01 management
An answer, not a report
«How many leads this week, where from, and why fewer than last week». Instead of asking three people for three exports and merging them in the evening, you ask once and get the breakdown.
About
A colleague who reads
all your systems at once
You write to it in Telegram in plain words: «how many leads this week and where from», «draft a quote for this request», «how is the ad account doing». It goes to the CRM, the mail, the ad account and the phone system itself, does the work and reports back.
Installed on your own server. Conversations, documents and keys stay with you.
In plain words
What this is
Picture an assistant who knows where everything of yours is kept. It can open the CRM and look at deals, go into the ad account and see the spend, pull up call history, find the right price list in the mail. And it does all of that from a single request written in ordinary language, not from a twelve step instruction.
It does not replace people. It takes away the part of the work that consists of «open four places and add it up by hand»: collect the numbers, find the document, prepare a draft, remind, check whether something was dropped.
It runs inside your perimeter. This is not a website you upload your customer base to. It is a program on your own server holding your own credentials, and the data does not travel anywhere.
Who
Whose work it takes off
02 sales
A quote in minutes
A manager forwards the customer request, the agent finds the items in the price lists, calculates and assembles a draft. The person checks it and sends it.
03 marketing
Ads in human language
«What dropped this week and where did the money go». Spend, campaigns, calls by source, in words rather than a two hundred row export.
04 tenders
Procurement watched
Follows new tenders against your criteria, reads the documentation and brings only what fits.
05 inbox
The flow sorted
Reads incoming mail and requests, groups them by topic, lifts the urgent to the top and reminds about what has been left hanging.
06 meetings
Minutes on their own
Turns a call recording into a list of decisions and tasks instead of a wall of text.
Scenarios
What it looks like in practice
| Task | The usual way | With the agent |
|---|---|---|
| Weekly lead summary | export from CRM, export from call tracking, merge in a spreadsheet | one question in chat, an answer broken down by source |
| A commercial quote | find the price list, calculate, fill the template | the agent assembles a draft, a person checks and sends |
| Advertising report | open the account, export, match against deals | a question in words, an answer with spend and drops |
| Finding tenders | walk the platforms, open the documentation, filter | the agent brings the fitting ones with the requirements read |
| Going through mail | read everything, sort, remember to reply | sorted by topic, urgent on top, hanging items with a reminder |
| Meeting minutes | listen to the recording again, write out decisions | transcript and task list right after the call |
| A recurring task | remember it and do it by hand every week | done on schedule, the result arrives in chat |
Economics
Where the gain comes from
We will not quote percentages: in someone else's company they are invented. You can work it out from your own numbers, and here are the four places where it usually adds up.
- Time spent collecting data. Multiply how often a week someone assembles a summary by hand by how long it takes. That work disappears entirely: the agent reads the systems directly.
- Speed of answering a customer. A quote sent the same day and a quote sent three days later are different deals. A draft in minutes changes conversion, not convenience.
- What gets lost. A request with no answer, a mail with no reaction, a tender noticed too late. That is not saved time, it is revenue never earned, and the journal makes it visible.
- Decisions on data rather than memory. When a number can be asked for out loud in ten seconds, people ask. When it takes two hours and three people, they go with a feeling.
And a fifth, less obvious one: the work stops depending on the one person who «knows where to look». The sequence of steps is written down for the agent and repeats identically.
Strategy
How to roll it out so that it works
01
One process, not ten
Take the one that happens often and hurts: the lead summary, preparing quotes, triaging the inbox. Ten scenarios at once works for nobody: people do not have time to adapt and everything falls back to manual.
02
Measure how it is now
How much time the chosen process takes per week, how many requests go unanswered, how many days pass before a quote. Without that measurement there is nothing to compare against in a month, and the conversation ends in impressions.
03
A month with confirmations
At first the agent asks permission before every outward action: a mail, a change to a deal, a document sent. You see what it intends to do before it does it. Trust is earned in front of you rather than granted up front.
