DEV Community: MatIntellect

Digital employee: how to help your team grow

MatIntellect — Sun, 01 Nov 2026 00:18:10 +0000

I believe you shouldn't fire employees just to replace them with artificial intelligence. The exception is when they've earned that decision themselves

AI doesn't work on its own: someone has to press the buttons. If you fire a specialist and replace them with AI, you'll be the one pressing those buttons

Do you have the expertise of a copywriter, designer, assistant, accountant or lawyer? Probably not: as a business owner, you're setting the direction of the company and dealing with higher-level decisions

AI can help you scale. Instead of replacing people, teach them to use it, help them get excited about it and adapt to it. Try different formats, experiment, bring in AI specialists, then introduce it into your business and grow

What is a digital employee?

A digital employee is an AI system with defined responsibilities and rules. Depending on how it's set up, it can sort incoming messages, collect enquiries, prepare texts or summaries and pass difficult questions to a person. It can carry out some actions on its own when it has access to the right tools and clear boundaries

For example, AI can gather information from enquiries and draft replies. But who decides which questions are difficult, or spots an answer that sounds convincing but is wrong? That takes someone who understands the work and is responsible for the result

Introducing AI with your team

Pick a specific part of the work and discuss it with the people who do it every day. A specialist can help identify repeated actions, where data is needed and what shouldn't be left to a system without review

Then define what the digital employee does independently and what it passes to a person. Test it with the team on real work examples: is the result useful, how long does checking take, and what happens when it gets something wrong? Give employees time to learn the tool and a way to suggest changes

The NIST AI risk management framework discusses defining human roles and oversight. The level of autonomy depends on the task, so it's worth knowing in advance who can step in and when

I explain which responsibilities a system can take on and where people keep the decision on my AI employee for business page

The numbers

Suppose a good employee costs $2,000 a month and an AI employee costs $200. If together they handle the workload of a five-person team, that comes to $2,200 rather than $10,000 a month

This is an illustrative calculation, not a guarantee: costs and results depend on the work, the setup and implementation expenses. But it shows why combining human expertise with AI can be worthwhile

Why I see it this way

I experienced this myself. I spent about four years in a management role in a marketing department. After AI arrived, the director decided it could replace most employees. I was fired, and six months later the company closed: the director couldn't handle that workload at the same standard

Making money doesn't mean optimizing absolutely everything. It means finding places where you can scale and grow. AI can help you do that. Firing your employees won't

AI Agent Implementation: Where to Start

MatIntellect — Sat, 31 Oct 2026 00:39:19 +0000

As obvious as it may sound, start implementing AI agents with the simplest tasks. That's where to begin

If you've only used ChatGPT and haven't worked with complex automations, any agent changes the way you work. Change is often difficult. Reshaping a business takes time, both for your AI system and for you personally

A common mistake business owners make is jumping into complex development and trying to automate everything at once. Don't do that. Neither you nor your system is ready for it. People don't want to change, and familiar paths are easier to follow

Write down where your day goes

Try a simple exercise: for one day, starting when you wake up, write down everything you do. Note the decisions you make, the actions you take, what follows what and, most importantly, where your time goes

You don't need to turn the day into a minute-by-minute report. Write each action down so you can see later where it begins and what you get at the end. For example: open an incoming email, identify the client's request and draft a response

Then give the list to an AI tool or the specialist who will automate your processes. It gives you something to examine for recurring work and decide where to start

Choose one simple task

For a first step, look for actions with clear input and a clear result. For example, the input is an email and the output is a short summary of the client's request. You can compare that with the original email much more easily than you can evaluate a task like "deal with all my email"

Decide where the result appears and who checks it. With an email summary, a person can open the original alongside the agent's text and see whether an important request was missed. If information is missing, the agent should show that rather than make it up

Give yourself and the system time

The first automation changes a familiar routine: you need to hand over a task, receive the result and check it. Your team also needs to know who handles that step now and where to look when something goes wrong

Test the agent on several real examples, including an email that doesn't contain enough information. Keep a person reviewing the output while it still raises questions. Sending a message to a client or taking another important action is a separate step, and the agent doesn't need to do it independently from day one

Once the first task works in a way you understand and you're used to working with it, you can take on the next. That's how a system grows around your business, one step at a time. I've explained more about turnkey AI agents on the service page

