Every business runs on the same core processes: hiring, sales, production, quality control, client communication. And every one of them has routine work you can hand to an AI agent. Not "hook up ChatGPT," but build a digital employee who knows your business's context, has access to your systems, and carries out tasks on its own.
I've spent 12 years systematizing businesses. For the last year and a half I've been building agents like this, first for myself, then for clients. Everything below runs in production right now. With metrics, with real cases. This isn't theory, it's things you can actually implement.
1. Recruiting: hiring without the manual grind
Automatic screening of HH.ru applications
One of our clients, a design agency with 1,500+ projects and clients on the scale of Gazprom, Rosatom, and Sberbank, had a project manager drowning in applications. 200-300 responses a day for a single designer role. Manually reviewing portfolios, filtering out freelancers, weeding out irrelevant candidates was eating 3-4 hours of her day.
What the agent does:
- Pulls in every application through the HeadHunter API (OAuth, incremental loading, adaptive throttling)
- AI scores each candidate against the client's criteria: 1-3 years of in-studio experience (not freelance), an actual portfolio, realistic salary expectations
- Assigns a rating: 1 star (exact match), 2 stars (with reservations), 3 stars (not a fit)
- Automatically rejects category-3 candidates through the API, with a polite template
- Sends the PM a Telegram report: "15 new, 3 relevant, here are the top 3 candidates"
Result: out of 200-300 applications overnight, the agent surfaces 5-10 relevant ones. The PM reviews 10 profiles instead of 300. That's 3 hours saved every day.
Onboarding a new employee
When someone joins the company, there are 15+ things to do: create a CRM account, set them up in Notion, grant access to folders on Google Drive and Yandex.Disk, add them to Telegram groups, assign a mentor, send a welcome message with instructions.
I used to do all of this by hand, and I always forgot something. Now the agent gets a command like "onboard Elena Sazonova, business analyst" and handles everything in 3 minutes:
- Creates the user in Bitrix24 under the right department
- Adds a card in Notion (employee database, start date, role)
- Creates a project folder on Google Drive and Yandex.Disk
- Creates a Telegram group from a template, adds members, assigns admins
- Sends a welcome message with links to the knowledge base, policies, and contacts
15 tasks in 3 minutes. It used to take an hour of switching between five different systems.
Offboarding: revoking every access in an hour
Offboarding is the reverse process, but it matters even more: you can't afford to miss a single access point. The agent audits every system: finds the employee across 49 Telegram groups, deactivates their CRM account, archives them in Notion, and produces a "what needs to be handed off" checklist. Instead of 2-3 days of manual work (with the inevitable "oh, we forgot about that group"), it's a one-hour audit, done completely.
2. Sales: from meeting to invoice without losing context
Transcribing and breaking down meetings
I have 3-5 client and candidate meetings every week. An hour-long meeting used to mean an hour of context that got lost in scribbled notes.
Now the process looks like this: record the meeting, the agent transcribes it via Groq Whisper (37 minutes of audio in 43 seconds, that's 50x real time), AI pulls out the key points, structured notes land in Notion.
From the transcript, the agent extracts:
- Pain points, the client's own words about their problems
- Budget signals, any mentions of budget or figures
- Buying signals, "when can we start?", "what do you need for the contract?"
- Action items, who promised to do what
A 59-minute meeting becomes a full breakdown in 9 minutes. And that breakdown then becomes the backbone of the proposal.
Creating the CRM deal straight from the meeting context
After breaking down the meeting, the agent creates the deal in Bitrix24 itself: company name, contact person, amount (if mentioned), funnel stage, a comment with the key points. All I have to do is review and approve.
Personalized proposals
A proposal isn't a template with the company name swapped in. The agent reads meeting transcripts, pulls the client's own quotes about their problems, picks relevant case studies from the portfolio (a construction client gets a construction case, an IT client gets an IT case), and frames the challenges through a "problem leads to consequence" structure.
The output is a PDF with genuinely personalized content. 15 minutes instead of 2 hours of manual work. And it lands, because it uses the client's own words.
Generating contracts and invoices
Contract: a Google Docs template, auto-filled with the details, AI checks the specifics (counterparty type: individual, sole proprietor, or LLC, plus a project-specific appendix), highlighted in yellow for manual review, sent to the client on Telegram.
Invoice: one command, and the agent creates a Smart Invoice in Bitrix24 with the right amount, linked to the deal, with line items. Two modes: standard monthly consulting, or a fixed-scope AI project.
Client onboarding: 7 phases, one command
The client has paid. What's next? Seven phases:
- Gathering context, the agent reads the Telegram history, the CRM deal, meeting notes in Notion
- CRM, updates the deal stage
- Notion, creates a project page with the standard structure
- File storage, folders on Google Drive and Yandex.Disk
- Assignment, links a business analyst to the project
- Telegram group, creates a group from the template "Name. Company. Systematization. Systemmatica," adds the client, the analyst, me
- Welcome message, a greeting, the links, the plan for the first step
Everyone's briefed, everything's created, nothing's forgotten. One command, 15 tasks across 5 systems.
3. Production: quality control and reporting
Assembling a brief for contractors out of chat history
This is a real case from a design-agency client. The pain: the project manager was manually piecing together revision notes from chats. The client writes some as text, sends some as voice messages, drops screenshots with annotations, all scattered across separate messages over 3 days.
