complex bugs at the intersection of multiple services as reliably as a human. It can find relevant code snippets, propose a convincing hypothesis, and still be wrong about how systems interact in actual runtime. The most unfortunate example is a situation where the bot confidently told a manager that Excel import functionality already existed, when in fact it wasn't in the product, and the team had to separately apologize to the customer.

There are also limitations that can't be solved by just improving the prompt. The assistant can gather information for a business decision, but shouldn't make it instead of the leader. It can suggest text for a response to an irritated customer, but won't replace the manager's confidence and support. Plus, long discussions in the work chat remain risky: if a topic stretches over several days, the model starts losing the thread because even a large context window isn't infinite.

To reduce the risk, the team introduced a simple but important rule: if there's no confidence, the bot should directly state that the answer needs to be verified with developers.

Over three months, the assistant left about 2,500 messages, created roughly 120 tasks, and helped close about 60% of manager questions without the CEO. But alongside this, there were three critical incidents due to hallucinations. This is a good result for a process accelerator, but a weak reason to consider such a system an autonomous source of truth.

What it means

This case well illustrates where corporate AI is moving in practice. The most useful format today is not a separate interface with a chat window, but an agent within an already existing work loop, where it has access to code, CRM, tasks, and discussion history. In this configuration, AI truly takes routine work off the leader's plate and accelerates the team. But trust in it is built not on beautiful answers, but on the quality of context, constraints, and the human's right to stop an error in time.