Start by looking at where your time goes. Then automate the simplest task first and move toward the more complex ones

AI maintenance cost: what you pay per month

MatIntellect — Fri, 30 Oct 2026 03:46:01 +0000

I think AI is underpriced right now, although plenty of people would argue with me. There are several ways to use it, and subscriptions are the most common. My recurring spending on three models and a server is around $300 a month

I pay $200 for Claude, $20 for ChatGPT, $20 for GLM and $30 for the server per month. I also track what using those models would cost if I paid directly for tokens through an API. My estimate is around $30,000 for the same volume of work. Tokens are pieces of text a model receives and generates. Subscriptions and token billing work differently, so that is my calculation, not a promise that everyone would save the same amount

Subscriptions, APIs and your own hardware

A subscription has a fixed price, but usage stays within the plan's limits. For example, Claude plans have usage limits, and ChatGPT Plus has limits too. Conversation length, model choice and tools can affect how much of the allowance you use

With an API, the bill depends on usage. According to OpenAI's billing guide, API charges are separate from a ChatGPT subscription and depend on the model, tokens and features used. Before choosing, estimate the workload: a subscription price and an API bill do not necessarily buy the same amount of work

Another option is to run a model on your own hardware. A local system using a powerful modern model needs computing resources, memory and storage. The exact requirements depend on the model; an ordinary laptop may not be enough for some larger ones

What goes into the monthly budget

My main expenses are subscriptions and a remote server. I buy additional tools when I need them. That spending lets me work on roughly 2-3 large projects a month while developing my own system

For a particular system, budget separately for the model, server and third-party services, as well as maintenance work: checking failures, applying updates and adapting it to new tasks. The mix depends on what the system does and which services it needs. I explain what AI system support and development covers on my site

The limits used to be much higher, but the grass used to be greener too

Multi-agent system for business: when to build a team

MatIntellect — Wed, 28 Oct 2026 07:45:56 +0000

If there are only a few tasks and scenarios, one general-purpose AI agent is enough. There's no need to create a complex architecture for that. When the work includes major areas, I give each agent a specific role

For me, the sign is a task that comes up regularly and needs its own knowledge base, memory, and tools. It's the same principle as hiring an employee: first, you need to know what work you'll give them

Where to start dividing the work

Take content production as an example. You regularly write posts and need images to go with them. Those are two areas: one handles the text, the other handles the visuals. You can assign a copywriter and a designer and define the result expected from each

The copywriter receives a topic and source material. Its job is to prepare the text. The designer receives the approved topic and the post's message, then creates an image based on them. Each agent has a clear assignment instead of just a job title on a list

What the copywriter needs

Telling an agent "write posts" isn't enough. It needs writing rules and examples of your style: how you address your reader, what expressions you use, and what should stay out of the text. Its writing skills and tools are set up for that work

When you edit a text, it's important for the agent to keep those lessons in memory. For example, if you remove a repeated idea, that can become a rule for the next post. This helps the copywriter gradually get closer to your style

How to connect the designer's work

The designer doesn't need the whole set of writing rules. It needs the message of the text and the image requirements: format, visual style, and what should appear in the frame. It has its own skills and tools for images and video

The handoff between agents matters too. If the designer starts with a draft and the text changes afterward, the image may no longer fit. That's why it's worth defining which version of the text goes to the designer and who checks the finished post and image together

Each agent is essentially a separate employee. You wouldn't expect your secretary to be an accountant, programmer, and designer all at once: they'd simply explode. The same applies to agents

An AI system architecture helps define these roles and connect them into a working process

AI implementation: start with lost revenue

MatIntellect — Tue, 27 Oct 2026 02:09:59 +0000

A business AI audit always starts with the main question: where is the business losing the most money, time and customers? A business exists to deliver an actual result for its owner. Every process we want to automate with AI needs to be examined from that perspective

The biggest misconception on the market is that artificial intelligence can solve any business problem. That's a lie. The most important thing AI can help a business do right now is scale

To scale it, you first need to examine the business thoroughly and find the weak spots where money, time and customers are being lost. Then automate slowly and carefully, from simple processes to complex ones. And most importantly, don't fire people. Someone still has to press the buttons