04
Drop confirmations where trust exists
When «yes» has been pressed a hundred times in a row for one specific action, that action moves to automatic. Everything else keeps asking. That is what a productivity gain actually is: not «the agent does everything», but «the person stopped taking part where they were not needed».
05
Expand one at a time
The next process is connected when the previous one has run without you for a week. That way a quarter produces four or five working scenarios instead of ten abandoned ones.
Concerns
What people ask first
Where does our data go?
Nowhere. The agent is installed on your server, the keys to your systems live there, and so do the conversations and documents. The only thing that leaves is what you send yourself, plus the request to the model if you chose a cloud model. There is no telemetry in the code.
How is this different from a normal AI chat?
A chat cannot see your deals and does nothing: it advises. The agent is connected to your systems with your own credentials and does the work. The difference is not the model, it is that the agent has hands.
What if it breaks something?
Dangerous actions require a human confirmation by default, and everything done is written to a journal with its outcome. You see what it intended, what it did and how it ended. The list of what requires confirmation is configurable.
Will it replace our staff?
It replaces moving data between browser tabs, not people. Experience, negotiation and decisions stay with the human. In practice the effect is usually that the same team carries more, not that somebody is let go.
Do we need a programmer for this?
Installing it needs someone who can put a program on a server: about an hour. After that, configuration goes through a web page and chat. If there is no such person, we install and configure it.
What happens if we stop paying?
The agent stays with you and keeps working: it runs on your machine, not in our cloud. What stops is support and updates, not the agent.
Systems
Connects to what
you already run
Nothing has to move: the agent comes to your systems. Next to each one it honestly says whether we verified it on a production account or so far only against the service documentation.
Bitrix24
portal structure, fields, dictionaries, deals and smart processes
verified in production
Yandex.Direct
campaigns, groups, reports, spend and API limits
verified in production
UIS / Comagic
calls, forms and the quality of source tagging
verified in production
Calltouch
calls, requests and advertising sources
verified live
Web search
Yandex Search API, SearXNG and Brave under one result shape
verified live
Browser
reading script-rendered pages, and actions behind confirmation
verified live
Yandex.Mail
folders, search and reading messages
ready, awaiting your access
ZenMoney
accounts, balances and transactions; writes require confirmation
ready, awaiting your access
VK
community messages and Long Poll state
ready, awaiting your access
Roistat
analytics and comparison of attribution models
ready, awaiting your access
Alice
a voice skill as one more surface for the conversation
ready, awaiting your access
MeetScribe
meeting transcripts and minutes
ready, awaiting your access
Roadmap
Q3now
2026
Launch
Open the public repository and bring self service to the point where a customer sets the agent up alone.
- Public repository
- Cabinet: the customer creates the cell
- Spend in numbers and a human readable work journal
- Voice through a live run
Q4
2026
Growth
Vision and work on other machines. The expensive part: a message inside the core is a string today, and an image does not fit into it.
- The agent sees images and screenshots
- Work on a laptop and a Mac through live runs
- Connecting third party extensions over the network
- Live runs on Russian models
Q1
2027
Reach
Beyond a single server: the phone, the English language and third party extensions.
- Mobile client
- English interface
- Third party extensions with review
Q2
2027
Directions
Dates will be set from the results of the previous quarters. No day level promises here.
- Industry specific builds for well understood processes
- Self service end to end without us
Honest
What is not there yet
- There is no mobile app. The phone works as a remote through Telegram, and for our scenarios that is enough.
- The agent does not see images. Sending it a screenshot and asking it to work things out is not possible yet: that is Q4 work.
- It does not count money in currency, it counts calls to the model. Provider prices differ and change, and an invented number in a report is worse than a missing one.
- Some systems are waiting for your access. The code is ready and covered by tests, but we have not run them against a production account. Which ones is written above, next to each.
Next
Start with one process
and look at your own data
Tell us what hurts and which systems you run. We will say whether it is worth taking on, and name a timeline. Installing it yourself and trying it out is free.