Find out in one month where AI pays off for your people

For companies of 10+ people whose employees already use AI on their own. Every employee gets their own agent in the messenger they already use. In a month you know which tasks the agent pays off on, which it doesn't, and how many tokens — that is, AI model usage — to buy next. If you don't continue, the pilot costs $300. What's in the report ↓

Request a pilot
Where does your data go?Company documents go into employees’ personal AI accounts. The company can’t see what was shared or who gave access.Where does your data go?COMPANYPERSONAL AI ACCOUNTSEmployeeClient basePersonal AIEmployeeContractPersonal AIEmployeeFinancial reportPersonal AIWhat was shared?Who gave access?The company can’t see what employees share with AI

Where does your data go?

Company

Client base · Contracts
Financial reports

↓ What was shared? Who gave access?
Employees’ personal AI accounts

The company can’t see what employees share with AI

TurnkeyYour server or our cloud
One monthFrom setup to numbers
$300if you don't continue

Everyone is rolling out AI, and almost no one knows if it paid off

95%

of corporate AI pilots showed no measurable impact on profit. Don't agree to an expensive pilot. There is no point paying a fortune upfront for a project that is this likely not to pay off.

MIT NANDA, “The GenAI Divide: State of AI in Business 2025”: 300 public deployments and a survey of 153 executives
Today
Your company today Eight employees inside the company perimeter. Five of them use their own personal AI outside the perimeter: ChatGPT, Claude, Claude Code, Codex, DeepSeek, Cursor, some of them two at once. Each one holds company secrets: deals and margins, the client base, source code and API keys, payroll, CRM exports and the database password. Three employees don’t use AI at all. company perimeter ChatGPT personal account deals & margin strategy CEO ChatGPT + Claude personal account client base discounts Anna sales no AI Igor marketing no AI Lena HR Claude Code + Codex personal account source code API keys Oleg engineering DeepSeek personal account payroll reports Masha finance no AI Pasha warehouse ChatGPT + Cursor personal account CRM exports DB password Sergey analytics Company secrets sit in employees’ personal AI: access, clients, code, payroll You can’t see what is where, and you can’t get it back when someone leaves 3 of 8 don’t use AI at all. What the rest can do, nobody knows
Everyone picked their own ChatGPT, Codex, Claude Code or Cursor, some use two.
14%

more support tickets resolved per hour with an AI assistant; newcomers improved by 34%

NBER, 5,179 support agents, 2023
40%

better work from consultants using AI on tasks it suits. On tasks it doesn't, 23% worse. The pilot shows which tasks are yours

BCG and Harvard, 758 consultants, 2023
13%

of employees say their organization performs significantly better thanks to AI

Glean Work AI Index, 6,000 office workers, 2026

Every employee gets an agent, on your server

With Fleetbox
Your company with agents Six agents inside the company, each employee has their own agent for their job. The agents message each other. company perimeter · your server Digests Ads Reports Hiring Code Invoices Company agents agents message each other CEO Igor Sergey Lena Oleg Masha log AI model only the question and needed context ChatGPT · personal Claude Code, Codex stay personal, can’t see the company
One agent — one employee and one of their tasks. Each agent has its own memory and access. Agents message each other and see a task through. The employee talks to their agent in the messenger. Everything sent to the AI model and received from it is written to a log on your server.

Works where your team already is: Telegram out of the box, other messengers on request. Nothing to install on computers: agents are available from a phone 24/7.

What it looks like: an employee messages their agent

Agents · Fleetbox
Темы
Pinned message
Working rules and instructions
Open
Do we have 40 units of item 1182 by Friday? The client is waiting.
11:05
W
Sales agentbot
Asked the warehouse agent. In stock: 32 units. A delivery of 20 arrives on Thursday. By Friday we will have 52. Reserve 40 for the client?
11:05
Yes, reserve.
11:05
W
Sales agentbot
Done: 40 units reserved until Friday. The warehouse agent confirmed.
11:05
Write a message…

Ads, leads and shipments: real conversations of our agents and an iiworker client, shortened and translated. Stock and digest: training examples.

What the agent does on its own, and what only after your “yes”

  • Works with its employee's permissions, no wider. What it did in the ERP or CRM shows up in the system history as the employee's actions.
  • Prepares the action: a text, a record, a task. Critical actions only after the employee confirms.
  • Without separate approval it makes no payments, legally binding sends or mass mailings, does not change financial data and does not delete records.
  • Everything sent to the AI model and received from it is logged on your server.

Where your data goes is your call

$

Global model

For example, GPT. Fastest start and lowest price. Questions go to the model provider's cloud.

$$

Regional model

Data stays in your country or region, which helps meet data residency rules. Costs more than a global model.

$$$

Model on your servers

Everything stays inside the company perimeter. The most expensive option.