How a clinic lost customers after paying for ads

One clinic where I helped set up and automate business processes was losing customers simply because it had no automated analysis of incoming leads. Ads were running and money was being spent, but nobody was handling new inquiries quickly. The clinic was losing money as a result

The first thing I started with was an audit of incoming leads, meaning inquiries from potential customers. This directly affected the clinic's money. How quickly those inquiries were handled affected whether people brought in by the ads would reach a booking. In this case, handling inquiries was the starting point for automation: advertising was already bringing people in, and the clinic was losing them at the next step

After that, we built a complete automated workflow around a reception bot. It began answering customers instantly across all channels, taking them through to a booking and reserving time in the clinic's calendar. That immediately started bringing in money

As a result, I helped increase the clinic's revenue several times over. Simply because I looked in the right place: the part of the process where customers and money were already being lost

My recommendation to you is to always start with money, time and customers. That is where a business AI audit begins: understand where the business is losing results, then move on to automation

AI voice agent: why it needs a system

MatIntellect — Mon, 26 Oct 2026 13:04:31 +0000

An AI voice agent is the future I imagined for working with artificial intelligence. Like in "Iron Man": the main character speaks to his system, and it carries out his commands in real time

That moment has arrived. Codex lets you work with files and code on your own computer and set up interactions with your system. OpenAI's desktop app supports folders, tools and tasks. Having a tool available, though, does not mean the whole system is already built

The system has to exist in the first place. A voice conversation with ChatGPT does not, by itself, give you a complete AI architecture. Conversation can be useful on its own. If you want to give instructions that lead to action, something behind the voice needs to carry them out

What sits behind the voice

That architecture includes agents handling regular tasks, along with code and scripts for recurring actions. The voice assistant connects to them and passes your instruction to the part of the system that can handle it

A conversation and a completed task are different things. Voice lets you express a command without a keyboard. Turning that command into an action requires access to the relevant program or data and a configured way to perform it. A microphone alone cannot provide that

Recurring work also needs its own setup. If an action has to run repeatedly, the system needs to know what to launch and when. Speaking the same instruction all over again each time is a different way of working from setting it up to run as a recurring task

Build the system you want to talk to

So "AI voice agent" can describe very different setups. In one, you talk to a model and get an answer. In another, the same voice gives you access to a system of agents and actions that are already configured. You cannot tell the difference just by listening to how natural the voice sounds

That connection matters when building an AI voice agent: which tasks can it pass on, and what will actually perform them? Without that connection, a polished conversation remains a conversation

A voice agent without infrastructure is a pretty toy. A voice assistant connected to a developed, multilayer system is a serious tool, and that is where the future lies. You choose which path to take

AI for business problems: scale without firing people

MatIntellect — Sun, 25 Oct 2026 01:22:51 +0000

The problem with AI in business, as usual, is inflated expectations. I was exactly the same after watching AI ads and hype videos on social media

Bloggers promise that a couple of clicks will launch a hundred agents that will practically run your business for you. But you find out what is true by going and trying it yourself. So I went and tried it

You can scale a business. Firing everyone is a bad idea

You really can scale your business with AI. But firing all your employees is the worst thing you could do right now. There is still a person behind artificial intelligence who makes decisions and presses the buttons

That is why it makes sense to look at AI in terms of scale: processing more orders, speeding up individual processes, and improving reporting, finance and accounting workflows. These are areas where help with the workload is useful, while your employees' experience is still needed to judge the results

People have the practical experience to tell good work from mediocre work. Remove the people who can do that, and processing a larger number of tasks will not replace their judgment. Handling an order faster and knowing whether the work was done well are different parts of the process

Which tasks are a better fit for AI?

Routine tasks lend themselves well to automation with AI. Repetitive work is where it makes sense to look for applications: order processing, preparing reports, and individual operations within finance and accounting workflows

It helps to separate the routine operation itself from the decision someone makes based on its output. Preparing a report is one part of the job. Assessing it and deciding what to do next is another. Human experience is still needed after preparation becomes faster

In my experience, the more complex, multi-part and intellectually demanding a task is, the harder it is to optimize this way. The promise of running an entire business independently bundles many of those tasks into one appealing advertising line

That distinction is where a conversation about AI consulting begins: which routine work can be automated, and where you need a person who makes decisions and can tell a good result from a mediocre one

So I wish everyone a more thoughtful approach to artificial intelligence, using it in ways that actually help them. More to come...