The agent: reads the chat for the period, transcribes every voice message, reads the screenshots with annotations, and assembles it all into a structured brief in Google Docs. With links to Figma, revisions grouped by page. The PM gets a finished document and sends it straight to the designer.
A 78-minute client meeting (a 2.9GB recording) became a full transcript plus a structured brief with 9 sections of revisions, delivered to the designer in a single session. Instead of 2-3 rounds of "wait, what did the client actually say?"
Reviewing the deliverable against the brief
The designer delivers a mockup. The agent takes the PDF or the design file, takes the original brief, and checks: were all the revisions made, is anything missing, does the result match the requirements. Intermediate quality control, no human required.
A report on the work done
The client needs to see the results. The agent pulls context from every source (chats, Notion, CRM) and builds a 7-page report:
- Cover, big numbers (12 interviews, 47 sections, 8 instructions), so the client immediately sees the scope
- What was agreed, the scope, a baseline for comparison
- What was done, specific document names, not abstractions
- Above and beyond, bonuses, called out visually. The client doesn't know it's a bonus unless you say so
- What was found, expert observations, a neutral tone
- Recommendations, a numbered, actionable checklist
- Next steps, the logical continuation
A finished HTML page turned into a PDF. One report instead of an hour spent laying it out in Google Docs.
4. Messaging: automatic communication control
Monitoring work chats
We have a rule: business analysts report in the team's work group daily by noon. The agent watches the group, and if a report doesn't show up, it sends a reminder. The rules (who, when, in which chat) live in Notion and load automatically.
An important detail: the agent reads the last 5-10 messages, not just the latest one, because a person might report across several messages, and you can't judge from just one.
Auditing an employee's tasks
Want to know what your employee actually got done this week? The agent exports the chat history (work groups plus direct messages), finds every task that was assigned, and builds a report: what's done, what's in progress, what's stalled. Voice messages get transcribed automatically.
Crossposting
Write a post once, and the agent adapts it for each platform's format (Telegram, VK, Instagram, YouTube Community) and publishes it. The text itself doesn't change, only the technical adaptation: length, formatting, hashtags.
5. CRM analytics: the money you're losing right now
Connecting to the CRM and analyzing everything it holds
The agent has full access to Bitrix24 through the REST API. What that enables:
- Lost clients, finds deals stuck at the "Negotiation" stage for longer than N days, and builds a list with context: date of last contact, what was agreed, why it went quiet
- Funnel control, how many deals sit at each stage, where the bottleneck is, average time to conversion
- Data audit, finds duplicate contacts, deals not linked to a company, contacts missing a phone number
This isn't a dashboard you have to open and interpret yourself. It's an agent that analyzes the data on its own and comes to you with conclusions: "12 deals have been sitting at the proposal stage for over 2 weeks. Here's the list, want me to nudge the managers?"
Morning briefings
Every morning at 7:00, the agent pulls together a digest: new messages in work chats, today's tasks, deadlines, reminders. All in one Telegram message. No need to open five apps: you open your phone, read the briefing, and know how the day looks.
6. What I wouldn't recommend doing
For the love of god, don't outsource your "expert content" to a neural network.
If you position yourself as an expert and run a personal blog, share your own experience. Don't generate articles out of "information gathered from the internet." A neural network compiles beautifully, but compiling someone else's experience isn't expertise, it's noise.
I see it every day: companies publishing 10 ChatGPT-written articles a week on their site. Smooth, correct, and completely useless. Not a single real number, not one case study, not one verifiable detail. That's not content, that's noise.
An AI agent should help you process and structure your own experience, transcribing your meetings, analyzing your data, building reports out of your own work. Not write texts for you that you never actually lived through.
Every automation in this article comes out of my daily workflow. I can show you the code, the screenshots, the metrics. If you can't name a single concrete figure from a real project in your article, that's worth pausing on before you publish it.
How this is actually built
The main difference from n8n, Make, or Zapier: those tools automate rigid chains. Our agent makes decisions. It reaches for tools as the situation calls for them, whether that's one function or ten, depending on the task. This isn't automation, it's a genuine assistant.
Under the hood:
- Shared memory, the agent remembers the context of the company, its processes, its clients
- Company profile, not a role prompt, but structured context
- A terminal/CLI, free-form chat, no commands to memorize
- An auto-insight module, the agent records what it learns and applies it going forward
- Full portability, every file, script, and instruction stays with the client
Integrations: Bitrix24, amoCRM, Notion, Google Docs/Drive, Telegram, WhatsApp, HeadHunter, Yandex.Disk, and any other system through its API.
Where to start
Two paths, pick based on your situation:
Want to figure it out yourself
I have an online course on AI agents, where I walk through everything: from picking a model to setting up automations. Good fit if you have the time, a technical interest, and want to control the process yourself.
Want the result without the hassle
My team and I handle the whole process end to end. You don't need to learn prompts or APIs, you get a working AI agent built for your business:
- AI Audit, a single meeting where we go through your processes and identify what to automate first
- AI Business Process, we take one process and build a working automation for it: interviews, formalizing the process, training the AI, iterating on real data, 3 months of support
- AI Consulting, weekly meetings: data review, strategic decisions made with AI
First working automations, within 5 days.
More on the service: systemmatica.com/ai-agent