Agents, their memory and access live on your server or in our cloud, your choice. We set it up turnkey.

Week 1

Agents appear for employees. Each agent asks its employee about their work and suggests where to start. Together we look for repetitive manual work. Every week of the pilot we run a 2-hour training for employees.

Weeks 2–3

Based on real requests we connect spreadsheets, CRM, ad accounts, marketplaces, ERP.

Week 4

Report: who uses it and how often, what got automated, what an agent costs per employee per month.

An open core with our tools around it

Every agent runs on the open-source Hermes Agent. Tools plug into it over MCP, the common standard for AI agents. Pick a tool to see an example.

CoreHermes Agent

Open-source agent by Nous Research. A solid base: nothing to build from scratch, and it improves every day.

  • Built by about 400 developers
  • Over 250,000 GitHub stars, tested by thousands of people
  • Open source, MIT license

trip2g knowledge base

The agent's notes, searchable by meaning, with dashboards on top. Share a folder with colleagues' agents and the team works from one context.

ExampleThe head of support opens a dashboard: which requests have no answer yet and from which clients. The agent rebuilds it from the chats every night.

Chat indexer

An archive of the Telegram groups and channels the connected account is in. The agent searches it, analyses the chats and draws context from them.

ExampleEvery morning it sends posts from industry chats where someone is looking for a contractor, each with a link.

Agent mail

Company agents pass letters and tasks to each other. A letter that needs a person arrives in chat with buttons.

ExampleThe ads agent writes to the web agent: “We need a landing page for ‘warehouse rental’.” The web agent takes it on.

Bot proxy

One bot for all your agents, each in its own group topic. Telegram out of the box, others on request.

ExampleSales works in one group: the sales agent answers in “Deals”, the warehouse agent in “Stock”.

Web interface

Talk to the agent in a browser, with its files at hand. Sign in with Telegram. Available on request.

ExampleA big spreadsheet is easier on a laptop: open the agent in the browser and work through it there.

Secrets and authorizations

Secret and authorization management: the agent works with Google, Yandex and API accounts without seeing passwords or keys. Risky actions run only after a person confirms.

ExampleClosing the financial month: the agent prepares everything, a person presses the button.

Browser

The agent's own browser: opens sites, fills in forms, downloads files. Sites behind a login go through the desktop.

ExampleWent through five suppliers' sites and put their prices for one item into a table.

Desktop

A shared screen with a browser, Telegram and a terminal. The agent clicks and types; you watch or step in.

ExampleThe supplier portal has no API. You sign in yourself on the shared screen, no password in chat, then the agent works.

Telegram account

The agent acts as a person: reads chats, writes first, creates groups. Reading is rate-limited.

ExampleFound someone in a founders' chat looking for a contractor and drafted a message to them.

Tools are switched on one by one, to fit each employee's work. Passwords and keys are stored outside the agent.

More MCP

connected per client

We connect your systems for each client, based on what employees actually ask for; 1C and Bitrix24 per project.

  • Google Sheets & Drive
  • Yandex Direct
  • Yandex Metrica
  • Company check by tax ID
  • 1C
  • Bitrix24
  • Ozon
  • Wildberries
  • Your system

In a month you decide on numbers, not impressions

What you keep after the month

  • How many tokens to buy: usage per employee and per task.
  • Which tasks the agent pays off on, and which it doesn't.
  • What employees tried with the agent, what they didn't, and what they should have.
  • Working agents. They stay with employees until you decide.

Pilot for 25 people

Pilot cost
Server for 25 agents, yours or in our cloud, set up turnkeyabout $100/mo
AI model during the pilot$200/mo
Platform license during the pilot$0
Your risk if you don't continue$300

What it costs to find out another way

Integrator project, paid upfront, no guaranteed resulttens of thousands $
Courses for 25 people: you get trained people, not agents$8,900–29,600
Fleetbox pilot if you don't continue$300

After the pilot

Company license, one-time, no per-user fees$4,700
Tokens for the models you choosedirect, no markup
Serveryours or our cloud, at cost
Supportyour IT team or us, $60/h

Who builds it

Founder

Alexey Yurchenko

15 years in software. Accounting and warehouse systems for Fresh, MoscowFresh and Family Friend, about a thousand orders a day. Author of the open-source project trip2g, a knowledge base for agents.

CTO

Pavel Yurchenko

CTO of a product for transport companies: 200,000 carriers in the database, pricing across 130,000 localities. Two years of building agents into a logistics system.

  • A pilot is running now at a Moscow agency with 40 employees.
  • iiworker.pro sells agents to individuals on the same platform. An iiworker client creates Ozon shipments with one message and saves about two hours a day.

Start a pilot

We reply within a couple of hours. In 30 minutes we'll show a live agent in your messenger on your task.

Or message us on Telegram: t.me/jrpcd ↗