Telegram bot development: business work in topics

MatIntellect — Sat, 24 Oct 2026 02:21:30 +0000

For business, Telegram is one of the most capable platforms available right now. I find it especially convenient to create groups, add bots and keep each area of work in its own topic

What you can build in Telegram

Buttons, cards and different reply flows, all the way up to custom applications inside the messenger. The format depends on the task: sometimes a button beneath a message is enough; sometimes you need a separate screen with a form

The Telegram documentation describes message buttons, commands and support for files, text and voice messages. A button can send a command to your program or open a link. Mini Apps let you build your own interface inside Telegram when a conversation alone is no longer enough

Separate workstreams in one group

We create topics for the tasks we need and work on each area in its own thread. It is convenient and helps keep the work organized

For example, I have a group called "Factory". It has separate topics for ideas, posts, design, copywriting, marketing, video, publishing and archives. And that is far from the full list

There is a straightforward benefit to this arrangement: design discussions stay in their topic, and marketing discussions stay in theirs. To return to a particular area, you open its topic. All those conversations still belong to the same group

What an AI chatbot needs behind it

Topics organize conversations; they do not carry out the work themselves. Buttons and messages are the interface. Behind that interface, a program has to receive the request and perform an action. Open-ended AI conversations require a connected model; work involving other services requires the relevant integrations

Building a bot therefore involves two separate parts: what a person sees in Telegram and what the system does after their message. Having a button does not, by itself, mean the process behind it has been set up. That part belongs to AI system development

Which messenger to start with

If you are planning to build an automation, start by looking at the messenger where you already spend most of your time. In my case, that is Telegram. What is yours?

Instagram DM automation: booking with checks

MatIntellect — Fri, 23 Oct 2026 01:13:09 +0000

Answering every question is the hardest part of booking clients through Instagram. There can be a huge range of conversation paths, just as in any business. The AI receptionist's job is to keep the client engaged throughout that conversation

Your client is a real person, your conversation partner. They can tell whether they are talking to AI. So when the conversation is complex, I recommend being open about the fact that AI is answering. In my view, that builds more trust than a chatbot pretending to be a real person

What goes to the model, and what goes to code

You can give the model the varied conversational work. People ask questions in their own words, clarify details and return to something discussed earlier. The model can handle this changing flow without reducing the entire exchange to a rigid set of lines

But the same conversation contains things that need strict checking: prices from your price list, dates from your calendar and appointment times. I recommend building those parts in Python and making code a participant in the conversation

The model keeps talking to the client while code checks prices and availability against the price list and calendar, rather than leaving the model to recall them. Each has a specific job: the model phrases the response, and code checks the actual booking details

Rules for the conversation

Having many possible conversation paths does not mean giving the model permission to say anything. Set boundaries in advance. Whatever the client asks, the reply should stay within the rules you defined before launching the chatbot

I recommend adding a second model to check the replies. It compares the first model's answers with those rules. One model handles the conversation; the other checks whether the prepared response follows the restrictions you set

Checking a reply and checking calendar availability are different tasks. The reviewing model checks the conversation rules, while code handles exact data. That way, the model that wrote an answer is not the only component responsible for reviewing it

A Direct booking involves both a conversation and specific appointment details. With an AI receptionist, define where the model talks, where code checks the data and which rules are used to review the reply to the client

AI content factory: scaling, not replacement

MatIntellect — Thu, 22 Oct 2026 00:10:09 +0000

After a year of working with a content factory, I have reached this conclusion: it does not fully replace and automate content work. It helps you scale that work, make it simpler and get it done faster

When I compare real Reels videos, photographs and posts with material generated by artificial intelligence, the real content gets several times more views and audience reactions. That is why I recommend using a content factory to scale your work across social networks, but never replacing all the content you make with AI. People notice and react poorly

What gets the weaker response

I mean entirely generated content with no real person in it at all. This is not about Instagram's "AI info" label: the weaker response is to the material itself. Those are different things, and attributing that response specifically to the label would be wrong

The comparison above includes Reels, photographs and writing. The question goes beyond whether a neural network can write a post: a video can be entirely generated too. The fact that you managed to make something without filming it says nothing on its own about how an audience will receive it

What to hand over to a content factory

In my content factory, an agent collects topics from channels I follow. I can pick a suggested idea or bring my own. The material then goes through fact-checking: does the thing being discussed exist, is there an official announcement, can the claim be verified?

Next come versions for different social networks. Each platform has its own audience, so the style of the writing changes. Once the material is ready, it can be published to all connected platforms with one click. I described that process in "Why do you need an AI content factory?"

Where the person stays involved

None of my posts goes out automatically without my supervision. A suggested topic is not yet a decision to publish, and a prepared text does not become my position just because a neural network wrote it

That is where an AI content factory belongs: making it simpler to run several social accounts. Real videos, photographs and the author's own posts remain part of the work. I do not recommend removing them entirely for the sake of automation

AI in sales: where to start

MatIntellect — Tue, 20 Oct 2026 03:54:36 +0000

In my view, AI delivers the most value in sales where inputs and outputs are clearly defined. Tasks that are easy to measure and involve fewer steps are the easiest to automate

If you hand all your sales and customer conversations over to AI, it will most likely be a complete failure. There are too many ways an interaction with a real person can unfold

Giving it strict, repetitive steps from your existing process is a reasonable move. My recommendation is to automate gradually, starting with the simplest work instead of handing everything over at once

Start by recording incoming inquiries

The first task should be recording every incoming inquiry and tracking how it is handled. Not even answering customers yet. Just recording inquiries so none go missing and you can see the progress of each one. At this stage, it is ordinary systems analysis

This task has clear boundaries: a customer inquiry comes in, and the output is a recorded inquiry with visible handling progress. You can check the result against that specific inquiry: was it recorded, and can you see what is happening to it next?

Receiving an inquiry and handling it are different things. A record of an incoming inquiry does not, by itself, show whether work on it has moved forward. That is why tracking needs to cover how each inquiry is handled, rather than just listing everything that arrived

Why not start with conversations?

Recording an inquiry can be separated from talking to the customer. The first task is to preserve incoming information and track handling. The second requires choosing how to respond as the conversation develops. That variety of possible situations is precisely what makes handing every conversation to AI risky

How to choose the next step

Once inquiries are being recorded, choose the next part of the process using the same criteria: clear incoming information, a defined output, and a short, repetitive sequence. This keeps the task for an AI employee limited to a specific action, rather than turning it into an instruction to run all sales independently

The outcome of the first stage is specific too: incoming inquiries are not lost, and the progress of each one is visible. That is the starting point for gradual automation

How much does an AI agent cost: what shapes the price

MatIntellect — Sun, 18 Oct 2026 02:30:10 +0000

Before answering how much an AI agent costs to develop, you need to understand how it differs from an ordinary GPT chat. The price depends on the functions you want it to have

What are we actually building?

An AI agent is a system running on your remote server or your computer. It has its own memory and working tools, and it can run around the clock

In an ordinary chat, you ask a question and get a response. A separate agent can keep working while the app on your phone is closed. To do that, it needs connected tools and configured tasks it can carry out without your involvement

That makes developing an agent a little more involved than talking to an AI in an app. The system needs to be set up for a particular job: what it remembers, which tools it uses and which services it connects to

How functions affect the cost

Say you want a personal assistant. It puts together your daily task list, reviews your conversations and helps you stay productive. Connecting it to a large number of different services increases the cost. But adding connections does not, by itself, mean the price will multiply

Even the word "assistant" can describe different jobs. Task lists, reviewing conversations and service connections are separate functions that need to be listed before an estimate. Otherwise, we could be using the same name to discuss very different amounts of work

The description of custom AI agent development is a starting point. The final cost depends on the particular set of functions you need

Why I recommend separate agents

For each area - marketing, copywriting, design or customer service - I recommend a separate agent. That keeps one assistant's functions and memory from becoming a jumble

The more different roles we load onto a single agent, the harder it becomes for it to do a good job in each individual area. Customer tasks, writing and marketing end up inside one assistant, even though they are different kinds of work

It is the same at work. We cannot give one assistant the jobs of an accountant, economist, receptionist and developer and expect them to do all of it well. We need separate specialists

The same approach applies to agents: first define the functions you need, then distribute them among assistants. That lets us discuss the cost of a specific scope of work, rather than an abstract agent that is supposed to do everything